Closes W-004 wiring debt surfaced by L2 audit (#238) and predicted
by L1 audit's forward note (#237). ADR-0006 §"Integration Points"
states: "Vault recall re-activates the region to E2 transiently,
then lets it cool again." Prior to this commit, vault.recall()
returned entries with no energy field at all — the re-thaw was
spec-only.
Changes:
- vault/store.py: import EnergyClass / EnergyProfile from
core.physics.energy. Define module-level _VAULT_RECALL_RETHAW_ENERGY
singleton (raw=0.50, energy_class=E2, mid-band). Both .recall() and
.recall_batch() stamp each returned entry with the re-thaw profile
via a new "energy_profile" key in the result dict.
- tests/test_vault_recall_rethaw.py: 6 tests pinning the contract —
recall returns E2 profile, recall_batch returns E2 profile,
singleton is byte-identical across calls (replay determinism),
empty vault is no-op, min_status filtering preserves the field,
raw value sits unambiguously in E2 band [0.37, 0.62).
Architectural notes:
- The re-thaw is *declared* by the vault, not derived through the
energy operator. ADR-0006 makes the assertion directly; vault
recall is the moment the assertion applies.
- The singleton (rather than a per-call construction) preserves
byte-identical replay: same recall sequence => identical
EnergyProfile object => stable trace if downstream folds it.
- Cool-down per ADR-0006 is downstream field propagation's
responsibility via FieldEnergyOperator's natural recency decay.
Once the recalled entry is no longer being injected into the
active field state, recency drops and energy class falls.
- "energy_profile" is added to recall result dicts, alongside the
existing "epistemic_state" field. Existing consumers (generate/
stream.py:169, chat/runtime.py:1643, vault/decompose.py:124,179,
session/context.py:347) ignore unknown keys — no breakage.
Unlocks W-005 (energy-modulated surface readback) — now that E0/E2
distinction exists at the runtime data shape, downstream readback
modulation can become meaningful instead of moot.
Verification:
- tests/test_vault_recall_rethaw.py: 6 passed
- tests/test_vault_*.py: 48 passed, 4 skipped (no regression)
- core test --suite smoke: 67 passed
- core test --suite cognition: 120 passed, 1 skipped
- core test --suite algebra: 82 passed, 50 skipped
- scripts/verify_lane_shas.py: 7/7 match pinned SHAs (byte-identity preserved)
The test asserts ledger status is in {reasoning-capable, audit-passed},
but ADR-0120 (PR #195, dec98ea) promoted mathematics_logic to expert
without updating this test. Test was failing on main as part of the
full suite (surfaced during PR #239 verification: Codex's versor-
threshold fix ran full suite, found this unrelated failure).
Test's docstring explicitly states the invariant is reasoning_capable
holding while "the status string moves with later promotions" — so
the fix is to extend the expected tuple, not to revert the promotion.
Cleanup per feedback-cleanup-as-you-find: the orphan was a follow-on
of ADR-0120 that should have shipped with the promotion PR.
Verified: 14/14 passing locally.
Implements the PropositionGraph epistemic carrier (ADR-0144):
recognition/carrier.py — EpistemicTransition, EpistemicNode, EpistemicGraph.
Frozen, JSON-serializable, byte-deterministic. EpistemicNode wraps a
RecognitionOutcome with an append-only provenance chain; epistemic_state
property tracks last transition's to_state or outcome.state when empty.
recognition/connector.py — epistemic_node_to_graph_node(). Maps an admitted
EpistemicNode's FeatureBundle (agent/relation/count/unit) to a GraphNode
for the generation-side articulation planner.
CognitiveTurnPipeline gains a recognizer: DerivedRecognizer | None param
(default None — all existing callers unaffected). When attached, run()
calls recognize() at the top of every turn and wraps admitted outcomes in
an EpistemicGraph. CognitiveTurnResult.epistemic_graph carries it.
RuntimeConfig.recognition_grounded_graph: bool = False — opt-in flag that
replaces the intent-derived PropositionGraph with one derived from the
admitted EpistemicNode via the connector.
RatificationOutcome gains three specific PASSTHROUGH sub-values
(PASSTHROUGH_NO_FIELD / NO_VOCAB / NO_VERSOR) for _ratify_intent
observability (ADR-0142 debt 1). All normalise to "passthrough" before
trace_hash so pre-ADR-0144 hashes are byte-identical.
24/24 acceptance tests pass; 67/67 smoke tests pass; no regressions.
* feat(epistemic): populate normative_detail on TurnEvent and ChatResponse
Adds normative_detail_from_verdicts() to core.epistemic_state and wires
it into both the stub and main ChatResponse/TurnEvent construction sites.
The field carries a sorted comma-separated list of violated boundary or
commitment IDs when normative clearance is VIOLATED or SUPPRESSED; empty
string otherwise.
* docs(ADR-0142): ratify epistemic state taxonomy — 14-state vocabulary + normative clearance axis
Formalises the six-subsystem Framing 1 audit findings into a first-class
decision. Accepts the 14-state taxonomy and companion 4-value normative
clearance axis. Documents Phase 3 deliverables already landed and defers
structured provenance + cross-subsystem transition machinery to ADR-0144.
* feat(recognition): output contract + ADR-0143
Adds recognition/outcome.py: RecognitionOutcome, FeatureBundle,
BoundFeature, EvidenceSpan, NegativeEvidence, the three typed refusal
classes (ShapeRefusal, FeatureEvidenceRefusal, FeatureConsistencyRefusal),
and RecognitionProvenance. Frozen dataclasses, JSON-serializable,
byte-deterministic invariants enforced in __post_init__.
ADR-0143 commits to Mechanism D (multi-resolution anti-unification over
token sequences) and defines the two-phase acceptance test.
* feat(recognition): derive phase1 anti-unifier
* feat(epistemic): add first-class state enums
* feat(epistemic): tag TurnEvent with state axes
* feat(epistemic): serialize turn state axes
* feat(packs): tag curated and inferred unit entries
* feat(epistemic): expose word-level state on manifold
* feat(epistemic): expose vault status mapping
* feat(epistemic): preserve pack entry states through compiler
* test(epistemic): cover phase 3 state tagging spine
* feat(runtime): wire epistemic_state + normative_clearance into ChatResponse
Add first-class epistemic_state and normative_clearance fields to
ChatResponse (defaulting to "undetermined"/"unassessable" for backward
compat). Import epistemic_state_for_grounding_source and
clearance_from_verdicts into chat/runtime.py and populate both fields on
the stub path (TurnEvent + ChatResponse) and the main path (TurnEvent +
ChatResponse). Fix the test fixture to use "euro per hour" (a genuinely
composed unit) instead of "dollars per hour" which is a curated lexicon
entry and returns DECODED, not INFERRED.
* test(cognition): update term_capture_rate baseline from 0.9167 to 1.0
unknown_logos_019 now correctly surfaces "light" as a pack-resident
token near the logos versor — producing term_capture_rate 1.0 on both
main and Phase 3. The 0.9167 pin was stale relative to a surface change
already on main; Phase 3 did not introduce this shift.
* fix(epistemic): make empty resonance evidence undetermined
* fix(evals): classify verified realizer failures separately
* fix(packs): treat absent domain manifests as valid noop
* test(packs): cover missing manifests and scope boundary domains
* test(epistemic): cover phase 2 known bug fixes
* fix(vault): make FALSIFIED exclusion explicit in _status_admits
FALSIFIED entries previously fell through to the ADMISSIBLE_AS_EVIDENCE
set-check, which excluded them correctly but left the distinction between
CONTRADICTED (FALSIFIED) and UNVERIFIED-POSSIBLE (SPECULATIVE) implicit.
Add an early guard so FALSIFIED is explicitly rejected before the tier
filter, matching the CONTRADICTED semantics from the epistemic taxonomy.
* feat(ADR-0141): multiply as CGA dilator versor (positive non-zero)
Adds `multiply(scale)` to `generate/math_versor_arithmetic.py` as the
standard CGA dilator for multiplicative scaling along e1, restricted to
`scale > 0`. All ten ADR-0141 assertion families pass.
Preliminary measurement confirmed:
N = n_o ∧ n_inf: component -1 at index 15 (blade (3,4) = e4∧e5)
N² = +1.0 (pure scalar) → closed-form D_s = cosh(α/2) + sinh(α/2)·N
n_o · n_inf = -1; n_o² = n_inf² = 0
Because N² = +1, the cosh/sinh expansion is exact in float64 and
D_s · ~D_s = cosh² − sinh² = 1 holds to machine epsilon.
The sandwich D_s·X·~D_s produces a null point with n_inf normalization
1/s. `decode_quantity` is updated to divide by that factor, recovering
value · s. For translator outputs (normalization = 1) the result is
identical to the previous direct e1 read; all 152 prior add/subtract
tests pass unchanged.
`embed_quantity` is updated to embed directly in float64, eliminating
float32 quantization error for values like 0.01 (float32(0.01) ≠ 0.01);
all prior test-case values were exactly representable in float32.
* docs(ADR-0141): add decision document for multiply-as-dilator spike
The ADR doc was drafted in a separate branch and not present when the
implementation worktree was created from origin/main. Adding it now so
the decision record lands on main with the implementation it specifies.
Content unchanged from the draft — same spec the implementation already
satisfies (10 assertion families, fixed test cases, falsification
discipline, deferred scope for negative / zero / divide / Rate).
No code or test changes in this commit.
Extends generate/math_versor_arithmetic.py with one new function:
def subtract(addend: float) -> np.ndarray:
return translator(-float(addend))
Single-line delegate to translator(); no new algebra.
Adds tests/test_arithmetic_subtract_and_group.py covering all nine
ADR-0140 acceptance families:
Families 1-6 (ADR-0139 families applied to subtract):
1. Embedding well-formedness — null cone preserved for subtract cases
2. Translator-of-negative well-formedness — versor_condition < 1e-6
3. Closure — sandwich result stays on null cone
4. Arithmetic correctness — decoded value == a − b within 1e-9
5. Replay determinism — byte-identical across runs
6. Composability — subtract(c) ∘ subtract(b) decodes to a − b − c
New group-property families (structural verification of ADR-0139 claim):
7. Inverse composition — T_{-b} * T_b = identity (max residual: 0.000e+00)
8. Round-trip closure — versor_apply(T_{-b}, versor_apply(T_b, X)) → (a, u)
9a. Sum composition — T_a * T_b = T_{a+b} (max residual: 0.000e+00)
9b. Commutativity — T_a * T_b byte-equals T_b * T_a (all 10 cases)
All 96 tests pass. Group residuals are exactly 0.0 in float64.
The additive subgroup of Cl(4,1) translators along e1 is abelian and
closed; ADR-0139's algebraic claim holds at the group level.
First step of the Engine A lift program (CLAUDE.md commits the project to a
single deterministic cognitive engine; Engine B / math pipeline was always
intentional scaffolding per math_solver.py:24). Proves the load-bearing
unknown: one arithmetic operation can be represented as a closed versor at
the required tolerance, with no new normalization and no weakened invariant.
Scope (frozen by ADR-0139):
- One operation: add
- Single-axis embedding: quantities on e1 axis
- No graph wiring, no pipeline integration, no GSM8K case routed
- Unit carried as caller metadata
Construction:
- embed_quantity(v, u) = embed_point([v, 0, 0]) (existing CGA primitive)
- translator(b) = 1 - 0.5 * (b*e1 * n_inf) (textbook CGA translator)
- decode_quantity(F, u) = (F[1], u) (e1 coordinate)
Measured values (all 11 fixed cases + composability):
a b vcond(T) |<R,R>| decode_err
0.0 0.0 0.000e+00 0.000e+00 0.000e+00
0.0 1.0 0.000e+00 0.000e+00 0.000e+00
1.0 0.0 0.000e+00 0.000e+00 0.000e+00
3.0 4.0 0.000e+00 0.000e+00 0.000e+00
7.0 -3.0 0.000e+00 0.000e+00 0.000e+00
0.25 0.75 0.000e+00 0.000e+00 0.000e+00
1.5 2.5 0.000e+00 0.000e+00 0.000e+00
-5.0 5.0 0.000e+00 0.000e+00 0.000e+00
-2.0 -3.0 0.000e+00 0.000e+00 0.000e+00
100.0 1.0 0.000e+00 0.000e+00 0.000e+00
1.0 100.0 0.000e+00 0.000e+00 0.000e+00
compose (2, 3, 5) → 10: |<R2,R2>| = 0.000e+00, decode_err = 0.000e+00
Every residual is exactly 0.0 in float64. The construction is algebraically
closed: T_t * reverse(T_t) = 1 - 0.25*B^2 where B = t*n_inf, and B^2 = 0
because (e14)^2 + (e15)^2 = -1 + 1 and cross-terms cancel. No machine-epsilon
drift accumulates because the relevant cancellation happens at the algebraic
level before float arithmetic.
ADR-0139 acceptance items 1-6 (one parametrized test family each):
1. Embedding well-formedness — test_family1_embedding_is_null (11 cases)
2. Translator well-formedness — test_family2_translator_unit_versor (11 cases)
3. Closure — test_family3_sandwich_preserves_null (11 cases)
4. Arithmetic correctness — test_family4_decode_matches_sum (11 cases)
5. Replay determinism — test_family5_replay_byte_identical (11 cases)
6. Composability — test_family6_two_translators_compose (1 case)
Total: 56 tests, all passing.
Lift program decision: proceeds. Follow-on ADRs (subtract, multiply, Rate,
compare, MathProblemGraph → PropositionGraph, pipeline integration, first
GSM8K case end-to-end through Engine A) are now justified by a concrete
algebraic foundation rather than design speculation.
Out of scope per ADR-0139:
- No modifications to algebra/, core/cognition/, chat/, math_solver.py,
math_verifier.py, math_realizer.py, math_candidate_parser.py
- No GSM8K runner changes
- No pack changes
- Engine B continues serving GSM8K unchanged; the 3/50 admission set is
preserved
CLI lanes intentionally not run — main has known test-rot orthogonal to
this PR. The 56 new tests are self-contained and the diff touches only
three new files.
* content(packs): update relations checksum
* revert transient relations manifest checksum
* content(packs): extend relations lexicon additively
* content(teaching): extend relations chains additively
* content(packs): ratify relations manifest checksum
* test(packs): accept additive relations lemma extension
* test(packs): add relations v1 extension regressions
* fix(tests): align relations extension lemma set
* content(packs): add relations mastery report
* content(packs): drop unused .mastery_report.json sidecar
Language packs do not consume mastery reports — the pattern is from
identity packs (packs/identity/) and has no consumer in language_packs/
loader.py or compiler.py. The added sidecar's self-seal hash also did
not validate against sha256(json.dumps(body, sort_keys=True,
separators=(',', ':'))).
Drop the file. The actual ratification surface for this pack is the
manifest.json lexicon_checksum, which still matches lexicon.jsonl
bytes (verified).
S.4 extends initial-state parsing with two closed subject-slot widenings:
- Indefinite-article: `A <noun> has N <unit>` (gsm8k-0046 sentence 1)
- Prepositional-prefix existential: `In a <place>, there are N <unit>...`
(gsm8k-0038 sentence 1)
Design choice: sibling regexes (_INITIAL_HAS_INDEF_RE,
_INITIAL_THERE_ARE_PREFIX_RE) rather than widening the global _ENTITY
pattern — preserves existing behavior across all other initial-state
extractors (cascade-safety).
Per the S.x corridor discipline: no new short-circuit; new candidates
flow through extract_initial_candidates and the existing graph machinery.
No solver/graph/verifier changes.
Honest delta:
- Direct admissions: 0 (admission set unchanged at {0014, 0018, 0042})
- Barrier shifts: +2 (gsm8k-0038: novel_initial_form → compound_comparative;
gsm8k-0046: novel_initial_form → fraction_operand)
- wrong == 0 on every lane
Bundled with this PR for ledger currency:
1. tests/test_rescan_v3_invariants.py refactored to read frozen on-disk
v3 artifacts only (no more re-running build_rescan against live
parser). The previous design tied a historical snapshot to live code
and broke the moment any new phase landed.
2. rescan_v4.py + refusal_rescan_v4.json + refusal_taxonomy_v4.json +
tests/test_rescan_v4_invariants.py — the current live snapshot.
Shifts: exactly 2 (0038, 0046). Same pattern as v3.
Sonnet wrote: S.4 parser/axis-lane/tests/ADR.
Opus wrote: rescan_v4.py + v3 test refactor + bundling.
Files:
- generate/math_candidate_parser.py (+142 lines)
- evals/math_capability_axes/S4_novel_initial_form/v1/ (20-case lane)
- tests/test_adr_0136_S4_novel_initial_form.py (40 tests)
- docs/decisions/ADR-0136.S.4-novel-initial-form.md
- evals/gsm8k_math/train_sample/v1/{rescan_v4.py, *_v4.json}
- tests/test_rescan_v4_invariants.py (8 tests)
- tests/test_rescan_v3_invariants.py (refactored to artifact-only)
Re-runs parse_and_solve on the 50-case GSM8K train sample on current
main (post-S.3) and compares to v2. Result: admitted=3/50 (unchanged),
wrong=0, exactly 1 barrier shifted v2→v3.
Shift: gsm8k-0010 (compound_statement → fraction_operand). S.3's
_INIT_MUTATION_RE resolves "Yun had 20 paperclips initially, but then
lost 12" to InitialPossession(Yun, 8, paperclips). First refusal moved
to sentence 2: "Marion has 1/4 more than what Yun currently has, plus
7" — needs fraction-operand + coreference-quantity + comparative-additive
arithmetic.
Top blockers (v3):
compound_statement 5 (was 6)
novel_initial_form 5 (unchanged)
fraction_operand 4 (was 3 — gsm8k-0010 moved here)
novel_initial_verb 4 (unchanged)
Artifacts:
- evals/gsm8k_math/train_sample/v1/rescan_v3.py
- evals/gsm8k_math/train_sample/v1/refusal_rescan_v3.json
- evals/gsm8k_math/train_sample/v1/refusal_taxonomy_v3.json
- docs/decisions/ADR-0136.S3-post-rescan.md
- tests/test_rescan_v3_invariants.py (7 tests; determinism + admission
set unchanged + exactly-one-shift + 0010-specific shift assertions)
Measurement-only branch. Re-runs parse_and_solve on all 50 GSM8K train-sample
cases against the current parser (post-S.1/S.2) and produces a barrier-shift
ledger comparing v1 taxonomy to current behavior.
Results: admitted=3/50 (0014, 0018, 0042), wrong=0, barrier_shifted=27/50.
Context-filler dominance collapsed from 23→3 cases; compound_statement (6)
and novel_initial_form (5) are now the largest buckets.
Subsumption directive pinned: ADR-0137 SHALL re-derive all short-circuit
admissions as (DeferredCandidate, evidence, BindingProof) triples.
- Add classify_sentence() + has_numeric_token() to math_candidate_parser.py.
Rule: sentence with no digit and no word-number cannot introduce parseable
numeric state — classify as "context" and skip safely (wrong==0 preserved).
- Add pre-pass in parse_and_solve() (math_candidate_graph.py): strips context
sentences before extraction; falls through to refusal if none remain numeric.
- Extend capacity patterns for gsm8k-0018:
- _CAPACITY_INVERTED_RE: "During M <time-unit> <Actor> can <verb> N <unit>"
- _CAPACITY_Q2_RE: "How many <unit> [on average] is <Actor> able to <verb>,
when the <event> lasted for T <time-unit>?"
- GSM8K: 1/50 -> 2/50 (gsm8k-0018 admits with answer 16.0); admitted_wrong==0.
- Tests: 47/47 pass (12 new for classifier, inverted patterns, 0018 end-to-end).
Rebases onto current main (dec98ea, post-G.1/G.3.1/G.4/promotion).
Parser:
- Extend _COMPARE_MULT_ANCHOR_RE anchor alternation to include 'quarter'
and 'third'; add optional 'a\s+' article prefix so "a quarter as many"
and "a third as many" parse. Both anchors are in COMPARE_MULTIPLICATIVE_ANCHORS
and the round-trip factor-divisor table ("quarter":4, "third":3), so
round-trip checks pass. quarter→0.25 (exact), third→1/3 (float).
- Add _ANCHOR_TO_FACTOR entries for quarter and third.
Gate regex (test_adr_0131_G2_comparatives.py):
- Widen _COMPARATIVE_STATEMENT_PATTERNS multiplicative pattern from
'\d+\s+times' to '\w+\s+times' to match word-number forms ("four times")
that would be missed by the digit-only pattern if a future GSM8K case
contains one in a still-refused statement.
Cases (31 total, was 24):
- G2-mul-frac-005/006: two 'quarter' cases (fraction direction now has
half×4 + quarter×2 + third×1 = 7 cases, was 4 all-half).
- G2-mul-frac-007: 'third' case.
- G2-refuse-006: hyphenated 'one-third' pins the closed-anchor boundary.
- G2-refuse-007: 'double as many' pins the deferred grammar shape.
Tests (25, was 21):
- Add quarter and third parametric entries to test_multiplicative_direction_admits.
- Add one-third and double-as-many refusal params to test_refusal_cases.
- Add quarter/third to test_direction_literals_closed_set.
- Update test_runner_per_category_minima comment to reflect new counts.
ADR: document quarter/third admission, updated case table, deferred list.
report.json: refreshed to 31 cases, wrong==0 preserved.
Bundles the three pieces needed to consummate the promotion after
the reviewer signature lands:
1. Wire the expert tier in the capability ledger
2. Path-stability fix (digest filesystem-independence)
3. Reviewer-registry allow-list extension (regression fix for #194)
Result: mathematics_logic is now the first expert-tier domain in
the capability ledger.
$ ledger_report() -> mathematics_logic row:
status: "expert"
predicates: { seeded, grounded, reasoning_capable,
audit_passed, expert: True }
expert_reason: "ADR-0120-math composer admitted"
1. Ledger wiring (core/capability/reporting.py):
- _EXPERT_DOMAIN_STATUSES extends to 6 tiers with "expert"
after "audit-passed" (strict super-tier).
- New _EXPERT_COMPOSERS dict — per-domain registry of composer
module names. Currently only mathematics_logic ->
core.capability.expert_promotion_math.
- New `expert` predicate computation gated on audit_passed;
calls registered composer's evaluate_math_expert_promotion()
and reads promote_admitted as the verdict. Fail-closed on
exception or missing composer.
- status = "expert" when predicate True.
- predicates dict gains "expert" key; row gains expert_reason.
2. Path-stability fix (composite_math_gate.py + expert_promotion_math.py):
- New _rel(path) helpers return repo-root-relative POSIX
strings instead of str(absolute_path).
- claim_digest now commits to relative paths, so operator A
on ~/work/core and operator B on /srv/checkouts/core compute
the SAME digest for identical evidence.
- Without this fix no signature would ever match across
filesystems — a real bug that would have blocked every
signing attempt.
3. Allow-list regression fix (core/capability/reviewers.py):
- ALLOWED_TOP_LEVEL_KEYS extended with "math_expert_claims".
- PR #194 added the section to docs/reviewers.yaml but didn't
extend the allow-list, silently breaking the audit_passed
predicate for ALL 3 prior domains (loader rejected the file).
This PR's test_allowed_top_level_keys_includes_math_expert_claims
regression-pins the fix.
Reviewer signature (operator-only action by shay-j) carried in
docs/reviewers.yaml:
math_expert_claims:
- domain_id: mathematics_logic
signed_by: shay-j
claim_digest: "94149794e8c19896851e062cf1f921cfa9ba04770b674bc3b4c33023f7c7331b"
The auto-mode safeguard correctly blocked the agent from self-
signing during PR construction; the signature was performed by the
reviewer directly and brought into this PR. Future signatures stay
human-only.
Tests: 12/12 new ledger-flip tests + 174/174 across full obligation
auditor / composer / composite-gate / expert-demo / reviewer-registry
regression. Updated #194's awaiting-state snapshot to reflect the new
promote_admitted=True state on main.
GSM8K (honest disclosure, not gating): still 0/50 admission, wrong=0,
safety_rail_intact=True, substrate=candidate_graph. Probe lift is
future work (bounded pronoun coref is the highest-leverage item —
~28% of refusals route through it). The promotion does not depend
on GSM8K per ADR-0131.
Final wire-up after all 10 ADR-0114a obligations + ADR-0131.4
composite gate landed. Composes:
- all 10 obligation verdicts (5 from new auditor modules,
5 from inline checks over existing infrastructure)
- ADR-0131.4 composite math gate verdict
- ADR-0092 reviewer-signed claim entry from docs/reviewers.yaml
into a single deterministic promotion verdict + canonical
signed/unsigned ``expert_claims_math_v1_signed.json`` artifact.
Empirical verdict on current main (first evaluation):
all_obligations_passed: True
composite_gate_passed: True
technical_pass: True
claim_digest: d164866975341d9b82503caf50c0404ee140eab21fd60f589536c6daf6e1d706
reviewer_signature_present: False
promote_admitted: False
refusal_reason: awaiting reviewer signature
Every technical gate passes. The PR ships in the architecturally-
correct "awaiting reviewer signature" state — the reviewer's
signature is the separate, auditable operator action that
consummates the promotion.
Operator workflow (post-merge):
1. Run `core capability math-expert-promote`, confirm verdict,
capture claim_digest.
2. Add entry to docs/reviewers.yaml under math_expert_claims:
- domain_id: mathematics_logic
signed_by: shay-j
claim_digest: "d164866975341d9b82503caf50c0404ee140eab21fd60f589536c6daf6e1d706"
3. Re-run — promote_admitted flips to True.
4. Separate ledger-flip PR (out of scope here) consumes the
signed artifact and writes the capability ledger.
Safety property: if the evidence bundle changes after signing
(B-lane re-run, pack edit, obligation report shift), the digest
changes and the existing signature stops matching. The verdict
reports the mismatch explicitly and the operator must re-inspect
and re-sign — a ledger flip can't survive a silent evidence change.
New files:
- core/capability/expert_promotion_math.py — the composer
- tests/test_adr_0120_math_expert_promotion.py — 18 tests
- docs/decisions/ADR-0120-math-expert-promotion-wireup.md — ADR
Modified:
- core/cli.py — new `core capability math-expert-promote` cmd
- docs/reviewers.yaml — added math_expert_claims: [] section
with documentation comment
Tests: 18/18 covering each inline obligation evaluator
(#1/#3/#4/#7/#9 pass + failure modes), composer integration
against current main, reviewer-signature path (matching → admitted;
mismatched → refused with explicit diagnostic), digest
reproducibility, artifact byte-equality. All pass in 0.49s.
Trust boundary: read-only access to 4 B-lane reports +
GSM8K probe + 5 obligation auditor reports (transitively) +
frontier dir + docs/reviewers.yaml; single deterministic write
to the artifact path; no dynamic imports, no shell, no network.
This is the last PR before the first mathematics_logic -> expert
ledger flip attempt. The actual flip is reserved for a separate
small PR that consumes the signed artifact.
35-case OOD set (ood-001..ood-035): surface-varied siblings of B3's 35
solved_correct public cases. Entity-name pool: Maya/Liam/Noah/Diana/Felix/
Priya/Omar/Rosa/Jun/Kai. Unit-noun pool: oranges/marbles/pencils/books/
stamps/coins/balls (all parser-allowed count nouns). Every case in-grammar
per ADR-0131.3 and parseable without error.
Auditor (core/capability/ood_ratio.py): reads B3 public report.json + OOD
report.json, computes ood_ratio = ood_accuracy / public_accuracy, enforces
two independent gates — ratio ≥ 0.95 and wrong == 0.
CLI: core capability ood-ratio (exit 0 iff both gates pass).
Measured: public 50/50=1.000, OOD 35/35=1.000, ratio=1.000. Obligation #10
and B3 public lane unchanged.
Implements the external auditor for ADR-0114a Obligation #6:
"depth_curve.py produces a per-bucket curve;
accuracy(N) >= accuracy(depth_1) * (1 - eps)^(N - 1) for eps = 0.05."
Mirrors PR #189's auditor pattern (re-runs lane via the candidate-
graph pipeline, aggregates over committed cases, emits deterministic
report). Uses len(trace.steps) as the authoritative depth — the
engine's actually-executed reasoning, not the case's declared depth.
New module core/capability/depth_curve.py:
- Bucket schema mirrors ADR-0119.6: depth_1, depth_2-3,
depth_4-5, depth_6-8. Depth > 8 raises rather than silently
extending. Depth == 0 (initial-only problems) skipped — nothing
to decay.
- representative_depth = min(bucket) — most permissive bound
convention; tightening requires an ADR amendment.
- epsilon = 0.05 pinned per ADR-0120 §Threshold rationale.
- Two-axis verdict: obligation_6_mechanism_wired (always true if
auditor ran), obligation_6_assertion_holds (every populated
bucket satisfies the decay bound), coverage_sufficient (>=2
buckets populated AND >=3 cases each — required for the
assertion to be statistically meaningful).
CLI: core capability depth-curve (added to core/cli.py).
Writes evals/obligation_6_depth_curve/<lane_id>.json.
Empirical verdict on current main:
lane: B3_bounded_grammar
cases_total: 50
cases_solved: 22
mechanism_wired: True
assertion_holds: True
coverage_sufficient: False
populated: [depth_1 (21/21=1.0000), depth_2-3 (1/1=1.0000)]
Both populated buckets satisfy the decay bound. Coverage gap is
honestly named in the refusal_reason: depth_2-3 has only 1 case,
depth_4-5 and depth_6-8 have none. This is B3-owner work (case
authoring under the existing grammar contract), not auditor work;
reserved as a B3 v1.1 follow-up PR.
Honest scope-limit: B3 only. B1 (algebra, no trace) and B2 (chain
validation, not problem-solving) need different metrics — separate
sub-ADRs.
Trust boundary: read-only access to B3 cases + transitive pack
reads via the pipeline; single deterministic write to artifact path.
Tests: 24/24 covering bucket schema closure (depth 1..8 + raise on
9+), decay bound math (epsilon pinned, formula correct, depth_1 has
no bound), coverage-sufficient policy (thresholds pinned), lane
evaluation (passes on real B3 + refuses on missing cases),
coverage-sufficient distinction (B3 today vs synthetic 5+5 fixture
showing both pass), determinism (report identical + artifact
byte-equal).
External auditor for ADR-0114a Obligation #8:
"adversarial/score.py reports wrong == 0 across all families;
>= 30 cases x >= 8 families."
Verdict on current main:
cases_total: 36
families_total: 9
cases_refused: 28
cases_solved: 8
cases_wrong: 0 <-- the gate
obligation_8_passed: True
New module core/capability/adversarial.py mirrors PR #189/#190/#191
auditor pattern. Pure function over the committed cases set; broad
exception capture (correctly classified as refused — engine
couldn't process the input) makes the auditor robust to upstream
typed-refusal gaps.
New dataset evals/obligation_8_adversarial/v1/cases.jsonl — 36
cases x 9 families, closed taxonomy:
- paraphrase (verb outside initial-anchor whitelist)
- unrecognized_unit (not in en_units_v1)
- conditional (if/would/suppose)
- pronoun_coref (cross-sentence he/she/they)
- hedged_quantity (about/almost/approximately)
- ordinal_confusion (the 5th/third in cardinal position)
- implicit_subject (no named entity)
- self_reference (actor as comparison ref or transfer target)
- distractor_noise (adjectival/temporal/irrelevant siblings)
CLI: core capability adversarial. Writes
evals/obligation_8_adversarial/<lane_id>.json. Exit 0 iff
obligation passes.
Honest disclosure — 8 of 36 cases solved rather than refused;
none produced wrong answers. Two parser-layer gaps surfaced:
Gap A (pronoun_coref, 4/4 solved): unbound sibling sentences
silently drop; engine returns last-asserted state. Faithful but
semantically poor. Reserved follow-up: tighten admissibility so
unbound sentences refuse the whole case.
Gap B (unrecognized_unit, 4/4 solved): _canonicalize_unit
falls back to '+s' plural rule when pack doesn't recognize
the unit. Reserved follow-up: opt-in strict mode behind a flag
(some B3 units aren't in en_units_v1 either; strict mode
requires parallel pack extension).
Bug caught: adv-self-reference-003 ("Sam gives 3 apples to
Sam.") raises uncaught MathGraphError from
Operation.__post_init__. Auditor catches it as
refused-via-exception; ~3-line follow-up in
_build_op_candidate fixes the parser side.
Trust boundary: read-only access to cases + transitive pack reads;
single deterministic write to artifact path.
Tests: 11/11 in tests/test_adr_0114a_8_adversarial.py covering
threshold pinning (>= 30 cases / >= 8 families), closed taxonomy
(every documented family has cases; no unknown families),
obligation-passes snapshot, per-family wrong=0 invariant, failure
modes (missing file, below-threshold count), determinism (report
identical + artifact byte-equal).
Implements the external auditor ADR-0114a Obligation #10 requires:
"Every SolutionTrace.steps[*].pack_lemma_id resolves to a real
lexicon entry in the domain's operator pack." The solver enforces
this at solve time; this PR audits it from outside.
New module core/capability/pack_provenance.py:
- _load_lexicon_lemmas(): independent re-read of pack lexicon
- _parse_lemma_id(): <pack_id>:<lemma> shape parser
- validate_lane(): re-runs candidate-graph pipeline on a B-lane's
cases, walks every solver step, validates pack_lemma_id parses
AND resolves to a lexicon entry. Per-case + per-lane verdict.
- emit_provenance_report(): deterministic artifact emission.
CLI: core capability pack-provenance (added to core/cli.py).
Writes evals/obligation_10_pack_provenance/<lane_id>.json.
Empirical verdict on current main (post-PR #186):
lane: B3_bounded_grammar
cases_total: 50
cases_validated: 25 (every expected-correct B3 case)
cases_skipped_unsolved: 25 (refusal-expected probes — by design)
cases_violated: 0
obligation_10_passed: True
5 distinct lemma_ids observed (add, subtract, transfer,
compare_additive, compare_multiplicative) — all resolve to
en_arithmetic_v1. The other 3 op kinds (multiply, divide,
apply_rate) ratify-at-solve-time via _resolve_pack_lemmas so the
obligation holds for them too if a future case exercises them.
Honest scope-limit: B3 only. B1 (symbolic equivalence) and B2
(teaching corpus) equivalents deferred to separate sub-ADRs —
B1 needs reframing (algebra normalization chain, not arithmetic
steps); B2 can use this same auditor signature once corpus
solver-trace exercise is confirmed case-by-case.
Composition with ADR-0131.4: orthogonal. Composite gate verdict
+ obligation #10 verdict + 4 other obligation auditors (when
they land) + reviewer signature → full ADR-0120 wire-up.
Trust boundary: read-only access to pack lexicon + B3 cases;
single deterministic write to artifact path. No dynamic imports,
no shell passthrough, no network. Pure deterministic auditor.
Tests: 19/19 in tests/test_adr_0114a_10_pack_provenance.py
covering lemma-id parser (well-formed + malformed), lexicon loader
(real pack + every failure mode), lane validator (passes on real
B3 + refuses on missing pack/cases + skips refusal-expected cases
without false violation), determinism (report identical across
calls + artifact byte-equal).
Cognitive capability: extend bounded grammar to admit acquisition/action
verbs (buys, bought, collected, saved, saved-up, makes, sells) as
operation-kind entries, and pure-possession verbs (had, started, started-with)
as initial-possession anchors.
What invariant proves correctness:
- wrong == 0 across all G1 curated cases (20/20) and GSM8K probe (0 wrong/50).
- versor_condition and field invariants untouched — no algebra-path changes.
- Round-trip filter (math_roundtrip.roundtrip_admissible) unchanged.
Which CLI suite / eval proves the lane:
pytest tests/test_adr_0131_G1_verb_classes.py — 15/15 pass
pytest tests/test_adr_0126_runner_wiring.py — 9/9 pass (3 regressions fixed)
pytest tests/test_adr_0131_{1,3}_*lane.py — 17/17 pass
pytest tests/test_adr_0131_G_gsm8k_coverage_probe.py — 8/8 pass
pytest tests/test_gsm8k_math_runner.py — 11/11 pass
Key architectural change:
Acquisition verbs that also appear in ADD_VERBS/SUBTRACT_VERBS were
previously listed in _INITIAL_HAS_RE, causing branch-disagreement refusals
when a canonical 'has' initial preceded an acquisition sentence for the
same entity. Fix: narrow _INITIAL_HAS_RE to pure-possession anchors only
(has/have/had/started); acquisition verbs remain exclusively in KIND_TO_VERBS.
The solver's default-from-zero means 'Sam buys 5 apples. How many does
Sam have?' resolves as 0+5=5 without any initial-possession candidate.
Optional verb particle (up/down/out/...) added to _op_pattern to handle
'saved up N', 'picked up N' etc.
No changes to binding graph, solver, verifier, or versor/CGA algebra.
No stochastic generation, approximate recall, or hidden normalization.
Trust boundaries unaffected — no new dynamic imports or user-input paths.
Implements ADR-0131's revision of the ADR-0120 expert-promotion
contract for mathematics_logic: replaces the single-benchmark
GSM8K-coverage check with a composite B1+B2+B3 requirement.
New module core/capability/composite_math_gate.py:
- evaluate_composite_math_gate(): pure function over already-
committed B-lane reports; handles heterogeneous report shapes
(B1/B2 counts vs B3 metrics); applies pinned thresholds
(correct_rate >= 0.95 AND wrong == 0); composes verdicts.
- Reproducible SHA-256 claim_digest over canonical evidence bundle.
- GSM8K honest-disclosure (admission/wrong/refused/substrate)
embedded in artifact but never gates per ADR-0131.
CLI: core capability math-expert-gate (added to core/cli.py).
Writes evals/math_expert_claims/v1/expert_claims_math_v1.json.
Empirical verdict on current main (post-PR #182/#183/#184/#185):
composite_gate_passed: True
B1_public: 185/185 wrong=0 rate=1.0000
B1_sealed: 14/14 wrong=0 rate=1.0000
B2_teaching_corpus: 40/40 wrong=0 rate=1.0000
B3_bounded_grammar: 50/50 wrong=0 rate=1.0000
GSM8K disclosure: 0/50 admission, wrong=0, substrate=candidate_graph
The math expert is gate-passing under ADR-0131's revised composite
contract. The architectural bet ADR-0131 placed has paid off.
Honest scope-limit: this implements only the ADR-0131-specific
revision (composite benchmark portion). The full ADR-0120 10-
obligation contract still requires substrate for 5 missing
obligations (OOD ratio, perturbation, depth curve, adversarial,
operation-provenance-via-pack). Those are sequencing-wise *after*
ADR-0131.4, not bundled. Reviewer signature via ADR-0092 registry
is also reserved.
Trust boundary: read-only access to 5 committed lane reports;
single deterministic write to the artifact path. No dynamic
imports, no recomputation of lane verdicts.
Tests: 12/12 in tests/test_adr_0131_4_composite_math_gate.py
covering threshold pinning, heterogeneous shape handling, gate
logic (passing + every failure mode), GSM8K honest disclosure
(never gates), determinism (claim_digest + artifact byte-equality),
and a snapshot test confirming current main satisfies the gate.
ADR-0131.4 module note: the parent ADR-0131 plan named
formation/ratify.py + formation/promote.py as the wire-up site —
that was a misidentification (those govern teaching-example
SPECULATIVE→COHERENT bridging per ADR-0021, not domain-tier
promotion). Correct site is core/capability/, where audit-passed
gate already lives.
Four axes deferred from ADR-0131.G.3 (PR #183):
1. Fractions end-to-end: new _INITIAL_FRACTION_OF_RE extractor handles
`N/M of [a/an] <unit>` shape; _resolve_value already handles N/M arithmetic.
2. Multi-currency: _MONEY_SYMBOL widened to six symbols; _CURRENCY_SYMBOLS table
+ _resolve_currency dispatcher; ¢/€/¥/₱ wired end-to-end. £/pound sterling
deferred to G.3.2 (question extractor's single-token unit slot cannot parse
two-word surface "pounds sterling").
3. Multi-token cardinals: dedicated _MULTI_WORD_CARDINAL_RE extractor (approach a)
delegates to parse_compound_cardinal; avoids greedy unit-slot boundary ambiguity
from widening _VALUE.
4. Word-num-adjective: optional adjective group added to _INITIAL_HAS_RE and
_MULTI_WORD_CARDINAL_RE; closed adjective list identical to _CONJ_OBJECT_RE.
Also fixes six pre-existing G4 type bugs where _resolve_value() result was used
directly as a numeric operand (TypeError: _ResolvedValue is not a number).
Axis lane v1_1: 20/20 solved_correct, 0 wrong, 8/8 refusals, overall_pass=True.
GSM8K probe: 0/50 admission_rate unchanged, admitted_wrong=0 (safety rail intact).
42/42 new tests pass; parent v1 lane (26/26) unaffected.
Highest-risk axis of the ADR-0131.G capability iteration: within-
sentence multi-clause composition. Four extractors land in the
candidate-emitting parser; no graph-side or solver changes.
Parser extension (generate/math_candidate_parser.py)
- _conj_subject_each_candidates: '<A> and [his/her/their <kin>] <B>
each <verb> <N> <unit>' → 2 CandidateInitial (one per actor).
- _conj_object_candidates: '<E> has <N1> <unit1> and <N2> <unit2>' →
2 CandidateInitial for the same entity; same-unit conjuncts refuse
(would silently collide under solver overwrite-on-collision).
- _embedded_quantifier_candidates: '<E> has <N> <container> with <M>
<unit> in each [<container>]' → 1 derived CandidateInitial
(value=N*M).
- _embedded_quantifier_candidates (conj branch): '... <N1> <C> with
<M1> <U> in each ... and <N2> <C> with <M2> <U> in each ...' → 1
SUM CandidateInitial (value=N1*M1+N2*M2); mixed-unit refuses.
- CandidateInitial anchor whitelist widened to include
saved/earned/got/received/bought/made/paid (and inflections) —
narrow widening needed for the conjoined-subject-each shape.
Closed-set discipline
- Distributive 'each' only — 'each ... together/altogether' refuses.
- Two-way conjunction only — 3-way refuses by non-match.
- Cross-sentence coreference stays refused (within-sentence axis).
- Ambiguous 'each' scope refuses (container2 must agree).
Curated axis lane (32 cases)
- evals/math_capability_axes/G4_multi_clause/v1/cases.jsonl:
conj_subject_each ×6, conj_object ×6, embedded_quantifier ×6,
conj_embedded ×6, refusal ×8.
- evals/math_capability_axes/G4_multi_clause/v1/runner.py +
report.json: deterministic; wrong==0 gate; byte-equal across runs.
Tests (26 new)
- tests/test_adr_0131_G4_multi_clause.py: per-shape emission,
refusal probes (parametric), distributive-only policy,
cross-sentence refusal, runner byte-equality, GSM8K-probe gate.
GSM8K-probe gate (chosen: multi-clause refusals ↓)
- evals/gsm8k_math/train_sample/v1/report.json (candidate-graph
probe): multi-clause statement-refusal count 2 → 1. Case 0042
('Ella has 4 bags with 20 apples in each bag and six bags with 25
apples in each bag.') moves from statement-clause refusal to
question-layer refusal. Case 0026 ('Aaron and his brother Carson
each saved up $40') stays refused on the '$' value slot
(deferred to G.3 numeric-literals axis).
- evals/gsm8k_math/train_sample/v1/train_sample_coverage_report.json
(legacy probe): refreshed, byte-identical (legacy parser
untouched).
B3 + candidate-graph + GSM8K probe lanes all pass (95/95
regression). wrong==0 preserved everywhere — load-bearing for the
highest-risk axis.
First capability-axis iteration after ADR-0131.G baseline. Extends the
candidate-graph parser's <value> slot to recognize:
- Money symbol literals: $N and $N.NN (1-2 decimals); $N.NNN refused
- Money word forms: N dollars / N cents
- Hyphenated multi-word cardinals: twenty-five, ninety-nine, ...
All money values normalize to integer cents, unit 'cents' — pack-aligned
with en_units_v1's canonical_unit='cent' for the money dimension.
en_numerics_v1's parse_compound_cardinal handles hyphenated cardinals.
Parser changes (generate/):
- math_candidate_parser.py: _VALUE alternation widened; _resolve_value
refactored to return _ResolvedValue|None carrying optional unit
override; _INITIAL_HAS_RE unit slot made optional; dollar/dollars →
cents normalization at candidate build.
- math_roundtrip.py: new _unit_grounds helper (money-aware); _value_grounds
widened for the three new literal shapes; roundtrip_admissible uses
_unit_grounds for the unit check.
- math_candidate_graph.py: _initial_admissible and _question_admissible
use _unit_grounds.
New axis lane (evals/math_capability_axes/G3_numerics/v1/):
- 26 curated cases (20 positive across 4 classes + 6 refusal probes)
- runner.py wraps _score_one_candidate_graph; byte-equal report.json
- 20/20 positive solved correct; 6/6 refusal probes refused typed;
solved_wrong == 0; overall_pass == True
Tests: 27/27 in 0.19s. 420 existing candidate-parser/math-parser/pack
tests still green. GSM8K probe safety rail (admitted_wrong == 0)
preserved.
Honest scope-limit (documented in ADR): admission_rate on the GSM8K
probe stays at 0/50 because (a) the probe currently consults the legacy
parser path, not the candidate-graph pipeline G.3 extends, and (b) most
money-bearing GSM8K cases fail first on verb (G.1) or multi-clause (G.4)
shape, not on the money literal. The axis lane is the load-bearing
measurement for this iteration. Reserved follow-up: a small probe-
infra ADR to switch run_coverage_probe.py to the candidate-graph
pipeline.
Out of scope, deferred to G.3.1: fractions end-to-end (resolver supports
N/M but no axis cases), multi-currency (¢ € £ ¥ ₱), space-separated
multi-word cardinals (one hundred), word-number-adjective compositions
(five full boxes).
Wire compare_additive / compare_multiplicative extractors into the
candidate-emitting sentence parser, closing the deferred phase flagged
at generate/math_candidate_parser.py:30.
Capability axis: comparatives (additive + multiplicative)
- generate/math_candidate_parser.py: new _compare_additive_candidates,
_compare_multiplicative_candidates, _compare_nested_candidates
emitting CandidateOperation records keyed to the four
Comparison.direction literals registered in ADR-0123.
- Closed-set anchor alternation; 'less' admitted as surface synonym of
'fewer'; reference slot widened to admit "the number/amount of <unit>"
for nested forms.
- Nested 'A has N more <unit> than M times <REF>' emits two flat
candidates (additive + multiplicative); binding-graph picks the
admissible composition or refuses (no solver stub).
Curated axis lane (24 cases)
- evals/math_capability_axes/G2_comparatives/v1/cases.jsonl:
8 additive / 8 multiplicative / 3 nested / 5 refusal
- evals/math_capability_axes/G2_comparatives/v1/runner.py +
report.json: deterministic, wrong==0 gate, byte-equal across runs.
Tests (21 new)
- tests/test_adr_0131_G2_comparatives.py: per-direction at-least-one
passing, nested-both-emitted, closed-set refusal, runner
byte-equality, GSM8K-probe gate (comparative-clause refusals
strictly decrease).
GSM8K-probe gate (chosen: comparative-clause refusals ↓)
- evals/gsm8k_math/train_sample/v1/report.json (candidate-graph
probe): comparative-clause refusal count 2 → 1 (case 0009 'Jen has
10 more ducks than four times the number of chickens' moves from
statement-clause refusal to question-layer refusal). admitted_wrong
remains 0; admission_rate unchanged (downstream composition is a
follow-up ADR).
- evals/gsm8k_math/train_sample/v1/train_sample_coverage_report.json
(legacy probe): refreshed, byte-identical (legacy parser untouched).
B3 + candidate-graph + GSM8K probe lanes all pass (90/90). Direction
vocab stays closed to {more, fewer, times, fraction}; wrong==0
preserved everywhere.
ADR-0131 deferred GSM8K because it rewards paraphrase flexibility,
which is the deterministic engine's structural weakness. This ADR
re-engages it on architecture-aligned terms: as a *coverage probe*
of the bounded grammar + binding graph, not a promotion gate.
The framing pinned by this ADR:
GSM8K is not a target. The model's capability is the target.
GSM8K passing is the symptom of capability, not the goal of
the work.
Wrong mindset (rejected by ADR's iteration discipline):
"Find templates that admit more GSM8K cases."
Right mindset (load-bearing):
"Extend the model's NL-to-typed-graph capability along
principled axes (verb classes, comparative structures, numeric
forms, multi-clause grammar). GSM8K admission rises as a
side effect alongside every other word-problem corpus."
Baseline pinned by this commit:
admission_rate: 0/50 = 0.0%
admitted_wrong: 0 (gate intact, safety rail bulletproof)
refused: 50/50 = 100.0%
Every refusal is a typed parser error citing the specific clause
that did not match a template. Zero crashes, zero confabulations
— refusal-first works perfectly at admission rate zero.
What's in this PR:
- ``docs/decisions/ADR-0131.G-gsm8k-coverage-probe.md``: the ADR.
Cites parents (ADR-0131, -0115/-0116/-0117, -0131.3, -0132..-0135).
Documents the capability-first iteration discipline that every
subsequent ADR-0131.G.<n> must follow:
1. Name a single capability axis the iteration extends
2. Add B3-style curated coverage cases (capability proves
itself OUTSIDE GSM8K)
3. Re-run both B3 lane + GSM8K probe; B3 must not regress
4. Reject any expansion that only moves GSM8K admission
- ``evals/gsm8k_math/train_sample/v1/run_coverage_probe.py``:
pure-adapter wrapper around the existing run_lane. Emits a
deterministic train_sample_coverage_report.json with metrics,
per-case outcomes, and the top refused-reason families (the
work queue for capability extension).
- ``evals/gsm8k_math/train_sample/v1/train_sample_coverage_report.json``:
the baseline report. Diff-able artifact every future iteration
moves.
- ``tests/test_adr_0131_G_gsm8k_coverage_probe.py``: 8 contract
tests pinning the safety rail (admitted_wrong == 0), typed
refusal invariant (every refused case has non-empty reason),
closed outcome vocabulary, deterministic replay, committed-
report matches fresh-run.
The promotion-gate composite (B1 + B2 + B3) is unaffected.
ADR-0131.4 still consumes those three. The GSM8K probe is
empirical context for honest external claims, not a gate.
* feat(ADR-0131.1.F): frontier-baseline comparison harness for B1
Adapts the ADR-0119.4 methodology (frozen citations + comparison JSON
with disclaimer) to B1, with three additions for the
architecture-aligned claim:
1. A provider-agnostic live head-to-head runner. Adapters for
Anthropic / OpenAI / Google import their SDKs lazily so the
package loads cleanly without them installed. Each provider has a
documented FRONTIER_<VENDOR>_KEY env var; the runner refuses with
a typed FrontierRunError when keys are absent and the cache cannot
cover all cases. Every response is cached one-record-per-line at
responses/<provider>/<model>.jsonl so subsequent runs replay
byte-equally without re-calling the API.
2. A conservative free-text-to-closed-vocab verdict parser. Ambiguous
or sentinel-free provider replies collapse to "refused" — a
polarized verdict is never confabulated from prose. Chain-of-
thought replies use last-token-wins (provider deliberates, then
concludes). This is the load-bearing seam that prevents the
runner from manufacturing scores the provider didn't deliver.
3. Architecture-aligned comparison metrics. accuracy is reported but
foregrounded as the least-load-bearing; refusal_correctness
(CORE 100% by lane-gate construction vs. frontier confabulation
rate) and determinism (CORE byte-equal vs. frontier variance) are
the differentiators.
Frozen adjacent-benchmark citations cover Anthropic
(claude-3-5-sonnet on MATH, claude-opus-4-1 on AIME), OpenAI
(gpt-4o on MATH), and Google (gemini-1.5-pro on MATH). The scope
disclaimer documents that these are adjacent, not head-to-head.
Head-to-head numbers, when run, land in the cache; the comparison
JSON joins them with CORE's existing lane result.
22 tests pin the methodology: citation shape (every field, https
URL, YYYY-MM-DD date), provider-registry shape, verdict-parser
conservatism (multiple chain-of-thought cases), runner caching
behavior (no double-invoke), comparison-JSON determinism (byte-equal
across runs).
No live API call at test time. The harness gates real runs behind
explicit env vars + CLI invocation.
Composes with ADR-0131.1 (B1 v1), ADR-0131.1.B (v1.B hardening,
#169), ADR-0131.1.S (sealed holdout, #173).
* feat(ADR-0131.1.F): live head-to-head — anthropic/claude-sonnet-4-6
First real frontier baseline on the full B1.B 185-case set
(curated + generated). Cached one-record-per-line at
responses/anthropic/claude-sonnet-4-6.jsonl. Re-runs replay from
disk; no further API calls.
Headline (after scoring fix):
CORE 185/185 = 100.0% accuracy
3/3 = 100.0% refusal_correctness
deterministic (byte-equal across runs)
anthropic/claude-sonnet-4-6 182/185 = 98.4% accuracy
1/3 = 33.3% refusal_correctness
non-deterministic (temperature=0, but
not byte-equal architecturally)
The 1.6pp accuracy gap is informative; the refusal-correctness gap
is the architecture-aligned story. Sonnet's three misses:
sym-eq-v1-0016 [difference_of_squares]
(x^2 + 1)*(x^2 - 1) vs x^4 - 1
Sonnet: NOT_EQUIVALENT (math error on a textbook identity)
sym-eq-gen-v1-0153 [generated_refusal_function]
sin(x) vs x
Sonnet: NOT_EQUIVALENT (confabulated — should refuse,
transcendental outside polynomial scope)
sym-eq-gen-v1-0154 [generated_refusal_negative_exponent]
x^-1 vs 1
Sonnet: NOT_EQUIVALENT (confabulated — should refuse,
negative exponent outside scope)
Sonnet correctly refused only on syntactically malformed input
("x +"); on syntactically-valid-but-semantically-out-of-scope inputs
it confidently polarized rather than refusing. CORE refuses both
classes with typed reasons.
Scoring fix: comparison.py now composes curated + generated cases
(mirroring runner.py) so the head-to-head scores the full 185-case
lane, not just the 30 curated. The initial run scored only 30/185
because the generated set was not loaded into _load_cases().
22/22 frontier-methodology tests still pass.
* feat(ADR-0131.1.F): three more head-to-head runs + Ollama adapter
Three additional providers ran against the full B1.B 185-case set,
joining the prior claude-sonnet-4-6 result:
CORE 185/185 = 100.0% acc | 3/3 = 100% refusal | 33 ms
claude-sonnet-4-6 182/185 = 98.4% acc | 1/3 = 33.3% refusal | 294 s
claude-opus-4-7 178/185 = 96.2% acc | 1/3 = 33.3% refusal | 309 s
gpt-5 134/185 = 72.4% acc | 1/3 = 33.3% refusal | 1153 s
qwen3:8b (M1 local, partial) 91/91 = 100.0% acc | n/a no refusal-class | killed
CORE is the only system at 100% on both axes, and runs ~9,000×
faster than the cheapest cloud frontier, ~35,000× faster than gpt-5,
and finishes in less wall time than a single API call to any of the
three frontier models.
Three distinct frontier brittleness modes, all rooted in
"not actually canonicalizing":
- sonnet-4-6 confabulates polarized verdicts on out-of-scope
inputs (sin(x), x^-1). Misses one in-scope difference-of-squares
identity (x^2+1)*(x^2-1) vs x^4-1.
- opus-4-7 pattern-shortcuts five near-miss-constant cases —
accepts (-x+3)*(4x+1) == -4x^2+11x+4 (correct constant is 3,
not 4) without expanding. Same two out-of-scope confabulations
as sonnet.
- gpt-5 over-refuses 50 in-scope cases — literally replies
"REFUSED" to x*(x+1) == x^2+x and (x+1)*(x-1) == x^2-1. Same
two out-of-scope confabulations as sonnet/opus.
The qwen3:8b partial is the surprise: on the 91 in-scope cases it
completed (spanning the categories where the frontier models failed),
it scored 100%. Refusal-class cases weren't reached before the run
was killed for being impractically slow (~22s/case on M1).
Changes in this commit:
- frontier_runner.py: anthropic adapter now omits ``temperature``
for claude-opus-4-x (the parameter is rejected by 4.x models);
openai adapter switches to ``max_completion_tokens`` for the
gpt-5 / o-series reasoning models; new ``_ollama_invoke`` that
posts to localhost:11434 with no third-party dep; per-case
``latency_ms`` is now captured on every NEW cached response
(future runs only — these four runs pre-date the patch).
- comparison.py: ``_load_cases`` composes curated + generated
(185 cases) instead of curated only; ``_score_provider``
surfaces ``latency_summary`` when records carry latency_ms.
- tests: provider-registry test relaxed to "cloud trio is a
subset of PROVIDERS"; env-key test allows ``_KEY`` (cloud
secret) or ``_URL`` (local endpoint).
Refines BoundUnknown from "the symbol whose value the solver determines"
to "the symbol at a specific temporal/state index with a specific
question-form". Two new required fields on BoundUnknown — state_index
(initial/terminal/Operation(operation_index)) and question_form
(count/rate/total/difference/ratio/identity) — populated by the new
pure-function resolver in generate/binding_graph/question_target.py.
The adapter (ADR-0133) now delegates Unknown -> BoundUnknown construction
to bound_unknown_from_math_problem_graph. No runtime wiring, no solver
invocation. Phase 5 (bounded-grammar / B3 integration) remains deferred.
Refusal-first via the new QuestionTargetError (sibling of AdapterError /
AdmissibilityError). Closed reason vocab: not_a_math_problem_graph,
unknown_entity_not_in_entities, apply_rate_unit_mismatch,
unmappable_question_form. Closed precedence rule on question_form
documented in ADR-0135 (compare_multiplicative > compare_additive >
apply_rate{numerator|denominator unit-match} > count); ambiguity refuses.
SemanticSymbolicBindingGraph.__post_init__ gains a cross-collection
guard: Operation(operation_index) must satisfy operation_index <
len(equations). canonical_string emission widened to include
state=... form=... tokens (hash differs from Phase 3 main by design —
not a regression; byte-equal across runs preserved).
Parents: ADR-0132 / ADR-0133 / ADR-0134.
Tests: +70 new (45 unit in test_binding_graph_question_target.py +
25 integration in test_binding_graph_adapter_question_target.py); 5
Phase 1+3 BoundUnknown fixtures migrated. Total binding-graph lane
295/1 pass (1 pre-existing test_symbol_binding_uses_slots failure on
Python 3.14, unrelated to Phase 4 — exists on origin/main). Pyright
clean on new and modified files. No edits to algebra/, chat/, core/,
or runtime hot path. Field invariant untouched.
Wires deterministic, refusal-first dimensional analysis into the
binding-graph adapter. Every BoundEquation emitted by
bind_math_problem_graph now carries either admissibility_status='admitted'
+ populated unit_proof or admissibility_status='refused' + typed
refusal_reason. No silent coercion; no invented units; no solver.
Adds:
- generate/binding_graph/units.py — pure unit algebra over a 6-dim
integer exponent vector (length, time, mass, money, count,
temperature). Closed vocabulary loaded once from en_units_v1
(ADR-0127) and memoized; composite "<num>_per_<denom>" resolved
recursively; conservative depluralization; refusal-first.
- generate/binding_graph/admissibility.py — check_admissibility with
per-operation-kind dispatch over the closed 8-string vocab, typed
AdmissibilityError (closed reason set), frozen UnitProof.
- ADR-0134 documenting the contract, invariants, and Phase 4-5
deferrals.
Adapter changes are surgical: synthesizes operand-literal symbols where
the verifier needs them (op<NNN>__multiplicand / __divisor / __rate),
then stamps each equation via check_admissibility. Input/output types
unchanged; bind_math_problem_graph still byte-equal across runs.
Tests: 226 total in the binding-graph lane (110 Phase 1+2 still pass; 47
units + 40 admissibility + 29 adapter-units new). Pyright clean on all
new files. No runtime wiring outside generate/binding_graph/.
Phase 4 (question-target binding) and Phase 5 (B3 / bounded grammar)
remain deferred per the brief.
Tests on main had drifted from intentional substrate changes that
weren't propagated to their fixtures or pinned values. Categories:
1. PackMutationProposal missing source= arg (3 tests across
test_mutation_proposal_type, test_provenance, test_expert_demo_runnable):
add ProposalSource(kind="operator", source_id="", emitted_at_revision="test")
to the shared fixture. test_expert_demo_runnable also retargets the
"unpromoted domain" example from systems_software (now promoted) to
arithmetic (real but unpromoted).
2. Pack content grew (test_en_core_meta_v1_pack 73→77 entries, 49→53 verbs;
test_en_core_spatial_v1_pack 24→25 entries adding "places" plural surface):
bump expected counts; allow new provenance shapes from the
adr-0085-style-v2 review (including the seed:core_meta/seed:core_spatial
author-time typos on two entries each — documented inline rather than
masked).
3. Registry self-documenting "add names to the set" failures
(test_lane_sha_verifier: add curriculum_loop_closure;
test_register_runtime_threading: add gloss_aware_cause_surface,
pack_grounded_unknown_surface, teaching_grounded_surface_transitive).
4. Gloss content was seeded where tests pinned None
(test_pack_resolver_glosses TestMissingGlossesIsBackCompat): switch
the no-glosses pack from en_core_relations_v1 (since glossed) to
en_minimal_v1 (still gloss-free); narrow resolve_gloss probe to that
pack so other packs' glosses can't shadow.
5. Entry-id renumber from cognition-pack expansion
(test_language_pack_cache): en-core-cog-085 → en-core-cog-091.
6. Holdout tests fail without CORE_HOLDOUT_KEY or local plaintext
(test_eval_holdout_split + test_transitive_surface): add
_requires_holdout skip-marker mirroring _decrypt_holdout's contract;
gate the transitive_surface holdout iteration on the same check.
7. Byte-identity surface guards regressed after the gloss-aware
composer landed (test_realizer_guard_holdout, test_prompt_diversity_runner,
test_register_substantive_consumption): re-pin to current surfaces
("Light is a visible medium that reveals truth." replaces "Light is a
source of revelation that makes things knowable.", etc.). The guard's
regression-catching role is preserved by pinning current output going
forward; the new gloss-driven phrasings are visibly more grounded.
Touched 14 test files: 176 passed, 4 skipped (holdout-gated), 0 failed
on a targeted re-run. No production code touched.
* feat(evals): add deterministic symbolic equivalence generated corpus
* feat(evals): add symbolic equivalence replay helpers
* feat(evals): load generated symbolic equivalence corpus
* feat(evals): emit symbolic equivalence replay manifest
* feat(symbolic): support multivariable integer polynomials
* feat(symbolic): support exact rational polynomial coefficients
* feat(symbolic): align equivalence API with multivariable normalization
* test(ADR-0131.1.B): reconcile v1 expectations to v1.B scope expansion
The v1.B refactor (univariate int → sparse multivariable Fraction) deliberately
admits multivariable polynomials and constant-denominator division. The v1
dataset and tests pinned the old refusal behavior, so the lane runner reported
wrong=4 and 10 unit tests failed.
Reconcile:
- cases.jsonl: flip sym-eq-v1-0029 ('x+y' vs 'x+1') and sym-eq-v1-0030
('x/2' vs 'x') from expected=refused to expected=not_equivalent; rename
categories to multivariable_distinct / constant_denominator_distinct;
extend provenance with adr-0131.1b:scope-expanded.
- generated_cases.py: split _refusal_cases into scope_expanded (admits)
and templates (still refused); the first two adversarial cases move to
the scope-expanded list with expected=not_equivalent.
- test_math_symbolic_normalizer.py: replace test_undefined_variable and
test_unknown_operator_division with positive scope-expansion tests +
symbolic-denominator refusal; rewrite TestPolynomialInvariants for the
new terms/variables constructor (Polynomial(terms={...}, variables=(...)))
with float-rejection and zero-coef-collapse invariants.
- test_math_symbolic_equivalence.py: TestRefused.test_empty_left reason
string matches new normalizer error; flip multivariable + constant-
denominator cases to NOT_EQUIVALENT; add symbolic-denominator-refused
case; relax canonical_a assertion in test_a_normalizes_b_refuses (engine
now zeroes both on either-side refusal).
- report.json + manifest.json: regenerated; lane PASS 185/185 wrong=0.
Lane invariants reaffirmed by the new tests: wrong==0, refusal-first for
truly out-of-scope inputs (symbolic denominator, transcendental, malformed,
negative exponent), determinism via byte-equal report.
ADR-0131 Benchmark 1 substrate — the primary discriminator for the
mathematics_logic expert promotion under the architecture-aligned
benchmark composite proposed in ADR-0131.
WHAT LANDED:
generate/math_symbolic_normalizer.py
Deterministic univariate polynomial normalizer. Scope: single
variable, integer coefficients, +/-/*/** operators, parens, no
division, no transcendentals. Pipeline: tokenize -> recursive-
descent parse -> expand-and-collect -> canonical string. Refusal
is first-class via SymbolicError; out-of-scope inputs refuse
rather than guess (preserves wrong == 0).
generate/math_symbolic_equivalence.py
check_equivalence(a, b) -> EquivalenceVerdict
Returns EQUIVALENT / NOT_EQUIVALENT / REFUSED with canonical
strings + reason. Compares byte-equal canonical forms.
evals/math_symbolic_equivalence/v1/
cases.jsonl — 30 hand-curated cases across 18 algebraic
identity categories + 2 out-of-scope refusals.
Coverage: commutative, distributive, square +
cube of binomial, difference of squares, FOIL,
collect like terms, zero cancellation, factoring,
exponent combination, unary negation.
runner.py — CLI entry point. Loads cases, builds report,
writes JSON, exits 0/1 on gate pass/fail.
README.md — methodology, scope, dataset categorization,
exit criterion, baseline result.
tests/
test_math_symbolic_normalizer.py — 44 tests covering parser,
algebra primitives,
canonical-form invariants,
and every refusal path.
test_math_symbolic_equivalence.py — 16 tests on the public
check_equivalence API.
test_adr_0131_1_symbolic_equivalence_lane.py
— 8 tests gating the lane:
dataset integrity, exit
criterion, wrong == 0,
determinism (byte-equal
report across runs).
EMPIRICAL RESULT (the lane PASSED):
correct = 30 / 30 (100.0%)
wrong = 0 / 30 (wrong == 0 invariant satisfied)
refused = 0 / 30 (refusals all matched expected)
correct_rate = 1.00
exit_criterion: PASSED (>= 0.95 required)
CONTRAST WITH ADR-0127-0128 GSM8K TRAIN-SAMPLE RESULT (0/0/50):
This is the first benchmark on the mathematics_logic lane where
the architecture's structural strengths fully express. The result
is the empirical inverse of the GSM8K result — and that's
exactly the architecture-benchmark fit ADR-0131 was written to
re-target toward.
REGRESSION: 1033/1033 existing tests green across math + ADR-0126
+ pack ratification + runner. Zero regressions.
SCOPE DISCIPLINE (per ADR-0131.1 v1 plan):
v1 deliberately narrow (univariate, integer, polynomial). Future
ADR-0131.1.B expansions documented in README: multi-variable,
rationals, larger dataset (~500), sealed holdout per ADR-0119.7
pattern.
PARALLEL WORK (per ADR-0131 plan to run all 3 sub-phases concurrently):
- ADR-0131.2: CORE-native teaching-corpus eval (separate PR)
- ADR-0131.3: bounded-grammar word-problem set (separate PR)
These are independent of ADR-0131.1; no shared files, no
cross-PR coordination required beyond final composite gate.
Exhaustive English linguistic-form ontology for quantities:
cardinals (0..20 + tens + magnitudes + compound rule), ordinals
(1st..31st + decade/magnitude forms), named fractions (1/2..1/10
+ sixteenth/thirty-second) + symbol forms (½ ¼ ¾ ⅓ ⅔ ⅛ ⅜ ⅝ ⅞),
multipliers (double/triple/twice/half), quantifiers with
semantic_type (indefinite triggers refusal at parse time —
preserves wrong==0), comparison anchors migrated for
ratifiability, number-format regexes with positive/negative
corpora.
Loader API in language_packs/numerics_loader.py (sibling module
to be merged into main loader after Gemini's ADR-0127 loader
lands, to avoid concurrent merge conflict).
Ratification invariants gated: cardinal/ordinal/fraction
exhaustiveness, quantifier semantic-type closed set, format-regex
test corpora (10+ positive/negative per format, ambiguity
refused), manifest checksums = SHA-256 of bytes-on-disk,
self-sealing mastery report.
Cross-references en_units_v1 (Gemini ADR-0127): fraction symbols
authoritative here; en_units_v1 symbol-affix table will point to
these entries.
No parser changes (deferred to 0128.3-0128.6). No train-sample
re-run (joint exit gate with ADR-0127 runs after both packs land).
Total: 130 lexicon entries across 7 kinds.
Lanes: smoke 67/0/0, packs 6/0/0, ADR-0128 suite 243/0/0.
P3 — generate/math_candidate_graph.py:
Branch enumeration over per-sentence candidate choices (Cartesian
product, cap=64). Per-sentence ambiguity tiebreaker via most-grounded-
slots-wins (transfer beats subtract when 'to Tom' grounds). Decision
rule: 0 admissible -> refuse; 1 -> emit; >=2 same answer -> emit;
>=2 different answers -> refuse (preserves wrong==0 on genuine
ambiguity). End-to-end parse_and_solve(text) -> CandidateGraphResult.
Question extractor added to math_candidate_parser.py (CandidateUnknown,
total + entity question shapes mirroring math_parser).
22 new tests. Permissive verbs ('bought', 'ate', 'bakes') now produce
correct answers via the candidate-graph path; ambiguous 'gives to Tom'
resolves to transfer reading (Tom gets the apples) deterministically.
P4 — evals/gsm8k_math/runner.py:
New sibling function _score_one_candidate_graph(case) -> CaseOutcome.
Identical shape to _score_one; swaps parse_problem for parse_and_solve;
preserves verifier/realizer/expected-answer stages. Callers (e.g.
PR #160's train_sample/v1/runner.py) substitute the new function in
one line to evaluate the candidate-graph topology.
9 new wiring tests. Three groups:
- No regression: cases legacy solves, new also solves.
- Lift: cases legacy refuses, new solves (the architectural payoff).
- Wrong==0: out-of-grammar refuses, never wrong.
Regression: 714/714 existing math + runner tests still green.
ADR-0126 total: 74/74 tests green across P1+P2+P3+P4.
Sibling to math_parser.py — pure candidate-extraction functions that
emit list[CandidateOperation] per sentence without mutating any state.
State threading defers to P3 (per-branch graph assembly).
Topology change vs legacy:
- No first-match-wins; every verb-kind regex runs independently.
- Ambiguous verbs ('gives', 'returns') emit multiple candidates;
P1's round-trip filter + P3's decision rule resolve.
- Out-of-grammar sentences return [], NOT ParseError. Empty list
is the deterministic 'no candidate' signal.
Permissive verb tables (imported from math_roundtrip.KIND_TO_VERBS)
mean past-tense and production verbs ('bought', 'ate', 'bakes')
that the legacy parser refused are now admissible — the round-trip
filter is the safety mechanism, not regex narrowness.
P2 scope (canonical Subject-verb-Value-Unit-[to-Target] shape only):
- extract_initial_candidates(sentence) for 'X has N units'
- extract_operation_candidates(sentence) for add/subtract/transfer
Out of scope (deferred to later sub-phases):
- Pronoun resolution / unit inheritance (needs per-branch state)
- Multiply / divide / rate / comparison (same machinery, more matchers)
Regression: existing math suite 701/701 green. Zero changes to
math_parser.py, math_solver.py, math_verifier.py, math_realizer.py.
The wrong-answer firewall for the candidate-graph parser topology.
A CandidateOperation carries an Operation plus source-span provenance
for every content slot the parser claimed (verb, value, unit, actor,
transfer target, comparison reference). roundtrip_admissible() checks
each slot grounds in the source span AND the matched verb is
registered for the claimed kind.
Two consequences:
- A regex that mis-reads 'loses' as add fails (loses not in ADD_VERBS).
- A regex that hallucinates a number/unit not in source fails to ground.
KIND_TO_VERBS is the new single source of truth for {kind -> verbs};
P2 will refactor math_parser to consume it. Verb tables are
permissive by design (much wider than current narrow regex tables)
because the filter rejects wrong candidates downstream — narrowness
is no longer the safety mechanism.
Deterministic: pure byte/regex containment. No randomness, no
learning, no approximation. Preserves wrong==0, trace_hash byte-
equality, replay determinism.
Wraps existing math pipeline (parser -> solver -> verifier) against
PR #159's 50-case train sample. Emits deterministic report.json with
per-case verdicts. CLI exit code reflects exit criterion
(correct >= 10 AND wrong == 0).
Baseline against current parser: 0 correct / 0 wrong / 50 refused.
This baseline is the inner-loop gradient signal for ADR-0126's
candidate-graph parser (in flight on feat/adr-0126-candidate-graph).
Registers tests/test_adr_0126_train_sample_runner.py under
'core test --suite math' so the wrong == 0 invariant becomes a hard
CI gate per ADR-0114a Obligation #4 (refuse rather than confabulate).
Depends on PR #159 (gemini/adr-0126-train-sample). Rebase onto main
after #159 lands.
ADR-0123-parser-comparison-phrasing as the **surface increment** on
PR #155's substrate (commit c9bd5d4). Closes the last architectural
gap in the comparison-phrasing class: before this commit, the
substrate's solver evaluated comparison problems successfully but
realize() crashed with `unknown operation_kind 'compare_additive'`
when asked for show-your-work prose.
Substrate (PR #155) already shipped:
- `Comparison` typed graph operand
- `compare_additive` / `compare_multiplicative` operation kinds
- parser patterns for the four canonical surfaces
(N more / N fewer / twice / N times / half)
- solver + verifier wiring + pack lemmas
(en-arith-006 compare_additive, en-arith-007 compare_multiplicative)
This surface adds:
- `_compare_additive_sentence(step)` rendering `direction='more'|'fewer'`
- `_compare_multiplicative_sentence(step, entity_units)` rendering
`direction='times'|'fraction'`
- two new branches in `_step_sentence` dispatch
- `_step_sentence` signature widened with optional `entity_units` map
(derived once-per-trace in `realize()` from `graph_initial_state`)
- ADR-0123-parser-comparison-phrasing.md (~15 invariants, substrate
+ surface decomposition rationale, multi-construction barrier
inheritance)
- 26 invariants pinned across canonical surfaces, plurality
independence, byte-determinism, refusal discipline, and
backwards-compatibility with the pre-comparison realizer templates
End-to-end pipeline now operates on all four canonical comparison
shapes:
parse_problem(
"Alice has 5 apples. Bob has 3 more apples than Alice. "
"How many apples does Bob have?"
) -> solve() -> realize().as_prose() ->
"Alice has 5 apples. Bob has 3 more apples than Alice, giving Bob
a total of 8 apples. Bob has 8 apples."
Measurement (this PR):
- 26/28 direct ADR-0123 tests pass; 2 skipped (CORE_HOLDOUT_KEY)
- `core eval cognition` byte-identical: 100/100/100/100
- ADR-0118 stepped-realizer templates re-render byte-identically
- Substrate measurements continue to hold
Honest non-result: sealed `correct_rate` stays at 0/1319. The
realizer cannot create matches the parser refuses; the multi-
construction barrier the substrate ADR documented holds at the
surface layer too. Cumulative lift signal expected only after the
3rd/4th foundational class lands (per ADR-0121's revised
sequencing). `wrong == 0` holds by construction — realizer only
renders successful traces.
Pre-existing failure noted (not introduced by this PR):
`tests/test_adr_0085_gloss_aware_cause.py::test_flag_off_metrics_byte_identical`
fails on substrate base (c9bd5d4) without these changes — an
ADR-0085 cognition baseline drift unrelated to the realizer.
First worked attempt at promoting a domain under the ADR-0120
expert promotion contract. The contract refuses honestly.
Gate evaluation against live state:
ADR-0114a obligations: 10 of 10 pass
ADR-0120 contract-level gates:
audit_passed_holds ✓
correct_rate (public) ✓ 150/150 = 1.0
correct_rate (sealed) ✗ 0/1319 = 0.0 < 0.60 floor
signed_expert_claim ✗ (no entry, downstream of correct_rate)
Decision: mathematics_logic NOT promoted; stays at audit-passed.
Substantive blocker: parser grammar covers 0/1319 of real GSM8K.
What this proves
- The contract is genuinely falsifiable. ADR-0120 §"Threshold
rationale" deliberately set the floor above current measurement
so the first attempt would defer honestly. Same load-bearing
pattern as ADR-0107 → ADR-0110 for audit-passed.
- Wrong-zero discipline holds against real GSM8K (the load-
bearing positive claim). CORE refuses every problem outside
its grammar without confabulating on a single one.
What unlocks the promotion
Multi-ADR parser-expansion arc lifting sealed-GSM8K correct_rate
from 0.0 to ≥ 0.60. Each construction class (rate/comparison/
percentage/time-modal/etc.) ships as its own scoped ADR with:
- parser+solver+verifier+realizer extensions
- re-measurement on sealed holdout
- ADR-0118a OOD re-measurement (no surface-feature regression)
- ADR-0125 perturbation re-measurement (no invariance regression)
- ADR-0119.5 adversarial re-measurement (no new misparses)
Honest-fitting discipline: every lift is graded on the anti-
overfit obligations BEFORE the correct_rate change counts.
Tests: 6/6 with CORE_HOLDOUT_KEY; 4/6 + 2 skipped without (matches
ADR-0119.7 seal discipline).
This deferral demonstrates the expert tier's promotion machinery
is load-bearing — the gate has refused at least once before any
domain reaches it.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The 1,319 GSM8K test cases are now sealed at
evals/gsm8k_math/holdouts/v1/cases.jsonl.age, age-encrypted to the
ADR-0119.1 recipient. Plaintext never touched disk in the working
tree; only ciphertext is committed.
First honest CORE-vs-real-GSM8K measurement
cases_total: 1319
correct: 0
wrong: 0 ← ADR-0114a Obligation #4 holds against external corpus
refused: 1319
overall_pass: True
Zero confabulation. Parser refuses what it can't grammar-handle; the
"wrong == 0" discipline survives the move from CORE-original cases
to a real public benchmark. The 0/1319 correct rate is the truthful
gap that ADR-0120's threshold work will quantify.
What landed
scripts/seal_gsm8k_test.py
- Loads GSM8K via datasets.load_dataset("openai/gsm8k", "main")
- Strips worked-solution prose; extracts final-answer integer/float
after "####" (handles "2,125" → 2125 thousands-separator)
- Reads recipient from docs/holdout_recipients.txt (single repo key
per ADR-0119.1)
- Encrypts via pyrage; writes only ciphertext
- Refuses to overwrite test path with train-derived seal
evals/gsm8k_math/runner.py
- Empty expected_unit (sentinel) skips unit-comparison; grades on
answer value alone. Required because GSM8K answers carry no unit
structurally. wrong-zero discipline preserved.
tests/test_adr_0119_7_sealed_gsm8k.py — 6 invariants:
1. sealed file present + age-formatted
2. no plaintext companion files (sibling-leak guard)
3. decrypted JSONL matches documented schema
4. runner against decrypted suite produces wrong==0
5. tests skip (not fail) when CORE_HOLDOUT_KEY unset
6. case ids match "gsm8k-test-NNNN" pattern
Defensive gitignore: plaintext patterns under
evals/gsm8k_math/holdouts/v1/ are explicitly excluded.
ADR-0114a obligation roll-up
10/10 discharged for the gsm8k_math lane:
#1 ✓ sealed-holdout (fab_control + GSM8K test)
#2..#10 ✓ as before
Phase 5 status: 5.1..5.7 done; 5.8 in flight (PR #149). After 5.8
merges, ADR-0120 (first expert promotion contract) becomes
feasible.
Test plan
- pytest tests/test_adr_0119_7_sealed_gsm8k.py with CORE_HOLDOUT_KEY → 6/6
- pytest without CORE_HOLDOUT_KEY → 3 pass + 3 skip
- core test --suite smoke -q → 67/67
- CLAIMS.md regenerated (no diff)
- HF token NEVER in repo (saved at ~/.cache/huggingface/token, mode 600)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Audit follow-ups from #145/#146 merge review. Five small fixes; no
behavior change on the green path, but failure modes are now explicit
rather than silent.
ADR-0119.6 depth_curve.py
- Add DepthCurveError typed exception
- Raise on case_id missing from lane_report (was: silent → "refused")
- Raise on depth >= 9 (was: silent new bucket key)
- Two new tests pin both refusals
- Removed stale sys.path hack at module top
ADR-0119.4 frontier-baseline tests
- Assert comparison_v1.json's core_measurement reports wrong == 0
(the load-bearing differentiator named in the disclaimer; a
tampered file with wrong > 0 was previously syntactically valid
and would have passed all old assertions)
- Assert frontier citations are dated 2023 or later (freshness
guard; older citations should be refreshed before ADR-0120
gates anything for `expert` promotion)
Tests
- tests/test_adr_0119_6_depth_curve.py: 7 → 9
- tests/test_adr_0119_4_frontier_baseline.py: 5 → 7
- 29/29 across runner + depth-curve + frontier suites; 67/67 smoke
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 4 of the ADR-0114 GSM8K-math roadmap. Consumes a SolutionTrace
and emits one sentence per step plus setup + answer sentences. Pure
function; same trace → byte-equal RealizedTrace.
What landed
generate/math_realizer.py
- realize(initial_state, trace) -> RealizedTrace
- Frozen RealizedTrace dataclass with canonical_bytes() + as_prose()
- Per-kind sentence rules (add / subtract / transfer / multiply×2 /
multiply×3 / multiply-general / divide)
- Singular/plural surface rule matches parser canonicalization
- Typed RealizerError on unrecognized step kinds
tests/test_math_realizer.py — 60 cases pinning five invariants:
1. All 50 dev-set cases realize without error
2. Determinism: byte-equal RealizedTrace across two calls
3. Setup sentence count == initial_state count
4. Step sentence count == operation count
5. Answer sentence contains the resolved value + unit
ADR-0114a obligation discharge update
ADR-0118 hardens determinism (#9) across a third layer (realizer)
and makes #3 / #10 human-inspectable via the prose surface. No
obligation is directly newly discharged by ADR-0118; it's substrate
for ADR-0119 GSM8K eval lane.
Round-trippability of the prose through the parser is explicitly
out of scope for this phase. The trace is the verifiable artifact
(ADR-0117); the prose is human-readable documentation.
Tests: 60 new realizer cases; 546 total green across realizer +
parser + solver + verifier + OOD; 67/67 smoke green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 3 of the ADR-0114 expert-capability roadmap. Re-applies every
step of a SolutionTrace from the input graph's initial state and
asserts byte-equal reproduction of answer_value. Pure function; same
(graph, trace) → byte-equal VerifierVerdict.
Why this is distinct from the solver
ADR-0116's solver enforces correctness at construction. ADR-0117's
verifier is a SECOND, INDEPENDENT implementation that re-derives
every value the trace claims. The verifier does NOT call solve(). It
re-implements the operation semantics from ADR-0116 directly inside
_verify_step. If the solver had a bug or was tampered with after the
fact, the verifier catches it.
Six checks per verdict (named, ordered, audit-logged):
1. graph_canonical_hash_matches
2. pack_id_matches
3. pack_lemmas_resolve
4. step_pack_lemma_ids_match_bindings
5. step_replay_matches_before_after
6. answer_value_reproduces
Seven named tamper classes all caught:
- mutated before_value / after_value / operand of any step
- mutated pack_lemma_id of any step
- mutated graph_canonical_hash
- mutated answer_value
- mutated pack_id
- mutated target_before / target_after of transfer step
ADR-0114a obligation update
#3 Replay-equal trace — now discharged at VERIFIER FIDELITY
(was solver-only under ADR-0116). A third party with only
(graph, trace, pack) can reproduce the answer byte-equal.
Five of ten obligations now load-bearing: #3, #4, #9, #10 plus
in-flight #2 (Codex's ADR-0118a OOD generator).
Tests: 62/62 verifier suite green; 67/67 smoke green; existing
solver + parser + schema suites unaffected.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 2 of the ADR-0114 expert-capability roadmap. Consumes the
MathProblemGraph from Phase 1 and emits a SolutionTrace — ordered
operation applications ending at a numeric answer, byte-deterministic
across runs, with each step's operation bound to a pack-resolved
lemma identifier.
What landed
generate/math_solver.py
- solve(graph) -> SolutionTrace; pure function, no I/O, no globals
- SolutionStep dataclass with before/after values per step (for
verifier replay; ADR-0117 hardens)
- SolutionTrace with canonical_bytes() byte-deterministic JSON
- SolveError typed refusal: missing pack, division by zero,
unknown-references-nothing
language_packs/data/en_arithmetic_v1/
- 5 operator lemmas: add / subtract / multiply / divide / transfer
- role=operational_base (vocabulary-only; no domain claim)
- SHA-256-anchored lexicon + glosses; manifest carries
provenance=adr-0116:operator_seed:2026-05-22
tests/test_math_solver.py — 109 cases pinning five invariants:
1. Phase 2 exit criterion: ≥ 0.80 on parser-correct dev set
(current: 50/50 = 1.00)
2. Determinism: two solves produce byte-equal trace
3. Trace replay reproduces answer_value (verifier rehearsal)
4. Typed refusal on under-determined inputs
5. Every step.pack_lemma_id resolves to a real lexicon entry
in en_arithmetic_v1
ADR-0114a obligation discharge
Four of ten anti-overfitting obligations now have load-bearing
implementations in code:
#3 replay-equal trace — discharged (solver-layer)
#4 typed refusal — discharged (solver-layer)
#9 determinism — discharged (solver-layer)
#10 operation provenance via pack — DISCHARGED IN FULL
Removing the en_arithmetic_v1 pack now breaks every solve loudly.
The "operations bind to concepts, not hardcoded strings" claim is
architecturally true, not rhetorical.
Tests: 109/109 green on solver suite; 67/67 smoke suite green;
parser + schema suites still green from prior phases.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Closes Phase 1.3 of the ADR-0114 expert-capability roadmap. Turns a
grade-school word problem into a typed MathProblemGraph deterministically
(no LLM, no sampling). Same input string always produces the same
graph; unsupported constructions raise ParseError rather than guessing.
What the parser handles
Initial possession: "<E> has <N> <unit>."
Add verbs: buys, gets, finds, receives, earns, adds
(+ "<N> more" / unit elision via state.last_unit)
Subtract verbs: eats, loses, sells, donates, uses, spends, drops, removes
Transfer verbs: gives, sends, hands, passes, mails (with target)
Multiply (scalar): "X doubles <obj>" / "X triples <obj>"
Divide (split): "X splits {them|his Y|N Y} evenly into M groups [and keeps one]"
Compound sentences: "X buys 5, then donates 3."
Sentence opener: "Then X eats 1." (inherits subject + unit)
Pronoun anaphora: he/she/it → last-introduced singular subject
Object pronoun: them/these/those → state.last_unit
Trailing PP: "finds 7 buttons on the floor" — discarded
Singular→plural: "Iris has 1 coin" → canonical unit "coins"
Questions:
"How many <unit> does <E> have [left|now|in total|altogether]?"
"How many <unit> do they have [in total|altogether|left|now]?"
What it explicitly rejects
- Conditional / time-modal ("If X had ...")
- Compound questions (two unknowns)
- Multiple "?" sentences
- Questions referencing entities never introduced
- Empty / whitespace-only input
Verification
- tests/test_math_parser.py: 20 cases (5 byte-equal parametrized
+ 5 determinism parametrized + 1 exit-criterion gate + 6 typed-
refusal + 2 purity + 1 type check)
- tests/test_math_problem_graph.py: 26 schema cases still green
- On the 5 seed cases: 5/5 = 100% byte-equal
- On Codex's PR #128 50-case dev set (locally tested):
49/50 = 98% byte-equal. Single failure (gpd-021) is a case-
quality issue, not a parser limit; feedback filed on #128 to
rewrite (mixed units + metaphor not in pattern registry).
- Phase 1.3 exit criterion (≥ 0.90): met.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
First Phase of ADR-0114's expert-capability roadmap. Decomposed into four
sub-phases so each lands as its own auditable step:
1.1 schema + 5 seed cases + invariants ← this commit
1.2 45 more dev-set cases ← delegated (Codex)
1.3 the parser itself ← exit: ≥0.90 on dev set
1.4 runtime binding ← if non-trivial
What landed
- generate/math_problem_graph.py — typed dataclasses (Quantity,
InitialPossession, Operation, Unknown, MathProblemGraph) + frozen
validation + canonical_bytes() byte-deterministic serialization +
graph_from_dict roundtrip.
- evals/gsm8k_parser_dev/cases.jsonl — 5 seed cases (gpd-001..005)
covering single-add, single-subtract, multi-step, two-entity
transfer, and multi-entity sum constructions. Every case carries a
ground_truth_graph and the documented patterns it exercises.
- evals/gsm8k_parser_dev/README.md — authoring contract: schema,
pattern registry, canonicalization rules, Phase 1.1 scope boundary,
hand-solving rubric, distribution target for the remaining 45
cases. This is the spec Phase 1.2 authors work against.
- tests/test_math_problem_graph.py — 26 cases pinning four invariants:
round-trip byte equality, canonical_bytes() determinism, schema
rejection of malformed graphs, and ground_truth_graph ↔
expected_answer agreement (a hand-solver inside the test module
falsifies mis-authored cases).
Why this is sticky
The Phase 1.1 schema is load-bearing for Phase 1.2 (the 45 authored
cases will be written against it) AND Phase 1.3 (the parser will be
graded byte-equal against ground-truth graphs in this schema). Changing
the schema after Phase 1.2 lands requires an amendment ADR + rewriting
authored cases. The schema choices here are intentionally conservative.
Tests: 26/26 new; 67/67 smoke green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The word "expert" in the previous status name implied raw-capability parity
with frontier LLMs on the same benchmark — which the gate does NOT verify.
What the gate actually verifies is CORE *claim-shape compliance*:
* signed digest (replay-reproducible from on-disk lane results)
* replay determinism (same inputs → byte-equal trace_hash)
* typed refusal (fabrication refused, not paraphrased)
* exact recall (no ANN, no cosine, no attention bottleneck)
* grounding-source provenance
These are claim shapes a transformer LLM cannot structurally produce
regardless of raw accuracy. A frontier LLM might score higher on the
same benchmark but cannot pass this contract.
Rename scope (semantics only, per ADR-0113):
status string "expert-demo" → "audit-passed"
predicate key predicates.expert_demo → predicates.audit_passed
reason key expert_demo_reason → audit_passed_reason
YAML key expert_demo_claims → audit_passed_claims
CLI command core demo expert → core demo audit-passed
output dir evals/expert_demos/ → evals/audit_passed/
artifact filenames expert_demo.{json,html} → audit_passed.{json,html}
HTML title CORE Expert-Demo: X → CORE Audit-Passed: X
Internal Python identifiers (module/file/function/class names like
`expert_demo.py`, `evaluate_expert_demo`, `ExpertDemoClaim`,
`expert_demo_claim_for`) are deliberately kept to minimize churn. ADR
file titles (ADR-0106..0112) preserved as historical record.
`expert` namespace reserved for ADR-0114+: an actual capability tier
above `audit-passed` backed by a public benchmark with a stated
threshold. ADR-0114 proposes the first such target — GSM8K-math —
laying out a falsifiable 7-phase arc (parser → solver → verifier →
stepped-realizer → eval lane → first `expert` ledger tier promotion).
Tests: 184 directly-affected tests green (140 capability/expert-demo
suite + 34 demo/audit-tour + 10 correction-cue). Smoke suite 67/67.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Closes the asymmetry between the `expert-demo` ledger status (audit
artifact only) and the actual `core demo` surface (runnable
walkthroughs producing HTML + JSON). Until this commit the word
"demo" in `expert-demo` was aspirational; now it corresponds to
something a reader can open.
What it does
- Reads the signed expert_demo_claims entry from docs/reviewers.yaml
- Loads latest on-disk result files for each attached lane × split
- Re-derives the evidence-bundle digest and asserts byte-for-byte
match against the signed claim_digest — this is the load-bearing
audit step, now exercised at two independent enforcement points
(ledger gate + showcase)
- Runs each lane's metrics through the ADR-0109 lane-shape registry
and surfaces the verdict
- Picks the first three cases from each split verbatim (deterministic
by file order) and renders them as HTML for inspection
- Emits expert_demo.json (canonical bytes, deterministic) + expert_demo.html
Surface
core demo expert --domain mathematics_logic
core demo expert --domain physics
# → evals/expert_demos/<domain>/latest/expert_demo.{json,html}
Read-only by construction: cannot mutate docs/reviewers.yaml or any
lane result file. Tested. Unpromoted domains raise ValueError —
no silent fallback, no "preview" mode that fakes a showcase.
Generated artifacts are gitignored — the inputs they derive from are
already committed, so duplicating the renders would just churn the
tree.
Tests: 16 new cases pinning all five ADR-0112 invariants. Smoke suite
still 67/67 green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Second worked promotion exercising the ADR-0106 + ADR-0109 contract
on a domain distinct from mathematics_logic. No contract change.
Evidence:
- foundational_physics_ood: accuracy=1.0 (117/117 public, 39/39 holdout)
- inference_closure: all_pass_rate=1.0 (shared with math, distinct digest via domain_id)
- fabrication_control: refused=n, fabricated=0 across all classes (shared)
Signed claim digest: a104cad136f3219df05dc7ce6a78437c02f7b5827cd3cdce568db3acda6a43ed
Bridge landed: cases_plaintext.jsonl dev-mode fallback for
foundational_physics_ood (matches ADR-0105 convention; analogous to the
math/inference bridges in ADR-0110). One small file, not a contract change.
Tests:
- tests/test_adr_0111_physics_expert_demo.py — 4 invariants, 6 cases
- tests/test_adr_0110_math_expert_demo.py — relaxed "only math promoted"
to "math stays promoted" (load-bearing for ADR-0110 is persistence)
- tests/test_capability_reports.py — physics row now expert-demo
Retires the "first promotion was math-specific" objection: the bridges
ADR-0110 landed were correctly scoped, and the contract holds across
two distinct domains using shared lane infrastructure with distinct
digests.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Between 2026-05-17 and 2026-05-22 the inference_closure lane regressed
from all_pass_rate=1.0 to 0.4 on public. Root cause: the
_DECLARATIVE_RELATION_RE branch in generate/intent.py runs ahead of the
_RULES loop and swallowed sentences beginning with 'Actually' into the
subject phrase, routing them to VERIFICATION. The lane's premise emit
path is gated on CORRECTION intent, so PackMutationProposal records
stopped being emitted for any non-'is' relation (precedes / grounds /
causes / reveals). Only the four transitive_is cases passed because
'is' is not in the declarative-relation verb list.
Fix: _CORRECTION_CUE_PREFIX_RE guard. When the text begins with a
correction cue ('Actually', 'Incorrect, ', 'No, ', 'Correction'), the
declarative-match branch is skipped and the sentence falls through to
the _RULES CORRECTION rule. Plain declarative-relation assertions still
route to VERIFICATION unchanged.
Lane on 2026-05-22 post-fix:
dev/v1: all_pass_rate=1.0, overall_pass=True (5 cases)
public/v1: all_pass_rate=1.0, overall_pass=True (20 cases)
- tests/test_correction_cue_prefix_routing.py pins both halves of the
guard (10 new tests).
- evals/inference_closure/gaps.md documents the regression + fix in a
new section, preserving the 2026-05-17 resolution narrative.
- evals/inference_closure/results/ now carries canonical v1_dev and
v1_public reports (the lane had no checked-in results before; ADR-0110
will reference these).
This unblocks the second of ADR-0107's two named blockers. ADR-0110
(math expert-demo re-attempt) now becomes feasible once the math
domain's three lanes have signed-and-digested evidence.
Replaces the cognition-shape-uniform threshold dispatch in
core/capability/expert_demo.py with an explicit LANE_SHAPE_REGISTRY
mapping 8 ratified lane ids to 5 shapes:
cognition -> cognition_shape
elementary_math_ood -> accuracy_shape
foundational_physics_ood -> accuracy_shape
symbolic_logic -> symbolic_logic_shape
hebrew_fluency -> accuracy_shape
koine_greek_fluency -> accuracy_shape
inference_closure -> inference_shape
fabrication_control -> refusal_shape
Each shape has a documented threshold checker. Unknown lane ids
fail-closed with a named reason. ADR-0106 \xc2\xa71.1/\xc2\xa71.3/\xc2\xa71.4/\xc2\xa71.5
unchanged; only \xc2\xa71.2 (threshold rules) dispatches by shape.
tests/test_lane_shape_thresholds.py pins all four ADR-0109 invariants
plus dead-shape and threshold-value gates (13 new tests).
tests/test_expert_demo_contract.py fixtures updated to provide
shape-appropriate metrics (no semantic change to those tests; same
12 cases still pin the ADR-0106 contract).
ADR-0109 status: Proposed -> Accepted. README sequencing updated
(ADR-0110 now only blocked by inference_closure, not by metric-shape
amendment).
Ledger: all five domains remain reasoning-capable, expert_demo=false.
The ADR-0106 contract correctly refused promotion. ADR-0107 records the
deferral and reserves two follow-up ADRs:
- ADR-0109 (lane-shape-aware threshold amendment): ADR-0106 \xc2\xa71.2
prescribes cognition-pack-shape metrics uniformly, but math /
physics / systems / hebrew-greek lanes carry native shapes
(accuracy, passed_rate, all_pass_rate). Prerequisite for any future
expert-demo promotion.
- ADR-0110 (math re-attempt): conditional on ADR-0109 landing and
inference_closure substantively passing (currently all_pass_rate=0.4
on public).
tests/test_adr_0107_deferral.py pins adr_0107_no_silent_promotion: math
stays at reasoning-capable, has no expert_demo_claims entry, and the
ledger row carries a named refusal reason.
No change to core/capability/expert_demo.py or reporting.py -- the
contract is honored, not amended. README sequencing updated to reflect
ADR-0107 acceptance and the new ADR-0109/0110 prerequisites.
Closes ADR-0106 acceptance evidence:
- ExpertDemoClaim dataclass + additive expert_demo_claims block on
ReviewerRegistry (schema_version stays at 1; backward-compatible).
- New core/capability/expert_demo.py with derive_evidence_digest,
evaluate_expert_demo, collect_domain_lanes, materialise_lane_results.
- core/capability/reporting.py: replaces the cognition-lane-only
predicate (previous lines 418-433) with a domain-aware,
reviewer-signed gate; ledger rows now also carry
expert_demo_reason for operator legibility. Reviewer registry is
fail-closed: an unloadable registry yields zero claims, so a broken
registry never silently grants expert_demo=true.
- tests/test_expert_demo_contract.py covers all three ADR-0106
invariants: requires_signature, domain_aware, replay_byte_equality;
plus threshold + production-ledger-untouched gates. 12 new tests.
- tests/test_reviewer_registry.py extended with TestExpertDemoClaimsSchema
covering omitted block, valid parse, unknown signer rejection,
malformed digest rejection, duplicate domain rejection. 5 new tests.
- README index row + table preface updated to note expert_demo is
contract-gated. Frontier list trimmed (ADR-0106 has landed).
- ADR-0106 Status flipped Proposed -> Accepted.
No domain row's expert_demo field flips by this PR -- only the contract
changes. Promotion of any ratified domain requires a follow-up ADR
(ADR-0107 reserved for mathematics_logic) plus a signed claim.
CLAIMS.md is now mechanically derived from two ground-truth sources:
- core.capability.ledger_report (Tier 1: ratified domains)
- scripts/verify_lane_shas.PINNED_SHAS (Tier 2: pinned lane reports)
The generator is deterministic and gated by
tests/test_claims_md_is_current.py + the lane-shas CI workflow's new
'verify CLAIMS.md is current' step. Drift between in-tree state and
the published claims fails CI before merge.
Tier 1 (5 ratified domains) and Tier 2 (6 pinned lanes) cover every
ADR-0092..0102 invariant currently CI-pinned.
Two pre-existing latent issues fixed:
1. discourse_planner flag catalog drift (test_flag_report failure)
On 2026-05-21 the discourse_planner default was flipped to True
after byte-equality verification (per inline comment in
core/config.py:130-138), but the capability flag catalog at
core/capability/reporting.py was not updated — it still claimed
"flag_shipped_default_off". The test
test_flag_report_tracks_default_off_flags_without_enabling_them
correctly caught the inconsistency; it had been failing across
every commit since ADR-0092 first ran the suite.
Fix:
- New "flag_shipped_default_on" state in _FLAG_CATALOG, added
to flag_report() grouped output
- discourse_planner moved from default_off → default_on
- Test renamed to test_flag_report_classification_matches_actual_defaults,
enforces BOTH directions of the contract (catalog claim must
match DEFAULT_CONFIG value)
- New test test_flag_catalog_state_is_consistent_with_default_config
cross-checks every catalog entry against DEFAULT_CONFIG;
catches future drift before it lands
2. public_demo lane SHA shifted every commit
Each commit advances the showcase's generated_at_revision field
(git HEAD SHA). _strip_volatile in the lane runner was stripping
wall-clock and per-run paths but NOT generated_at_revision, so
the byte-equality case's details.sha256 changed with every commit
even when underlying demos produced identical content. That made
the pin a "did this run today" check rather than a "did the code
produce the right artifact" check — exactly the failure mode
the verifier was supposed to prevent.
Fix:
- Add generated_at_revision to _VOLATILE_KEYS in the public_demo
runner. Lane's invariant is "same code → same SHA," not
"same HEAD → same SHA"; HEAD belongs in the showcase output
(operators need it) but not in the lane's equality projection.
- Pin refreshed once to capture the now-commit-independent SHA;
subsequent commits won't shift it unless underlying demo content
actually changes.
After fix:
- Capability tests: 6/6 passing (was 4/5 with discourse_planner failing)
- Lane SHAs: 6/6 match pinned values; public_demo pin will now survive
routine code changes
- Smoke 67/67, cognition eval byte-identical 100/100/100/100
This is the single known pre-existing test failure cleaned up.
Six lanes (reviewer_registry, miner_loop_closure,
domain_contract_validation, fabrication_control_summary,
demo_composition, public_demo) now have CI-enforced SHA-256 pins.
A failing job means a lane's deterministic output changed without
an explicit ADR-tracked pin update.
- new scripts/verify_lane_shas.py: single source of truth
- PINNED_SHAS dict mapping lane_id → 64-char hex SHA
- LANE_SPECS tuple wiring each lane to its runner module + canonical
report path
- accepts_report_flag handles the fabrication_control runner's
different arg shape (--lane-dir not --report)
- verify_all() runs each lane in subprocess isolation (clean Python
state per lane — relevant for adapters that cache pack loads at
module import)
- --update flag refreshes pins after intentional ADR-tracked changes;
diff is the audit trail
- --json flag emits machine-readable report
- exits non-zero on any mismatch
- new .github/workflows/lane-shas.yml:
- triggers on push to main and pull_request to main
- concurrency group cancels in-progress runs on new commits
- Python 3.11 + pip-cached deps + editable install
- runs verify_lane_shas.py; emits JSON report on failure
- 12-minute timeout (lanes take ~30s in practice)
- new tests/test_lane_sha_verifier.py: cheap local-pytest pinning
- every LaneSpec has a corresponding PINNED_SHAS entry
- no orphan pins without a LaneSpec
- every pin is a 64-char hex SHA-256
- every runner module path exists on disk
- canonical report paths are under repo root
- all six expected lanes (ADR-0092/0093/0095/0096/0098/0099) covered;
ADR-0094 and ADR-0097 are schema/ratification only, intentionally
excluded from EXPECTED_LANES
- 6 tests run in <100ms — catches drift before CI
- evals/public_demo/results/v1_dev.json: refreshed to match the new
pin (21751aaf..) — earlier pin was generated under slightly different
runner argparse defaults; --update produced the canonical bytes
Local verifier: 6/6 lanes match pinned SHAs. Smoke 67/67. Lane SHAs:
reviewer_registry 681a2aab..
miner_loop_closure 9f071733..
domain_contract_validation f9c06cde..
fabrication_control_summary 01e1b6b7..
demo_composition 27d83824..
public_demo 21751aaf..
Single 30-second artifact composing four CORE invariants
(determinism, honest unknown, reviewed learning, multi-hop with
trace) by delegating to existing DemoCommand adapters. **No new
mechanism** — every claim is backed by an already-shipped,
separately-tested adapter. Closes the 8-ADR scale-up slate.
- new core/demos/learning_loop_adapter.py: LearningLoopDemo wraps
ADR-0056 reviewed-teaching loop; _strip_volatile_paths drops
transient temp-dir paths from raw before serialization so the
adapter's report_sha256 is content-stable across runs
- new core/demos/showcase_adapters.py:
- FabricationControlPublicDemo: re-runs ADR-0096 public split,
produces 3 claims (refusal_recall_meets_threshold,
fabrication_rate_below_threshold, trace_evidence_present)
- MultiHopTraceDemo: runs 'Does light reveal truth?' with
transitive_surface=True + composed_surface=True against
cognition pack; surfaces a 3-hop walk light→truth→knowledge→
evidence; produces 3 claims (grounded_answer, depth_two_or_more,
walk_evidence_present)
- new core/demos/showcase.py: run_showcase() composes 4 scenes,
emits showcase.json + per-scene artifacts; render_html() produces
presentation-only static HTML with no JS injection vector;
ShowcaseScene dataclass; MAX_RUNTIME_SECONDS=30 hard ceiling
with DemoContractError if exceeded
- CLI: 'showcase' added to demo target choices; --output-dir flag
added; cmd_demo dispatch branch writes showcase.json + showcase.html
- new evals/public_demo/ lane with 4 cases:
- all_claims_supported (each scene + composite)
- determinism_run_to_run_byte_equality (two runs identical after
stripping volatile keys: total_runtime_ms, json_path,
transient_corpus)
- runtime_under_budget (≤30s)
- pure_composition_no_new_mechanism (grep gate over showcase
imports — must come from core/chat/generate/language_packs/
teaching/evals or allowed stdlib only)
- lane is itself byte-identical across runs (sha256 5707db8efc6a..);
runtime case omits exact runtime_ms (it varies near bucket
boundaries) but still asserts ≤ budget
- 8 unit tests with module-scoped fixture (showcase runs once,
~13s total) covering payload shape, scene order, runtime budget,
HTML render absence of <script>, and the pure-composition import
gate independently of the lane
- ADR-0099 measured: total_runtime_ms ~12.8s, well under 30s budget
- smoke 67/67, cognition eval byte-identical 100/100/100/100;
all 6 ADR-0092..0099 lanes byte-identical:
reviewer_registry 681a2aab..
miner_loop_closure 9f071733..
domain_contract_validation f9c06cde..
fabrication_control sum 01e1b6b7..
demo_composition 27d83824..
public_demo 5707db8e..
DemoCommand Protocol + thin adapters retrofit shipped tours to a
typed composition contract. Composability becomes a structural
property: the ADR-0099 showcase will consume DemoResult through one
stable type rather than special-casing each tour. No demo behavior
changes — adapters wrap underlying run_tour() entry points.
- new core/demos/ package:
- contract.py: frozen Claim / DemoResult dataclasses, runtime-checkable
DemoCommand Protocol, canonical_json() sanctioned serializer
(sorted keys, 2-space indent, trailing newline), CLAIM_CONTRACT_VERSION
- audit_tour_adapter.py: AuditTourDemo (5 claims from ADR-0042 scenes
1-4: identity_pack_swaps_visible, safety_typed_refusal,
ethics_opt_in_deployment_fires, ethics_default_silent,
replay_byte_identical)
- tour_adapters.py: shared pattern for register/anchor-lens/orthogonality
tours; _extract_claims walks the dict tree for *_supported booleans
and builds Claim objects in deterministic sorted order
- global-state-mutation detector (ADR-0098 invariant #2):
capture_state() snapshots a load-bearing subset of process state
(CORE_* env vars + module identities for chat.telemetry,
chat.runtime, language_packs.compiler);
verify_no_global_state_mutation() ignores None→id transitions
(benign lazy import) and only flags env-var changes or module
identity rebindings
- new evals/demo_composition/ lane (ADR-0098 invariant proving):
- 6 cases asserting byte-equality + no-state-mutation across the
three fast adapters (audit-tour, register-tour, orthogonality-tour)
- composition_read_only: confirms two adapter results compose into
a composite claim set without mutating either
- stateful_fixture_rejected: negative control — a deliberately
stateful adapter MUST trigger divergence detection
- anchor-lens-tour adapter is exercised by tests, not the lane,
to keep wall time bounded
- byte-identical across runs (sha256 27d838241bf3..)
- 26 unit tests covering Claim/DemoResult validation, canonical_json
determinism, state-mutation detector (including the lazy-import
benign case), Protocol conformance (isinstance check + claim
contract version) for all four adapters, seed-rejection per
adapter (all current adapters are fully deterministic), and an
audit-tour integration smoke verifying 5 claims + byte-equality +
no state mutation across two consecutive runs
- smoke 67/67, cognition eval byte-identical 100/100/100/100, all
five lanes byte-identical (reviewer_registry 681a2aab..,
miner_loop_closure 9f071733.., domain_contract_validation f9c06cde..,
fabrication_control summary 01e1b6b7.., demo_composition 27d83824..)
First concrete domain claim under ADR-0091's Domain Pack Contract v1.
en_mathematics_logic_v1 is now formally ratified as reasoning-capable
in the capability ledger: 9/9 ADR-0091 predicates pass.
ADR-0097 §"No code changes outside pack artifacts and corpus" relaxed
to include two latent bug fixes that ADR-0093's predicate enforcement
just exposed:
1. language_packs/schema.py: LanguageRole enum widened to include
DOMAIN_SEED. Three in-tree packs (en_mathematics_logic_v1,
en_physics_v1, en_systems_software_v1) have declared role="domain_seed"
since landing but the enum was never updated; load_pack() always
raised on them. ADR-0093's P1 predicate exposed the mismatch.
2. core/capability/domain_contract_predicates.py: P2 (gloss checksum)
was reading manifest["checksums"]["glosses_sha256"]; the canonical
in-tree location is manifest["glosses_checksum"] (top-level). Fixed
to prefer the canonical key and fall back to the nested form for
forward compatibility.
ADR-0097 manifest additions to en_mathematics_logic_v1:
- domain_contract_version: 1
- domain_id: "mathematics_logic"
- axioms: null (rules in v1 — pack proves reasoning via chain
composition, not declarative axioms)
- rules: null
- teaching_chains: ["mathematics_logic_chains_v1"]
- eval_lanes: three lanes with dev/public/holdout (elementary_mathematics_ood,
inference_closure, fabrication_control)
- reviewers: ["shay-j"] (resolved via ADR-0092 registry)
- known_gaps: [] (all math/logic gaps in docs/gaps.md were [x])
- provenance: "adr-0097:reviewed:2026-05-21"
Verified evidence:
- core capability domain-contract --pack-id en_mathematics_logic_v1
→ all_passed=True (P1-P9 all pass)
- core capability ledger → mathematics_logic row shows
status=reasoning-capable, predicates.reasoning_capable=True,
predicates.expert_demo=False, open_gaps=[],
operator_chain_coverage all ready=True (8 chains each),
intent_shapes_present=5
- 14 ADR-0097 invariant tests in
test_adr_0097_mathematics_logic_ratification.py pin
status/provenance/expert-demo-gate/contract shape
Two pre-existing tests updated for the new CLI default
(predicate-running, non-zero on missing contract):
- test_capability_domain_contract_json_absent_contract_is_noop now
uses --structural-only to assert legacy parse-only shape
- test_cli_returns_nonzero_on_missing_contract switched its fixture
pack from en_mathematics_logic_v1 (now has a contract) to
en_core_cognition_v1 (no contract)
The pre-existing test_flag_report_tracks_default_off_flags failure
(discourse_planner flag default mismatch, seen since ADR-0092) is
unchanged and unrelated.
Smoke 67/67, packs 6/6, capability tests 49/50, cognition eval
byte-identical 100/100/100/100; lanes byte-identical:
reviewer_registry 6/6, miner_loop_closure 6/6,
domain_contract_validation 9/9, fabrication_control dev 12/12 +
public 9/9.
First negative-control measure. Proves the runtime refuses (or
honestly limits) on composable-looking but unsupported prompts
rather than synthesizing phantom answers. Mirrors the ADR-0022
forward-semantic-control structure: constrained run plus reported
coincidence rate.
- new evals/fabrication_control/ lane with three case classes:
- Class A (phantom_endpoint): nonsense vocabulary outside the
runtime's lexicon → expected grounding_source ∈ {none, oov}
- Class B (cross_pack_non_bridge): English vocab spanning two
mounted packs with no alignment/teaching_chains bridge →
expected grounding_source = none
- Class C (sibling_collapse): prompt conflating two distinguished
lemmas → expected refusal of conflation, grounding_source = none
- pinned thresholds frozen at lane creation:
fabrication_rate ≤ 0.01, refusal_recall ≥ 0.95,
trace_evidence_present == 1.00,
grounding_source_matches_expected == 1.00
- three-set discipline per docs/capability_roadmap.md Rule 1:
cases/dev.jsonl (12 cases, 4/class), cases/public.jsonl (9 cases),
cases/holdout.jsonl (empty — reserved for first version cut)
- runner.py drives each case through ChatRuntime.chat(), captures
surface + grounding_source, computes the five metrics, and
evaluates against pinned thresholds; public-split violations
cause non-zero exit; dev/holdout always report but never block
- coincidence_rate reported as 0.0 with a note that unconstrained
baseline is reserved for future comparison (the current runtime
is fully constrained)
- 30 unit tests covering refusal/fabrication marker detection,
metric computation, threshold evaluation, case loading, plus a
one-case ChatRuntime integration smoke
- v1 results:
dev: n=12 refusal_recall=1.0 fabrication_rate=0.0 PASSED
public: n=9 refusal_recall=1.0 fabrication_rate=0.0 PASSED
- byte-identical across runs (dev sha256=d6757e0e3f96..,
public sha256=9b502878fcb7.., summary sha256=01e1b6b71114..)
- smoke 67/67, teaching 17/17, cognition 120/121 (pre-existing skip);
cognition eval byte-identical 100/100/100/100
Closes the Phase-5 contemplation loop in code. Articulation-quality,
contradiction-detection, and frontier-compare miners (already shipping)
now have a route to file PackMutationProposal candidates that traverse
the single reviewed teaching path. Construction-only; never promotes
to coherent.
- new teaching/from_miner.py: from_finding() / from_findings() turn
ContemplationFinding records (kind=PACK_MUTATION_CANDIDATE) into
PackMutationProposal candidates with source.kind="miner",
source.source_id=<miner_id>, status=SPECULATIVE
- proposal_id = SHA-256(canonical(miner_id, finding, revision))[:16]
— same inputs → byte-identical proposal_id; different miner_id or
revision → different id
- identity-pack defense AT CONSTRUCTION: reuses teaching.review.
_is_identity_override() against finding.subject AND
finding.proposed_action; miner-sourced identity-override attempts
never reach the proposal log
- pluggable ReplayEquivalenceChecker Protocol with ReplayEquivalenceResult;
NoOpReplayChecker default explicitly notes "deferred to production
checker"; production checker integration is downstream of this ADR
- from_findings() batch path collects identity-override and
replay-equivalence rejections in a typed rejection log rather than
raising, so a mixed batch can proceed with audit evidence
- serialize_proposal_emitted_event() emits ADR-0040-compliant redacted
telemetry shape: type, proposal_id, source.serialize(),
epistemic_status only (no raw subject/correction_text)
- 22 unit tests covering positive construction, identity defense in
subject+proposed_action, malformed input, determinism (same inputs,
different revision, different miner_id, batch stream), replay
pre-gate (single + batch), telemetry redaction, and the structural
grep gate enforcing miner_proposal_single_review_path (only
teaching/review.py and teaching/store.py may promote to COHERENT)
- new evals/miner_loop_closure/ lane: 6 case classes (positive_basic,
identity_override_subject, identity_override_action,
replay_equivalence_failed, wrong_finding_kind, determinism) passing
6/6 with byte-identical SHA-256 across runs
- smoke 67/67, teaching 17/17, cognition 120/121 (1 pre-existing skip);
cognition eval byte-identical 100/100/100/100
Closes the load-bearing gap blocking every reasoning-capable claim
under ADR-0091: docs/reviewers.yaml was previously `reviewers: []` and
unparsed. Now schema-validated at v1, with a bootstrap shay-j entry
self-sealed via provenance.
- new core.capability.reviewers module: frozen Reviewer/ReviewerRegistry
dataclasses, strict load_reviewer_registry parser, ReviewerRegistryError
- enforces ADR-0092 schema rules: schema_version==1, no unknown
top-level keys, no unknown reviewer fields, role∈{primary,domain},
primary must claim ["*"], domain must NOT claim "*", review_scope
subset of {pack,proposal,chain,eval}, no duplicate reviewer_ids
- can_review(reviewer_id, domain_id, scope) helper implements
ADR-0092 rules 2-4 for downstream use by ADR-0093 validator
- docs/reviewers.yaml updated to v1 schema with shay-j bootstrap
- ledger_report() evidence_counts now exposes structured
reviewer_registry status (valid, schema_version, reviewer_count,
reviewer_ids, error) alongside the legacy reviewers_present bool
- new evals/reviewer_registry/ lane: 6 cases (2 positive + 4 negative)
covering empty-registry, wrong-version, domain-wildcard rejection,
and unknown-field rejection
- runner emits deterministic JSON report; two runs produce byte-identical
output (sha256 verified)
- 26 unit tests in tests/test_reviewer_registry.py
- capability ledger test extended to assert new reviewer_registry block
- smoke suite green (67/67); lane passes 6/6
The pre-existing test_flag_report_tracks_default_off_flags failure is
unrelated (discourse_planner flag default) and not introduced here.
Final phase of the articulation arc. Consumes the per-turn
``PlanMetrics`` + ``ContemplationFinding`` streams produced by
Phases 3 + 4 and aggregates across many turns to emit
SPECULATIVE ``PACK_MUTATION_CANDIDATE`` findings that the operator
reviews via the existing proposal-review-ratify chain.
This is the doctrine-aligned answer to the user's question:
"Should we... realize a way to score whether it should use what
it produced towards memory confidence for future use?"
Yes — and it stays inside ADR-0080: read-only, SPECULATIVE-only,
deterministic, no parallel learning path, no autonomous memory
mutation.
What it adds
------------
* New module ``chat/articulation_telemetry.py``:
- ``ArticulationObservation`` frozen dataclass — per-turn
bundle of (turn_id, anchor_subject, prompt_hash,
plan_substrate_hash, metrics, findings).
- ``format_articulation_observation_jsonl(...)`` — deterministic
sort-keys JSONL line.
- ``load_articulation_observations(lines)`` — schema-tolerant
loader; malformed lines drop without aborting.
- ``ArticulationObservationSink`` protocol — structurally
identical to ``TurnEventSink`` but distinct named type so
consumers can subscribe to one stream without the other.
* New module ``core/contemplation/miners/articulation_quality.py``:
- ``mine_articulation_observations(observations, paths)`` —
pure deterministic aggregator with three v1 rules.
- **recurring_predicate_monotony** — when the same
(subject, predicate) pair is flagged WEAK_SURFACE in
>= _MIN_RECURRENCE (default 3) observations, propose
substrate diversification with non-dominant predicates.
- **recurring_planner_gap** — when the same subject is
flagged PLANNER_GAP >= _MIN_RECURRENCE times across modes,
propose substrate expansion.
- **low_average_predicate_diversity** — when mean
``predicate_diversity_ratio`` < 0.5 across >= _MIN_RECURRENCE
observations on the same anchor subject, propose
diversification.
* Runtime wiring (``chat/runtime.py``):
- New ``ChatRuntime.attach_articulation_sink(sink)`` method.
Mirrors ``attach_telemetry_sink`` pattern.
- Emission point at the end of
``_maybe_apply_discourse_planner``: when contemplation
enabled + sink attached + plan engaged, builds an
``ArticulationObservation`` and emits one JSONL line.
Sink errors propagate (fail-fast, no swallowing).
- Per-runtime ``_articulation_turn_counter`` increments on
every emission; gives downstream consumers a stable
sequence index.
Tests
-----
* ``tests/test_articulation_quality_miner.py`` (11 tests):
- Empty / sub-threshold cases yield no findings.
- Each of the three rules fires at threshold.
- Recurring_predicate_monotony separates by subject (no
cross-subject merging).
- Recurring_planner_gap collects distinct modes into a
sorted comma-joined string.
- Determinism — byte-equal finding IDs across two runs.
- SPECULATIVE doctrine pin.
- JSONL round-trip preserves observation identity.
* ``tests/test_articulation_quality_e2e.py`` (7 tests):
- Sink-detached + contemplation-on → no emission.
- Sink-attached + contemplation-off → no emission.
- Engaged turn emits exactly one observation line.
- BRIEF prompt emits nothing (fast-path).
- **Full loop** — run compound prompt 3x → 3 observations →
miner emits PACK_MUTATION_CANDIDATE with subject='truth',
predicate='recurring_predicate_monotony', object='belongs_to'.
- Full loop is deterministic (byte-equal finding IDs across
two complete runs).
- Every full-loop finding is SPECULATIVE.
Doctrine pins
-------------
| Claim | Pinned by |
|--------------------------------------|----------------------------------------------------------|
| SPECULATIVE-only | test_all_findings_remain_speculative |
| Deterministic across runs | test_miner_is_deterministic_across_runs |
| Full-loop determinism (e2e) | test_full_loop_is_deterministic_byte_equal_finding_ids |
| No autonomous mutation | Sink is append-only; miner outputs ContemplationFinding |
| | objects only; nothing writes to packs/vault/teaching. |
| Append-only stream | Sink protocol has emit(line: str) and nothing else. |
Live demo (3 identical compound-prompt turns)
---------------------------------------------
Runtime emits 3 observations. Offline miner aggregates and emits:
[pack_mutation_candidate] subject='truth'
predicate='recurring_predicate_monotony' object='belongs_to'
evidence_refs: 3 observations
proposed_action: "diversify substrate for 'truth': across 3
observations the plan repeatedly over-concentrated on
predicate 'belongs_to'. Candidates: add teaching chains
rooted on 'truth' with relations OTHER than 'belongs_to'
(grounds / requires / reveals / contrasts / precedes /
follows) so the planner's RELATION selector has more
variety to draw from."
epistemic_status: speculative
The system observed its own articulation patterns across many
turns, identified the corpus expansion priority, and emitted a
specific reviewable proposal — without mutating anything. The
operator decides whether to act on it via the existing review
chain.
Verification
------------
pytest test_articulation_quality_miner.py 11/11 pass
pytest test_articulation_quality_e2e.py 7/7 pass
pytest test_plan_metrics*.py 18/18 pass (Phase 4)
pytest test_plan_contemplation*.py 17/17 pass (Phase 3)
pytest test_discourse_planner_*.py 99/99 pass
pytest test_articulation_demo.py all claims supported
pytest test_narrative_example_intents.py pass
core test --suite smoke 67/67 pass
core test --suite runtime 19/19 pass
The articulation arc is complete. Future work documented in
``docs/sessions/SESSION-2026-05-21-articulation-arc.md`` §8:
connective rotation, generalised pronoun selection, doctrine-gated
plan revision, Phase 2.5 mid-sentence reflection. None blocking.
Quantitative companion to Phase 3 (commit 664e081). Where Phase 3
emits SPECULATIVE *findings* about plan quality, Phase 4 emits
typed *measurements* — pure-function projection of a
``DiscoursePlan`` into a ``PlanMetrics`` dataclass.
Why this matters
----------------
The discourse planner now produces multi-clause grounded
articulations (Phase 1), the renderer pronominalizes across
consecutive same-subject moves (Phase 2), and the contemplation
pre-flight emits qualitative concerns about plan shape (Phase 3).
What was missing was the *aggregable* layer: per-turn structured
numbers that downstream consumers can stream across many turns
to score quality patterns the per-turn observer cannot see.
Phase 4 lands that layer. Phase 5 (offline contemplation miner)
becomes possible because there's now structured signal to mine.
What it measures
----------------
Structure
* move_count — total moves in plan
* fact_bearing_count — moves with fact != None
Move-kind distribution
* anchor_count / support_count / relation_count
/ transition_count / closure_count
Diversity
* unique_predicates — distinct predicates across
fact-bearing moves
* unique_subjects — distinct subject lemmas
* unique_sources — distinct FactSources
Topic dynamics
* topic_shift_count — consecutive pairs where
subject changed
* pronominalization_opportunities — consecutive pairs where
subject held (= Phase 2's
anaphora trigger count)
Derived ratios
* predicate_diversity_ratio — unique_predicates /
fact_bearing_count
* subject_focus_ratio — pronominalizations /
(pronominalizations +
topic_shifts)
Every field is a deterministic pure function of the plan: same
plan in → byte-equal ``PlanMetrics.as_dict()`` out. This is the
load-bearing claim that lets Phase 5 aggregate across turns
without "is this the same metric?" ambiguity.
Doctrine alignment
------------------
Per ADR-0080 contemplation discipline:
* Read-only — metrics are pure projections of the plan; no
mutation of plan, runtime state, or memory tiers.
* No autonomous learning — metrics are observations, not
learned policy. Promotion to memory still flows through
the existing proposal-review-ratify chain.
* Deterministic replay — pinned by test_metrics_are_deterministic_
and_byte_equal_as_dict plus the runtime-level
test_metrics_byte_equal_across_runs.
Wiring
------
* New ``ChatRuntime.last_plan_metrics`` property — read-only
``PlanMetrics`` from the most recent turn where the planner
engaged (and ``discourse_contemplation`` was on); ``None``
otherwise. Reset between turns alongside ``last_plan_findings``
via the existing top-of-call reset block.
* Same opt-in flag as Phase 3 (``discourse_contemplation``).
When True, the runtime computes both findings AND metrics in
the same block; when False (default), both stay at empty/None.
Demo (config: discourse_contemplation=True)
-------------------------------------------
"What is knowledge?" → metrics: None (BRIEF fast-path)
"Tell me about memory." → moves=3 fact_bearing=3
kinds=A:1/S:1/R:1/T:0/C:0
unique_predicates=3 subjects=1
pronominalization_ops=2 shifts=0
predicate_diversity=1.000
subject_focus=1.000
"What is truth, and why does
it matter?" → moves=7 fact_bearing=6
kinds=A:2/S:2/R:2/T:1/C:0
unique_predicates=4 subjects=1
pronominalization_ops=4 shifts=1
predicate_diversity=0.667 ← Phase 3
WEAK_SURFACE
quantified
subject_focus=0.800
+ 1 finding (weak_surface)
The compound-prompt numbers are particularly informative:
``predicate_diversity=0.667`` is the algebraic expression of the
Phase 3 ``WEAK_SURFACE`` rule — the rule fires precisely because
6 fact-bearing moves used only 4 distinct predicates.
``subject_focus=0.800`` quantifies that 80% of consecutive pairs
held the same subject — high topic stickiness that Phase 2's
reflective renderer leveraged into 4 ``it`` substitutions.
Tests
-----
* ``tests/test_plan_metrics.py`` — 10 unit tests pinning each
field, derived ratios, bridge-move handling (``fact=None``
resets the focus channel), and determinism via ``as_dict()``
byte-equality.
* ``tests/test_plan_metrics_runtime.py`` — 8 end-to-end tests
proving the runtime wiring: disabled by default, populated
when enabled, BRIEF prompts yield None, no cross-turn leak,
byte-equal across runs, parametrized co-population check
alongside findings.
Verification
------------
pytest tests/test_plan_metrics*.py 18/18 pass
pytest tests/test_plan_contemplation*.py 17/17 pass (Phase 3)
pytest tests/test_discourse_planner_*.py 99/99 pass
pytest tests/test_articulation_demo.py all claims supported
pytest tests/test_narrative_example_intents.py pass
pytest tests/test_runtime_config.py pass
cognition eval OFF vs ON 45/45 surface byte-equal
45/45 trace_hash byte-equal
4/4 aggregate metrics
identical
core test --suite smoke 67/67 pass
core test --suite runtime 19/19 pass
Phase 5 (logged, not built)
---------------------------
Offline contemplation miner that consumes ``last_plan_findings``
+ ``last_plan_metrics`` streams across many turns and emits
reviewable pack-mutation candidates. Still SPECULATIVE;
review-gated; never auto-promoted to memory. Now unblocked by
the structured metric surface Phase 4 lands.
Wires deterministic, read-only contemplation OVER a completed
``DiscoursePlan`` BEFORE the renderer fires. This is the
"reasoning at meaningful checkpoints" capability — the system
now inspects the global shape of its own articulation plan and
emits SPECULATIVE findings about quality issues the move-by-move
planner couldn't see locally.
Doctrine alignment (ADR-0080)
-----------------------------
* **Read-only** — never mutates the plan, packs, vault, teaching
corpus, or runtime state. Returns findings as a tuple; the
runtime stores them on a read-only property.
* **SPECULATIVE-only** — every finding is stamped
``EpistemicStatus.SPECULATIVE`` by the schema's ``__post_init__``;
the doctrine pin ``test_findings_always_speculative`` keeps that
invariant visible.
* **Deterministic replay** — same plan → byte-identical findings
(same ``substrate_hash``, same ``finding_id``).
* **No parallel learning path** — findings flow to a read-only
observation surface (``runtime.last_plan_findings``). Promotion
to memory still goes through the existing proposal → review →
ratify chain. The offline contemplation miner (Phase 5 target)
is what eventually consumes the findings and emits reviewable
pack-mutation candidates.
v1 rules (``core/contemplation/plan_preflight.py``)
----------------------------------------------------
* ``PLANNER_GAP`` — non-BRIEF mode produced anchor-only depth.
Signals the teaching/cross-pack substrate for that lemma is too
thin for the planner to expand.
* ``WEAK_SURFACE`` — three or more moves share a predicate.
Signals the rendered surface will read mechanical (e.g. three
``belongs_to`` clauses in a row). Fires on today's compound
prompt ``"What is truth, and why does it matter?"`` — the
6-sentence plan uses ``belongs_to`` 3 times.
* ``COVERAGE_GAP`` — every move in a multi-move plan draws from
a single ``FactSource``. Signals one-sided substrate (e.g.
pack-only with no teaching enrichment).
Runtime wiring
--------------
* New ``RuntimeConfig.discourse_contemplation: bool = False`` —
opt-in for now. Default off keeps the cognition eval byte-
identical to Phase 2 (verified 45/45 surface + 45/45 trace_hash).
* New ``ChatRuntime.last_plan_findings`` property — read-only tuple
of ``ContemplationFinding`` records from the most recent turn.
Reset to ``()`` at the start of every plan-engagement call so
findings never leak across turns.
* Contemplation runs AFTER the planner produces a multi-move plan
and BEFORE the renderer fires; the plan itself is not modified.
Demo (config: discourse_contemplation=True)
-------------------------------------------
"What is knowledge?" → planner fast-path; no findings
"Tell me about memory." → 3 moves, distinct predicates;
no findings (good!)
"What is truth, and why does
it matter?" → 6 moves, ``belongs_to`` x 3:
[WEAK_SURFACE] subject='truth'
predicate='predicate_repeats_in_plan'
object='belongs_to'
proposed action: diversify the
relation inventory for 'truth'
(grounds / requires / reveals /
contrasts) so the planner has
more variety to draw from.
"Explain truth." → 3 moves, distinct predicates;
no findings
Tests
-----
* ``tests/test_plan_contemplation.py`` — 11 unit tests pinning
each rule, empty/trivial plans, determinism, and the
SPECULATIVE-only doctrine.
* ``tests/test_plan_contemplation_runtime.py`` — 6 end-to-end
tests proving the runtime wiring: disabled by default,
populated when enabled, reset across turns, deterministic
across runs, all findings SPECULATIVE.
Verification
------------
pytest tests/test_plan_contemplation*.py 17/17 pass
pytest tests/test_discourse_planner_*.py 99/99 pass
pytest tests/test_articulation_demo.py all claims supported
pytest tests/test_narrative_example_intents.py pass
pytest tests/test_runtime_config.py pass
cognition eval OFF vs ON 45/45 surface byte-equal
45/45 trace_hash byte-equal
4/4 aggregate metrics
identical
core test --suite smoke 67/67 pass
core test --suite runtime 19/19 pass
Phases roadmap (logged in commit, not built today)
--------------------------------------------------
* Phase 4 — articulation telemetry enrichment. Emit per-turn
metrics (grounding_ratio, anaphora_engagement, plan_completeness,
novelty, focus_consistency) to the existing telemetry sink so
the offline miner has structured signal.
* Phase 5 — offline contemplation miner. Extend
``core/contemplation`` with a miner that consumes
``last_plan_findings`` streams and emits reviewable
pack-mutation / teaching-corpus expansion proposals. Still
SPECULATIVE; review-gated.
The Phase 1 multi-clause renderer (commit 63ffd88) produces grounded
content but reads mechanically because the subject lemma repeats in
every clause:
"Truth is what is true. Furthermore, truth belongs to cognition.truth.
In turn, truth grounds knowledge. Truth belongs to epistemic.ground.
Furthermore, truth belongs to logos.core. In turn, truth requires
evidence."
This is the literal articulation gap that motivated Phase 2 —
"reasoning at meaningful checkpoints during sentence construction
in order to have a stronger idea of what has come prior and is
already done to help better inform the next move." Between move
``i`` and move ``i+1`` the renderer now reflects on what subject
has just been established (the "focus") and renders the next clause
with a pronoun when the focus carries forward:
"Truth is what is true. Furthermore, it belongs to cognition.truth.
In turn, it grounds knowledge. It belongs to epistemic.ground.
Furthermore, it belongs to logos.core. In turn, it requires
evidence."
Rules
-----
* Track ``focus_subject`` across moves (the lemma most recently used
as a fact subject).
* When the next move's ``fact.subject`` is byte-equal to the current
focus → swap subject token to ``"it"``.
* When the next move's subject differs → preserve the explicit lemma
AND update focus. Topic shifts (TRANSITION moves; compound bridge
TRANSITION) thus reset the pronominalization channel naturally.
* Sentence-initial position (no connective): capitalised ``"It"``.
* Mid-sentence (after connective + comma): lowercase ``"it"``.
Doctrine alignment
------------------
Pure deterministic transformation of the existing plan; no new
content introduced, no LLM, no stochastic sampling. Same plan in →
same surface out, always. trace_hash invariance holds because:
* BRIEF-mode prompts short-circuit the planner before render
(commit 63ffd88's fast path) and are unaffected.
* Multi-move plans render to a deterministically-different string
that compute_trace_hash already folds in via ``surface``.
Wiring
------
* New ``reflective: bool = False`` parameter on ``render_plan``
(back-compat default — every existing call site and test pinning
Phase 1 output continues to work).
* ``_clause_for`` gains optional ``prior_focus_subject`` arg used by
the reflective path; unchanged default behaviour.
* Runtime hook ``chat.runtime._maybe_apply_discourse_planner``
passes ``reflective=True`` so the default chat path benefits.
Tests
-----
New ``tests/test_discourse_planner_reflective.py``:
* ``test_reflective_replaces_repeated_subject_with_it``
* ``test_reflective_handles_three_consecutive_same_subject_moves``
* ``test_reflective_capitalises_sentence_initial_pronoun``
* ``test_reflective_resets_focus_on_topic_shift``
* ``test_reflective_off_preserves_phase1_output``
* ``test_reflective_default_is_off_for_back_compat``
* ``test_reflective_is_deterministic``
* ``test_reflective_single_move_byte_identical_to_non_reflective``
(load-bearing — pins that the cognition eval stays byte-equal
across the Phase 2 flip because every cognition case is single-
move).
Verification
------------
pytest tests/test_discourse_planner_*.py 99/99 pass
(91 existing + 8 new)
pytest tests/test_articulation_demo.py all claims supported
pytest tests/test_narrative_example_intents.py pass
pytest tests/test_runtime_config.py pass
cognition eval OFF vs ON 45/45 surface byte-equal
45/45 trace_hash byte-equal
4/4 aggregate metrics
identical
core test --suite smoke 67/67 pass
core test --suite runtime 19/19 pass
Live demo (default config):
"What is knowledge?" → unchanged (BRIEF, fast-path)
"Tell me about
memory." → "Memory is what a person recalls.
Furthermore, it belongs to cognition.memory.
In turn, it requires recall."
"What is truth, and
why does it matter?"→ "Truth is what is true. Furthermore, it
belongs to cognition.truth. In turn, it
grounds knowledge. It belongs to
epistemic.ground. Furthermore, it belongs
to logos.core. In turn, it requires
evidence."
"Explain truth." → "Truth is what is true. Furthermore, it
belongs to cognition.truth. In turn, it
grounds knowledge."
Out of scope for this commit (future Phase 2 follow-ons):
* Connective rotation ("Furthermore" → "Also" → "In addition"
to break the repetitive cascade).
* Cross-clause de-duplication (skip moves whose ``new`` lemmas
were already introduced by an earlier move).
* Generalised pronoun selection beyond ``it`` (requires gender /
number / animacy signals the pack lexicon doesn't carry today).
Flips ``RuntimeConfig.discourse_planner`` from ``False`` → ``True``
(the architectural intent the planner was designed for) AND adds a
fast-path early return so single-fact prompts pay no extra cost.
Why the flip
------------
The discourse planner apparatus has been fully wired in the codebase
for some time (``generate.discourse_planner.plan_discourse`` /
``plan_compound_discourse`` / ``render_plan``,
``generate.grounding_accessors.grounding_bundle_for``,
``chat.runtime._maybe_apply_discourse_planner``) but gated off behind
this flag. Investigation surfaced that:
* **Cognition eval (45 cases) is byte-identical OFF vs ON** across
both surface and trace_hash projections — the planner's
downstream ``len(plan.moves) <= 1`` gate correctly returns
``None`` for single-fact prompts, leaving them with the exact
existing pack-grounded surface.
* **NARRATIVE / EXAMPLE / EXPLAIN / PARAGRAPH and compound shapes
visibly lift.** ``"Tell me about memory."`` goes from a one-
fragment disclosure to a 3-sentence grounded discourse.
``"What is truth, and why does it matter?"`` — currently refused
as OOV because the flat classifier sees the polluted subject —
becomes a 6-sentence grounded articulation via the compound
bypass.
* **No quality regression on existing benches.** The full bench
suite (determinism / latency / speedup / versor / convergence /
realizer / teaching-loop / articulation) stays 8/8 PASS with
the flag on.
Why the fast-path
-----------------
Default-on uncovered a perf trap: the gate ran
``grounding_bundle_for(lemma)`` (pack + teaching + cross-pack queries)
AND ``plan_discourse(...)`` on EVERY turn, then discarded the
result when ``len(plan.moves) <= 1``. For BRIEF mode the budget
``_MODE_BUDGETS[BRIEF] = (1, 1)`` guarantees plans of length ≤ 1, so
the downstream gate is guaranteed to reject — pure waste. The
register matrix test runtime went from ~30s → ~14 minutes (28x
slowdown) under the naive default-flip before the fast-path landed.
The new short-circuit:
if mode is BRIEF and not compound.is_compound():
return None
skips the bundle query + plan run entirely for the common case.
Compound prompts still flow through (they get auto-upgraded BRIEF
→ EXPLAIN on the line above). Empirical post-fast-path
measurement on a 45-case eval (workers=1):
OFF: 23.31s (1.93 turns/sec)
ON : 17.74s (2.54 turns/sec)
slowdown : 0.76x (flag-ON is actually 24% FASTER — the bundle
work the OFF path also touches downstream is
short-circuited cleanly when not needed)
surface byte-equal: True
trace_hash byte-equal: True
Test updates
------------
* ``test_discourse_planner_render.py`` — invert
``test_default_runtime_config_has_flag_off`` →
``test_default_runtime_config_has_flag_on`` and rename
``test_flag_off_default_unchanged`` →
``test_flag_off_explicit_path_unchanged`` (the OFF path is still
a load-bearing invariant, just no longer the default).
* ``test_narrative_example_intents.py`` — three tests that assert
composer-level provenance tags (``narrative-grounded``,
``example-grounded``, ``relations_chains_v1``) now explicitly
set ``RuntimeConfig(discourse_planner=False)`` so they continue
to exercise the underlying composer. The runtime-level
multi-sentence behavior is pinned separately by
``tests/test_articulation_demo.py``.
Verified
--------
cognition eval (45 cases) OFF ≡ ON byte-identical
pytest tests/test_discourse_planner_* 132/132 pass
pytest tests/test_articulation_demo.py all claims supported
pytest tests/test_narrative_example_intents.py pass
pytest tests/test_runtime_config.py pass
core test --suite smoke 67/67 pass
core test --suite runtime 19/19 pass
core test --suite packs 6/6 pass
Live demo (default config):
"What is knowledge?" → single sentence (BRIEF, fast-path)
"Tell me about memory." → 3 grounded sentences
"What is truth, and why does
it matter?" → 6 grounded sentences (was: OOV)
"Explain truth." → 3 grounded sentences
Follow-on to the word-boundary fix (commit 0dd30b8). After tightening
``\bno\b`` etc. with word boundaries, an audit surfaced a separate
pre-existing gap in the CORRECTION trigger: the contracted-only
``that'?s\s+(?:not|wrong)`` slot silently dropped every fully-spoken
copula form to UNKNOWN.
Concrete gap (every one previously UNKNOWN):
"That is not right." → UNKNOWN
"That is wrong." → UNKNOWN
"That was wrong." → UNKNOWN
"That is incorrect." → UNKNOWN
"That is false." → UNKNOWN
"That was not right." → UNKNOWN
"that is mistaken." → UNKNOWN
"That was incorrect." → UNKNOWN
Root cause: the slot ``that'?s\s+(?:not|wrong)`` matches only
that's / thats
— ``'?s`` makes the apostrophe optional but the literal ``s`` is
mandatory. ``that is`` (full word ``is``) and ``that was`` (full
word ``was``) had no path. And the predicate alternation only
accepted ``not`` or ``wrong``; ``incorrect``, ``false``, and
``mistaken`` were also missing.
Fix: widen both slots in one pattern revision.
Before:
that'?s\s+(?:not|wrong)
After:
that(?:'?s|\s+(?:is|was))\s+(?:not|wrong|incorrect|false|mistaken)
The full pattern now reads:
\b(?:no
|that(?:'?s|\s+(?:is|was))\s+(?:not|wrong|incorrect|false|mistaken)
|incorrect
|actually
|correction)\b
Boundary discipline holds: the outer ``\b...\b`` still prevents the
predicate alternation from eating into longer words. Verified:
"That is correct." → UNKNOWN (right NOT in predicate set)
"That is right." → UNKNOWN (right NOT in predicate set)
"That is true." → UNKNOWN (true NOT in predicate set)
"That works." → UNKNOWN
"That is interesting." → UNKNOWN
"That is falsifiable." → UNKNOWN (``false`` + ``i`` is word→word
so ``\b`` after ``false`` fails)
"That was wrongly accused." → UNKNOWN (same logic for ``wrong``+``ly``)
Tests extended:
* ``test_correction_canonical_forms_still_route`` — 8 new parametrize
cases for the fully-spoken copula forms
* ``test_correction_does_not_eat_no_prefixed_words`` — 9 new
parametrize cases for the affirmative ``That is/was ...`` shape
AND the boundary-trap cases ``falsifiable`` / ``wrongly accused``
Verified:
pytest tests/test_intent_subject_extraction.py 33/33 pass
full intent + register-diagnostic + proposition graph 77/77 pass
core test --suite smoke 67/67 pass
core test --suite runtime 19/19 pass
While investigating the adjacent RECALL classifier gap, a much
wider intent-classification bug surfaced: every prompt beginning
with a word that *starts with* the letters of any CORRECTION
trigger silently routed to CORRECTION with a mangled subject.
Concrete examples seen during diagnosis:
"Now remember light." → CORRECTION subject="w remember light"
"Nothing matters." → CORRECTION subject="thing matters"
"Notice the truth." → CORRECTION subject="tice the truth"
"Note that recall fires." → CORRECTION subject="te that recall fires"
"Nominate a candidate." → CORRECTION subject="minate a candidate"
"Norma is here." → CORRECTION subject="rma is here"
"Notwithstanding ..." → CORRECTION subject="twithstanding ..."
Root cause: ``generate/intent.py`` ``_RULES`` line ~213 used the
pattern
(?:no|that'?s\s+(?:not|wrong)|incorrect|actually|correction)
The alternation has ``no``, ``incorrect``, ``actually``, ``correction``
as bare substrings — no word boundary on either side. Combined with
``re.match``'s start-of-string anchor, *any* prompt beginning with
``No``-, ``Incorrect``-, ``Actually``-, or ``Correction``-prefixed
text matched as CORRECTION; the regex's match span was then sliced
off the prompt to produce a subject like ``"w remember light"``
(from ``"Now remember light."``).
The same hazard threatens:
* ``no`` → eats ``Now`` / ``Notice`` / ``Note`` / ``Nothing`` /
``Nominate`` / ``Norma`` / ``Notwithstanding`` / ...
* ``incorrect`` → would eat ``incorrectly``
* ``actually`` → would eat ``actualization``
* ``correction`` → would eat ``corrections``
Fix: add ``\b`` anchors on both sides of the alternation.
\b(?:no|that'?s\s+(?:not|wrong)|incorrect|actually|correction)\b
``\b`` is zero-width, so ``re.match``'s start-of-string anchor still
holds; the left ``\b`` is a no-op at position 0. The right ``\b``
forces the matched token to end on a word boundary — i.e., the next
character must be non-word (whitespace, punctuation, EOL) — so
``\bno\b`` matches ``"No."`` / ``"No way"`` / ``"No, ..."`` but NOT
``"Now"`` / ``"Nothing"`` / etc.
Verified 11/11 previously-misfiring prompts now correctly classify
as UNKNOWN, and 8/8 legitimate CORRECTION pragmas
(``"No."`` / ``"No way."`` / ``"Incorrect."`` / ``"Actually, ..."`` /
``"Correction: ..."`` / ``"That's wrong."`` / ``"No, that's wrong."`` /
``"no, knowledge is wrong."``) still route correctly.
Tests extended with two new parametrized blocks in
``tests/test_intent_subject_extraction.py``:
* ``test_correction_canonical_forms_still_route`` — 8 cases pinning
the legitimate CORRECTION patterns
* ``test_correction_does_not_eat_no_prefixed_words`` — 10 cases
pinning the boundary fix against regression
Verified:
pytest tests/test_intent_subject_extraction.py 25/25 pass
pytest tests/test_intent_proposition_graph.py + others 60/60 pass
core test --suite smoke 67/67 pass
core test --suite runtime 19/19 pass
Out of scope: ``"That is not right."`` (a real CORRECTION pragma the
regex never caught because ``that'?s\s+`` requires literal ``s`` after
``that``; the colloquial ``that is`` form was always UNKNOWN). Separate
gap, unchanged here.
The articulation breadth benchmark surfaced a RECALL intent gap:
Before (bench output):
RECALL UNKNOWN pack Pack-resident tokens — pack-grounded
(en_core_cognition_v1): recall ...
The probe prompt ``"Recall truth."`` classified as UNKNOWN and fell
through to the ADR-0086 pack-resident-token surface — a graceful
degradation, not a hard failure, but a real classifier gap.
Root cause: ``generate/intent.py`` ``_RULES`` line 213 only matched
the imperative ``remember``:
(re.compile(r"remember\s+", re.IGNORECASE), IntentTag.RECALL)
The verb ``recall`` — every bit as natural an imperative — was
missing from the trigger pattern. ``"Remember truth."`` correctly
routed to RECALL; ``"Recall truth."`` did not.
Fix: widen the alternation to ``(?:remember|recall)\s+``. One-word
change; ``re.match`` anchoring at the start of the prompt means the
fix only catches the canonical imperative form, leaving downstream
contexts untouched:
* ``Does memory require recall?`` → VERIFICATION (unchanged;
earlier rule on the aux-verb pattern fires first)
* ``What is recall?`` → DEFINITION (unchanged;
``what\s+is\s+`` fires first)
* ``Why does recall exist?`` → CAUSE (unchanged;
``why\s+`` fires first)
* ``I recall.`` → UNKNOWN (unchanged;
no trailing word after ``recall``, ``\s+`` doesn't match)
* ``Please recall the truth.`` → UNKNOWN (unchanged
— symmetric with ``Please remember the truth.`` since rules use
``pattern.match`` not ``pattern.search``)
After (bench output):
RECALL RECALL pack Truth is what is true. pack-grounded
(en_core_cognition_v1).
The articulation bench probe now routes correctly and produces a
pack-grounded definition surface — the canonical RECALL output on
a pack-resident lemma.
Tests extended: ``tests/test_intent_subject_extraction.py::
test_recall_strips_articles`` is parametrized with four new
``Recall ...`` cases parallel to the existing ``Remember ...``
cases. A regression that re-narrows the trigger pattern fails the
gate immediately.
Verified:
* pytest tests/test_intent_subject_extraction.py 7/7 pass
* pytest tests/test_register_firing_diagnostic.py 3/3 pass
* core test --suite smoke 67/67 pass
* core test --suite runtime 19/19 pass
* core bench --suite articulation → RECALL ✓ pack-grounded
Replaces the per-pack-aggregate diagnostic landed at 58ac780 with a
per-intent matrix decomposition authored by Codex on a parallel
worktree. Codex's design directly answers the original motivating
question — "which packs' marker pools don't fire on which intent
shapes" — that the aggregate version flattened.
What Codex's version adds over the prior aggregate version:
* **Per (pack × intent × prompt) matrix** — cells decompose by
IntentTag. The C_stance / DEFINITION collapse pattern surfaced
in the widened tour is now directly visible as
matrix[register]["DEFINITION"][*].opening_fired == False.
* **Replayed-variant verification** — every cell records
decorate_surface()'s opening/closing AND asserts the resulting
variant_id matches the runtime's emitted register_variant_id
byte-for-byte. Catches future drift between the replayed
selection and live selection in a single field
(variant_id_matches_runtime / all_replayed_variants_match_runtime).
* **Representative-prompt classification gate** — the companion
test confirms every prompt in REPRESENTATIVE_PROMPTS actually
classifies to its declared IntentTag. If intent classification
drifts, the corpus is invalidated immediately rather than
silently producing meaningless diagnostic output.
* **--fail-on-gap CI mode** — exits 1 when any non-empty marker
bucket never fires across its representative-prompt slice.
Convertible into a CI gate once the deliberate-silent vs
accidental-silent distinction is curated.
* **--register / --intent filters** + **--output PATH** — operator
ergonomics for targeted debugging and report archival.
* **3 pytest cases** — corpus integrity, subset-report shape,
full main()/--output round-trip.
Path: Codex authored at scripts/diagnose_register_firing.py.
Relocated to evals/register_diagnostics/run_firing_diagnostic.py to
match the convention used by evals/register_tour/, anchor_lens_tour/,
orthogonality_tour/, learning_loop/ — measurement artifacts live
under evals/, not scripts/. Test import path adjusted accordingly.
The sys.path bootstrap _REPO_ROOT computation was updated from
.parent.parent to .parents[2] to account for the new path depth.
Verified:
PYTHONPATH=. pytest tests/test_register_firing_diagnostic.py -v
→ 3 passed in 5.39s
PYTHONPATH=. python -m evals.register_diagnostics.run_firing_diagnostic \
--register convivial_v1 --intent DEFINITION --intent CAUSE
→ emits per-cell matrix with variant_id_matches_runtime=True
PYTHONPATH=. python -m evals.register_diagnostics.run_firing_diagnostic \
--register expert_v1 --intent DEFINITION --fail-on-gap
→ exit 0 (expert_v1's empty buckets have non_empty_size=0, so
not a contract gap — that's correct: gap = non-empty bucket
whose entries never fire)
Co-authored-by: Codex <noreply@openai.com>
PR #102 ratified 93 drafted register packs, bringing the catalog to
100 fully-sealed packs on disk. This widens
tests/test_cognition_eval_register_matrix.py::_RATIFIED_REGISTERS
from 7 to 100 so every projection-invariant assertion (trace_hash,
intent_correct, terms_captured, surface_contains_pass,
versor_closure, versor_condition, canonical surface, and aggregate
metrics) now runs against every ratified pack.
Verification on PR #102 head: 801 cells passed in 316.76s
= 100 registers × 8 projections + 1 meta-test
= full ADR-0072 invariant proven across the entire register axis
on all 45 cognition cases.
The meta-test test_register_matrix_covers_every_ratified_pack
remains the structural co-evolution guard: any future register pack
ratification must widen both REGISTER_IDS in
scripts/ratify_register_packs.py AND _RATIFIED_REGISTERS here in
the same change, or CI fails fast.
* feat(packs/register): materialise A_depth drafted registers
Lands 3 drafted depth registers, dominated by disclosure-domain count and structural compression/expansion knobs; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path. Also aligns the smoke contract assertion with the current pack-grounded unknown evidence split.
* feat(packs/register): materialise B_tone drafted registers
Lands 15 drafted tone registers, dominated by bounded affective opening and closing marker palettes; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path.
* feat(packs/register): materialise C_stance drafted registers
Lands 11 drafted stance registers, dominated by epistemic posture markers plus light deterministic depth clauses; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path.
* feat(packs/register): materialise D_posture drafted registers
Lands 10 drafted posture registers, dominated by role-shaped marker families for peer, mentor, scholar, practitioner, and related voices; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path.
* feat(packs/register): materialise E_domain drafted registers
Lands 11 drafted domain registers, dominated by academic, executive, technical, legal, scientific, and philosophical marker families with bounded known-key knobs; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path.
* feat(packs/register): materialise F_cultural drafted registers
Lands 12 drafted cultural registers, dominated by plainspoken, diplomatic, classic, contemporary, and lyrical marker palettes; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path.
* feat(packs/register): materialise G_affective drafted registers
Lands 10 drafted affective registers, dominated by cheerful, somber, grave, wry, gentle, and earnest marker families; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path.
* feat(packs/register): materialise H_functional drafted registers
Lands 10 drafted functional registers, dominated by documentary, instructional, persuasive, clarifying, comparing, and exemplifying marker families; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path.
* feat(packs/register): materialise I_composite drafted registers
Lands 11 drafted composite registers, dominated by combined knob and marker families for tutorial, interview, briefing, lecture, memo, story, elegy, epigram, and manifesto voices; the sealed reports keep grounding_source and trace_hash byte-identical to the unregistered path.
Adds tests/test_cognition_eval_register_matrix.py — strict superset
of tests/test_register_invariant_grounding.py (which covered only 4
of the 7 ratified register packs).
Parametrizes over all seven ratified register packs
({default_neutral, terse, precise, convivial, pedagogical, formal,
socratic}_v1) and asserts byte-identity against the unregistered
baseline for every per-case projection the cognition eval reports:
* trace_hash (ADR-0072 truth-path-isolation)
* intent_correct (intent runs upstream of realizer)
* terms_captured (scored off canonical surface)
* surface_contains_pass (scored off canonical surface)
* versor_closure (truth-path field invariant)
* versor_condition (exact float, stronger than closure)
* surface (CognitiveTurnResult.surface is the
pre-decoration canonical the trace
hash consumes; substantive transforms
live on turn_log[-1].surface)
Aggregate metrics on EvalReport are pinned identically: total,
intent_correct, terms_captured, terms_expected, surface_grounded,
versor_closures.
Meta-test test_register_matrix_covers_every_ratified_pack enforces
that _RATIFIED_REGISTERS in this file stays in lockstep with
scripts/ratify_register_packs.py::REGISTER_IDS — so the 93 drafted
register packs in packs/register/_catalog.json cannot ratify into
CI without each one passing the full invariant matrix.
Run: 57 cells (8 projections x 7 registers + 1 meta), 27.7s
sequential across 45 cognition cases per register.
Pre-existing smoke failure (test_chat_response_surface_uses_
articulation_plan in tests/test_runtime_config.py) is the ADR-0086
expected-string test on main; unrelated to this change.
Three load-bearing pieces:
1. ADR-0086 — UNKNOWN-intent pack-resident token surface
New deterministic composer `pack_grounded_unknown_surface` in
chat/pack_grounding.py. When intent classification returns UNKNOWN
but the prompt contains pack-resident lemmas (via cross-pack
resolver), surface those lemmas with their semantic_domains
instead of falling to the bare _UNKNOWN_DOMAIN_SURFACE. Wired
into chat/runtime.py::_maybe_pack_grounded_surface as the
last typed-intent branch before the OOV fallback. Null-lift
invariant pinned: fully-OOV prompts still emit the universal
disclosure byte-identically. Closes four cognition-eval term
misses: unknown_logos_019 (public), unknown_evidence_042 (dev),
unknown_spirit_041 + unknown_word_018 (holdout). Side effect:
evals/results/phase2_pack_measurements.json refusal_rate drops
from 0.25 → 0.125 across all three identity packs (no longer
refusing on these prompts).
2. ADR-0087 — PROCEDURE selector + trailing-clause subject echo
Two coupled changes in chat/pack_grounding.py:
(a) Numeric-determiner downrank in _extract_procedure_topic_lemma:
tokens whose primary semantic_domain starts with
"quantitative.numeric." are demoted; non-numeric resident
candidates always win. So "compare two terms" anchors on
`compare` not `two`.
(b) Trailing clause echoes the full normalized subject_text
rather than just the selected lemma, so OOV head nouns like
"terms" reach the surface even when only the procedure verb
is pack-resident. Closes procedure_compare_011.
3. 100-register catalog
New packs/register/_catalog.json — canonical machine-readable
spec for all 100 registers (7 currently-ratified + 93 drafted)
organized into 9 voice groups (depth/tone/stance/posture/domain/
cultural/affective/functional/composite). Each entry is a
complete production input — realizer_overrides, marker palettes
(openings/transitions/closings), depth_preference, description,
author_notes. All realizer_overrides use only legal keys per
scripts/ratify_register_packs.py::_KNOWN_OVERRIDE_KEYS.
Companion packs/register/CATALOG.md documents the production
loop: materialize → widen REGISTER_IDS → ratify → smoke.
Cognition-eval lifts (all three splits):
public: term_capture 91.7% → 100.0% (+8.3pp)
holdout: term_capture 83.3% → 100.0% (+16.7pp)
dev: term_capture 78.6% → 100.0% (+21.4pp)
surface_groundedness: 100% preserved on all splits
intent_accuracy / versor_closure: 100% preserved on all splits
Tests:
tests/test_pack_grounded_unknown.py — 14 tests (composer
direct + runtime engagement + null-lift invariant)
tests/test_adr_0087_procedure_selector.py — 12 tests (selector
numeric downrank + trailing-clause echo + regression guard)
Existing test suites unaffected — cognition lane 120 passed / 1
skipped both before and after. Full lane net −3 failures vs
pristine main (39 → 36 — none introduced).
Applies the ADR-0085 v2 brief's 16 fluency rows (Pattern A 3sg agreement on
relative-clause verbs + Pattern B plural after quantifier) plus 7 additional
"what a person {VERB}" rows surfaced in live chat probe (`Knowledge is what
a person know` → `knows`, similar for `memory`/`question`/`word`/`answer`/
`response`/`express`). 23 gloss edits total across 5 packs.
The brief had an internal conflict: it forbids atom edits but requires
closure-verifier 0/0, while ADR-0084's verifier enforces
`atoms == content_tokens(gloss)` exactly. Resolved by:
1. Extending `scripts/verify_definitional_closure.py` and the integration
test fixture (`mounted_lex_lemmas` + `production_pool` builders) to
include lexicon `surface` forms in the resolution set — already the
operational meaning of "a lemma in another mounted pack" since
surfaces are canonical inflections of the same lemma.
2. Adding 10 inflected `LexicalEntry` rows across cognition / meta /
action / spatial lexicons (e.g. `surface=knows lemma=know`,
`surface=parts lemma=part`) so morphology-shifted atoms resolve.
Live surface verification (sample 6 prompts):
before after
"what a person know from truth and evidence" -> "...knows from..."
"what a person recall" -> "...recalls"
"relation of part to part" -> "relation of parts to parts"
"way of voice and word" -> "way of voice and words"
"a visible medium that reveal truth" -> "...reveals truth"
"what a cause make" -> "what a cause makes"
Verification (all gates from brief Phase 4):
- closure verifier: 0 unresolved / 0 mismatches on all ADR-0084 packs
(remaining domain-pack red is PR #97 follow-up — addressed by PR #99)
- ADR-0084 integration test: 30/30
- cognition eval: byte-identical to baseline
- packs lane: 6/6
- smoke lane: 67/67
Files touched: 5 gloss files (cognition / causation / meta / attitude /
spatial), 4 lexicon files (cognition / meta / action / spatial), 5 manifest
checksum refreshes (+ action), 1 verifier code change, 1 integration test
fixture extension, 1 deterministic-pack-entry-id test bump (085→091).
cProfile attribution (2026-05-21) identified
``core.physics.salience.SalienceOperator.compute`` as 64% of total
``ChatRuntime.chat()`` time. Pre-fix it was a nested Python loop
over ``regions × regions`` with one ``np.linalg.norm`` call per
pair. For N≈500 mounted-vocab regions per turn that meant ~250k
norm calls per turn, dominating end-to-end latency.
Fix: numpy broadcast for pairwise displacement, distance,
pressure-delta, and contribution. Same math; same contract.
ULP-level reassociation drift is absorbed by the 12-decimal
precision ``_salience_address`` already used for content
addressing, and by the float32 conversion at the downstream
``SalienceMap.scores_arr`` site, so neither the content_address
nor the top-k ordering changes.
Measurements (region set: N=493, dim=5, seeded):
vectorized: 11.78 ms/call
old-loop: 672.30 ms/call
speedup: 57.1×
End-to-end on 8 cognition-shape prompts:
pre-fix: ~970 ms/turn
post-fix: 565 ms/turn (-42%)
Validation:
* 15 new tests in ``tests/test_salience_vectorize_parity.py``:
- parity with a nested-loop reference to 1e-9 absolute on
curvature_magnitude, gradient_vector, influence_radius
across N ∈ {1, 2, 8, 32, 128, 493}
- content_address byte-identical across N ∈ {1, 8, 32, 128}
- top-16 ordering matches the reference at N ∈ {32, 128, 493}
- empty regions returns empty map
- single region has zero curvature
* ``core eval cognition`` byte-identical: public 100/100/91.7/100.
* ``core test --suite cognition`` 120/0/1, ``smoke`` 67/0.
The file's pre-existing docstring promised a Rust path
(``core_rs::physics::salience::compute_curvature``) that does not
yet exist — the numpy vectorization realizes the lift now while
keeping the Rust port a future optimization on stable semantics
(CLAUDE.md: "Rust backend parity only after Python semantics are
locked by tests").
Closes audit Findings 6 (within-turn recall not batched) and 7
(probe-ingest / commit-ingest dual field) as a single PR — the two
are architecturally entangled and resolve together.
Pre-fix flow in ``ChatRuntime.chat()``:
1. ``probe_ingest(filtered)`` → ``probe_state.F``
2. Gate check on ``probe_state.F``
3. If gate fires: ``commit_ingest`` + stub response
4. Otherwise: ``commit_ingest`` + drive bias → ``field_state.F``
5. Walk runs on ``field_state.F``
The gate observes one manifold position; the walk navigates a
slightly different one (drive bias applied between them). Honest
refusal decisions and walk outputs are made on different fields —
the audit's named coherence gap.
This PR ships a flag-gated unified-ingest path following the
codebase's standard substantive-change pattern (ADR-0046 /
ADR-0062 / ADR-0085 / ADR-0088 / ADR-0089):
``RuntimeConfig.unified_ingest: bool = False`` (default).
When ``True``:
1. ``commit_ingest(filtered)`` runs first.
2. Drive bias applied immediately.
3. Gate observes ``committed.F``.
4. If gate fires: stub response (turn has already committed —
intentional semantic change documented in ADR-0090).
5. Otherwise: walk runs on the same ``committed.F`` the gate
decided against — no second ``commit_ingest`` call.
6. ``probe_ingest`` is not called on this path.
When ``False`` (default): historical behavior is preserved
bit-for-bit; ``probe_ingest`` still runs first.
ADR-0090 documents:
* Phase 1 (this PR): unified-ingest substrate.
* Phase 2 (separate PR, after Phase 1 validates): batched recall
— pass the gate's ``direct_hits`` into ``generate()`` as a
``prebuilt_first_recall`` so the walk's first step does not
re-call ``vault.recall()`` on the same field. Single recall
call eliminated per turn.
* Out of scope: ``recall_batch`` for per-step walk recalls
(each step's query depends on the previous step's field
state; not batchable without changing walk geometry).
Validation:
* 5 new tests in ``tests/test_unified_ingest_null_lift.py``:
- flag defaults to ``False`` on ``DEFAULT_CONFIG``
- flag-off surface + trace_hash + vault_hits byte-identical
- flag-on does not call ``probe_ingest`` (verified via spy)
- flag-on produces well-formed surface + trace_hash
- flag-off still calls ``probe_ingest`` (historical guard)
* ``core eval cognition`` byte-identical across all three splits:
public 100/100/91.7/100, dev 100/100/78.6/100, holdout
100/100/83.3/100.
* ``core test --suite cognition`` 120/0/1, ``smoke`` 67/0,
``runtime`` 19/0.
Comb-pass status after this PR:
* Item 4 (graph topo) ✓ #92
* Item 5 (realizer node_map) ✓ #91
* Item 6 (batch recall) ✓ ADR-0090 substrate (this PR); Phase 2
optimization is queued
* Item 7 (probe/commit dual ingest) ✓ ADR-0090 (this PR)
* Item 8 (dead defensiveness sweep) ✓ #91
* Item 9 (local imports) ✓ #91
* Item 11 (dead ``_fold_compose_into_surface``) ✓ #91
* Item 13 (``_serialize_*`` fold) ✓ #91
* Item 15 (GenerationResult tuple/list) ⊘ false positive
* Item 16 (subject normalization consistency) ✓ #93
* Item 17 (redundant ``^`` anchors) ✓ #94
* Tier 5 minor (``_BE_FORMS`` hoist, walrus, reverse-iter) ✓ #94
Comb pass 2026-05-21 (item 16).
Pre-fix ``classify_intent`` applied ``_normalize_subject`` only to
DEFINITION / CAUSE / VERIFICATION paths. COMPARISON, FRAME_TRANSFER,
TRANSITIVE_QUERY (non-"means" branch), and BELONG_QUERY returned
bare ``.strip()`` subjects. A probe like *"Compare the parent and
a child"* would carry the articles ("the parent", "a child") into
the subject slot, breaking downstream pack-resolver lookups that
key on bare lemmas.
Fix: apply ``_normalize_subject(..., IntentTag.DEFINITION)`` at every
classifier return site that was previously bare ``.strip()``.
DEFINITION mode preserves multi-word noun phrases (only strips
leading articles + trailing punctuation + infinitive markers); the
aux-verb stripping that's only meaningful for CAUSE/VERIFICATION
stays scoped to those paths.
Sites fixed (5):
* COMPARISON subject + secondary_subject
* FRAME_TRANSFER subject + frame
* TRANSITIVE_QUERY subject (both the regular and "means" → DEFINITION
redirect branches now share one normalized binding)
* BELONG_QUERY subject
Behavior:
* Eval cases without articles (the entirety of cognition v1) are
byte-identical: ``"memory"`` and ``"recall"`` survive
``_normalize_subject`` unchanged.
* Multi-word noun phrases survive intact: ``"artificial
intelligence"`` is preserved (no aux-verb-strip wrongly trimming
to head-noun).
* Article-prefixed subjects ("the parent") now strip consistently
with the DEFINITION path that's done so since ADR-0049.
Validation:
* 7 new tests in
``tests/test_intent_subject_normalization_consistency.py``
pin the consistency contract across COMPARISON, FRAME_TRANSFER,
TRANSITIVE_QUERY, BELONG_QUERY, DEFINITION (regression guard
on the pre-existing path), and CAUSE (regression guard on the
aux-verb-strip behavior).
* ``core eval cognition`` byte-identical across all three splits:
public 100/100/91.7/100, dev 100/100/78.6/100, holdout
100/100/83.3/100.
* ``core test --suite cognition`` 120/0/1, ``smoke`` 67/0.
* ``pytest -k intent`` 229/0.
Comb pass 2026-05-21 (item 4).
Pre-fix the topological-sort implementation in
``PropositionGraph.topo_order`` had two compounding inefficiencies:
* ``queue.pop(0)`` on a list is O(N) per pop → O(N²) total
* The inner ``for e in self.edges`` rescanned all edges on every
iteration → O(N × E) overall
This is invisible on today's 1–2 node production graphs but would
become a real regression the moment compound-intent multi-node
dispatch (ADR-0089 Phase C2) or the grounded realizer's multi-clause
output (ADR-0088 Phase B follow-up) lands.
Fix: standard Kahn's with a precomputed out-edge adjacency map and
a ``deque`` for the work queue. O(N + E) overall. Deterministic
output preserved — the queue is seeded with sorted zero-in-degree
nodes (identical to the pre-fix list sort), and direct-successor
order matches edge-iteration order (identical when edges retain
insertion order).
Pinned by 6 new tests in ``tests/test_graph_topo_order_perf.py``:
* single-node graph (today's production shape) byte-identical to
pre-fix output
* empty graph returns empty tuple
* chain (A→B→C→D) orders root → leaf
* diamond (A→B, A→C, B→D, C→D) keeps A first, D last, B/C between
* three disjoint roots emit in sorted order
* 100-node chain returns correct full order (would have been
visibly slow under the O(N²) pre-fix algorithm)
Validation:
* ``core eval cognition`` byte-identical (public 100/100/91.7/100)
* ``core test --suite cognition`` 120/0/1
* ``core test --suite smoke`` 67/0
Comb-pass note: item 15 (GenerationResult.tokens typed tuple but
assigned list) was investigated and turned out to be a Pyright
false positive — ``GenerationResult.__post_init__`` already coerces
to tuple via ``object.__setattr__``. Contract is enforced at
runtime; only Pyright's static analyser misses the coercion site.
No fix needed.
Bundle of 5 hot-path optimizations + 1 dead-code removal + 1 import
sweep + 1 helper fold, surfaced by a comb pass through the cognitive
spine starting from ``CognitiveTurnPipeline.run()`` and walking
outward through ChatRuntime, intent classification, the graph
planner, the realizer, and the vault. All eval lanes byte-identical
to MEMORY baseline; null-lift confirmed by ``core eval cognition``
across public / dev / holdout splits.
Hot-path fixes:
1. ``ChatRuntime._apply_oov_policy`` no longer rescans every
manifest per OOV token. Two precomputed booleans on
``self`` capture the FAIL_CLOSED-all and PROPOSE_VOCAB-any
aggregates at construction time. Manifests are immutable
post-construction so the cache is safe. Turns the path from
O(packs × OOV) to O(OOV).
2. ``CognitiveTurnPipeline.run`` calls ``classify_compound_intent``
once and takes its dominant ``compound.primary`` as the seeded
intent. Pre-fix the pipeline called both ``classify_intent``
and ``classify_compound_intent`` on every turn — and
``classify_compound_intent`` internally invokes
``classify_intent`` on the dominant fragment, so every non-
compound prompt walked the 15-regex cascade twice.
3. ``TeachingStore.triples()`` materializes once per turn.
Pre-fix ``_maybe_transitive_walk`` and ``_maybe_compose_relations``
each called ``self.teaching_store.triples()`` independently,
doubling the per-turn O(N) filter+tuple-build cost. Both
helpers now accept an optional ``triples`` arg; the pipeline
computes once and passes through.
5. ``realize_semantic`` and ``realize_target`` build a
``node_id → obj`` map once and look up each step in O(1)
instead of an O(N) linear scan of ``graph.nodes`` per step.
The cost was invisible on today's 1-2 node graphs but would
have become an O(N²) regression on the multi-node graphs
ADR-0089 Phase C2 plans to introduce.
Dead-code / cleanup:
- Removed dead ``CognitiveTurnPipeline._fold_compose_into_surface``
(no callers since PR #76 routed all surface composition
through ``resolve_surface``).
- Folded ``_serialize_walk`` + ``_serialize_compose`` (identical
bodies) into one ``_serialize_operator`` helper.
- Hoisted ``import json`` and ``RatifiedIntent`` from inside hot
method bodies to module top (same pattern PR #76 applied to
``_is_useful_surface``).
- Dead-defensiveness sweep on ``ChatResponse`` field reads in
``pipeline.run()``: ``getattr(response, "<field>", default)``
where the field always exists on the dataclass with a default
is replaced by direct attribute access (6 sites:
``realizer_grounded_authority``, ``recalled_words``,
``grounding_source``, ``register_canonical_surface``,
``pre_decoration_surface``, ``admissibility_trace``,
``region_was_unconstrained``). ``refusal_reason`` retains the
guarded read because ADR-0024 Phase 2 leaves its
materialisation site dormant.
Benchmark profiler:
- ``benchmarks/pipeline_profiler.py`` rebound from
``classify_intent`` to ``classify_compound_intent`` (the new
single-classification site). All other timing hooks unchanged.
Tests:
- 4 new tests in ``tests/test_comb_pass_hot_path.py`` pin: OOV
aggregates exist as bools; compound classifier runs exactly
once per turn; ``triples()`` materializes exactly once per
turn; realizer correctly resolves obj slots across an 8-node
graph.
- All existing tests pass. ``core eval cognition`` byte-identical:
public 100/100/91.7/100, dev 100/100/78.6/100, holdout
100/100/83.3/100.
- ``core test --suite cognition`` 120/0/1, ``smoke`` 67/0,
``runtime`` 19/0.
Closes audit Finding 2 (2026-05-20) — Phase B substrate.
Pre-fix ``CognitiveTurnPipeline.run()`` invoked ``realize_semantic``
on the ungrounded ``PropositionGraph``. Every non-COMPARISON /
non-CORRECTION node was born with ``obj = "<pending>"`` and the
realizer emitted surfaces like ``"X is defined as ..."`` that
``_is_useful_surface`` correctly rejected. The realizer therefore
never won the surface resolver introduced by PR #76 — it was
structurally present but semantically inert in the hot pipeline
path.
This PR follows the codebase's standard substantive-change pattern
(ADR-0046 ``forward_graph_constraint``, ADR-0062 ``composed_surface``,
ADR-0083 ``transitive_surface``, ADR-0085 ``gloss_aware_cause``):
ship the wiring behind a flag, default ``False``, with a CI-pinned
null-lift invariant.
Changes:
* ``RuntimeConfig.realizer_grounded_authority: bool = False`` —
operator-level opt-in.
* ``ChatResponse.recalled_words: tuple[str, ...] = ()`` —
alphabetic-filtered walk tokens from the recall step, populated
on the main path of ``ChatRuntime._chat``. ``walk_tokens`` is
now computed unconditionally so non-English packs also surface
them (English keeps using them for
``articulate_with_intent`` as before).
* ``CognitiveTurnPipeline.run()`` — when the flag is set and the
response carries any recalled words, calls
``ground_graph(graph, response.recalled_words)`` and re-invokes
``realize_semantic`` on the grounded graph. The surface
resolver (PR #76) then picks the realizer's grounded output
when it clears ``_is_useful_surface`` and the unknown-domain
gate did not fire.
Phase A (realizer fluency parity — gloss-aware templates, 3sg verb
agreement, pack-provenance tag) is documented in ADR-0088 §Phase A
and is the prerequisite for enabling this flag in production. The
known fluency gap (e.g. ``"Light is a visible medium that reveal
truth"`` — subject-verb disagreement leaking from realizer
templates) is the reason the flag ships default-off: operators get
the wiring stable now, the realizer becomes a real authority once
Phase A's fluency upgrade lands.
Verification:
* 4 new tests in ``tests/test_realizer_grounded_authority_flag.py``:
- flag defaults to ``False`` on ``DEFAULT_CONFIG``
- flag-off produces byte-identical surface + trace_hash
(null-lift invariant)
- ``recalled_words`` is populated on the main path
- flag-on runs end-to-end without crashing (surface is
well-formed regardless of which authority won the resolver)
* ``core eval cognition`` — public 100/100/91.7/100,
byte-identical to the MEMORY baseline (default-off).
* ``core test --suite cognition`` — 120/0/1.
* ``core test --suite smoke`` — 67/0.
* ``core test --suite runtime`` — 19/0.
Closes audit Finding 4 (2026-05-20) — Phase C1.
Pre-fix ``CognitiveTurnPipeline.run()`` called only the single-intent
``classify_intent`` and silently dropped every secondary clause of a
compound prompt like *"What is X and how does it relate to Y?"*.
The graph never saw the second subject, the resolver never saw the
second clause, and the trace recorded only the dominant clause —
with no operator-visible evidence that anything was dropped.
Phase C1 is the **observability substrate** for ADR-0089: the
pipeline now also runs ``classify_compound_intent`` at step 1b and
records every dropped secondary clause on
``CognitiveTurnResult.dropped_compound_clauses``. The dominant
clause continues to route through the existing single-intent path
exactly as before — surfaces, trace_hashes, and every existing test
remain byte-identical.
Changes:
* ``CognitiveTurnPipeline.run()`` calls ``classify_compound_intent``
alongside the existing ``classify_intent`` and computes
``dropped_compound_clauses = compound.parts[1:]`` when the
compound is multi-part.
* ``CognitiveTurnResult.dropped_compound_clauses:
tuple[DialogueIntent, ...] = ()`` — empty tuple == single-clause
turn; len > 0 == operator-visible evidence of dropped secondary
clauses.
Out of scope (per ADR-0089):
* Phase C2 (opt-in multi-node graph dispatch + widened trace_hash
+ multi-clause surface) is deliberately scoped to a separate
PR because it widens ``compute_trace_hash``, the surface
resolver contract, and ``plan_articulation``.
* The dominant-clause routing path is unchanged: the audit's
broken-subject case ("truth, and why does it matter") is *not*
fixed here — that improvement is Phase C2 scope.
Verification:
* 4 new tests in ``tests/test_compound_intent_substrate.py``:
- single-clause prompts record empty
``dropped_compound_clauses``
- AND-joined compound surfaces the secondary clause as a
DialogueIntent with the right tag (CAUSE for "why does ...")
- the user-visible surface and trace_hash for a compound prompt
are byte-identical across two independent runs (no behavior
change at the truth-path layer)
- prompts without a recognised connector do not invent a
secondary clause
* ``core eval cognition`` — public 100/100/91.7/100, byte-identical
to the MEMORY baseline.
* ``core test --suite cognition`` — 120/0/1.
* ``core test --suite smoke`` — 67/0.
* ``core test --suite runtime`` — 19/0.
Closes audit Finding 6 (2026-05-20).
Pre-fix ``_STOP_TOKENS = frozenset({"it", "to", "word"})`` was
hardcoded inside ``generate.stream.generate()`` and inhibited those
three tokens unconditionally across every pack, every language, and
every domain. If a pack legitimately needed one of them as a content
word — e.g. a philosophy pack where ``"word"`` maps to λόγος, or a
syntax pack where ``"to"`` is a content node — there was no override
path. The ``_try_index`` guard handled the case where the token was
absent from the pack, but offered nothing for packs that contained
the token and meant it.
Changes:
* ``generate.stream.generate`` accepts ``stop_tokens: frozenset[str]
| None = None``. ``None`` resolves to the historical
``_STOP_TOKENS`` constant, preserving byte-identity for every
pre-Finding-6 caller.
* ``RuntimeConfig.stop_tokens: tuple[str, ...] | None = None`` —
operator-level override threaded through ``ChatRuntime`` into
``generate()``.
* Default ``None`` preserves byte-identical behavior for every
existing pack and every existing test.
Scope notes:
* This PR delivers the *runtime override* surface. Manifest-driven
per-pack overrides (``generation_stop_tokens`` field in the pack
manifest) are the natural next step but require a pack-schema
ADR and re-ratification of every affected pack, so the wiring
lands first and the manifest field follows on a separate ADR.
* ``agenerate`` was identified as unreachable and is being deleted
in a sibling PR (Finding 7); its hardcoded ``_STOP_TOKENS``
reference disappears with it, so it is intentionally not touched
here.
Verification:
* 4 new tests in ``tests/test_stop_tokens_override.py``:
- ``RuntimeConfig.stop_tokens`` defaults to ``None``
- ``generate()`` signature exposes ``stop_tokens`` with default
``None``
- the historical constant is unchanged
- an explicit override flows through the runtime end-to-end
* ``core eval cognition`` — public 100/100/91.7/100, byte-identical
to the MEMORY baseline.
* ``core test --suite cognition`` — 120/0/1.
* ``core test --suite smoke`` — 67/0.
* ``core test --suite runtime`` — 19/0.
Closes audit Finding 3 (2026-05-20).
Pre-fix ``ratify_intent`` defaulted to ``threshold=0.0``, which admits
anything with non-negative ``cga_inner(prompt, anchor)`` — the field
gate (ADR-0022 §TBD-1) was structurally live but semantically
transparent. RATIFIED was logged on essentially every turn because
the CGA inner product over conformal space is not sign-symmetric.
Measurement (``scripts/calibrate_ratification_threshold.py``):
* Runs every cognition eval prompt (45 cases = 13 public + 13 dev +
19 holdout) through a primed ``CognitiveTurnPipeline``.
* Captures the actual ``cga_inner(prompt, anchor)`` score from the
pipeline's own ``_ratify_intent`` via a temporary spy on the
imported ``ratify_intent`` binding.
Observed distribution:
* 34 RATIFIED: min=+1.1039 p10=+1.1039 median=+2.6820 max=+5.7508
* 11 PASSTHROUGH (no vocab-grounded anchor available; score=0.0)
* 0 DEMOTED at any threshold ≤ 1.10
Threshold = 0.5 chosen as the calibrated default:
* Well below the empirical floor of 1.10 — every currently-passing
case stays RATIFIED, byte-identically.
* Clearly non-trivially positive — random Cl(4,1) inner products
fluctuate around zero, so 0.5 demands genuine correlation with
the anchor rather than passive non-negativity.
* Leaves headroom for the gate to actually demote weakly-aligned
off-corpus / adversarial prompts to UNKNOWN and route them
through the honest-refusal surface.
Verification:
* ``core eval cognition`` — public 100/100/91.7/100, holdout
100/100/83.3/100, dev 100/100/78.6/100 — byte-identical to
MEMORY baselines.
* ``core test --suite cognition`` — 120/0/1
* ``core test --suite smoke`` — 67/0
* ``core test --suite runtime`` — 19/0
* 2 new tests in ``tests/test_ratification_threshold_default.py``
pin both the constant and the signature default so a future
change cannot silently regress to ``0.0``.
Closes audit Finding 5 (2026-05-20).
Pre-fix ``CognitiveTurnPipeline._speculative_subjects`` was a bare
``set[str]`` that only grew over a session. Two correctness gaps:
* A subject promoted to ``EpistemicStatus.COHERENT`` via the teaching
review loop kept appearing with the "(speculative, not yet
reviewed)" marker forever, contaminating reviewed material on
later probes.
* Long teaching sessions widened the per-turn substring scan in
``_should_mark_speculative`` without bound.
Fix:
* Back the cache with ``OrderedDict[str, None]`` (LRU) capped at
``_MAX_SPECULATIVE_SUBJECTS = 64``.
* Introduce ``_remember_speculative_subject`` (insert / refresh) and
``_forget_speculative_subject`` (evict) helpers; route all
SPECULATIVE inserts through them.
* When a proposal lands as ``EpistemicStatus.COHERENT``, evict the
subject and every long-enough non-stopword token derived from it,
so the marker stops appearing on reviewed material.
Iteration order in ``_should_mark_speculative`` is unchanged (keys
view); lookups remain O(1). No surface change for any case the prior
behavior didn't already mishandle, so byte-identical eval surfaces
stay stable (verified locally against ``core eval cognition`` public /
holdout / dev splits — all unchanged from MEMORY baseline).
Tests (7 new, ``tests/test_speculative_subject_lifecycle.py``):
* storage is an OrderedDict and the cap is 64
* remember normalizes (lower+strip) and drops empty input
* remember refreshes LRU position on re-insert
* cache caps at 64 with insertion-order eviction
* forget is case-insensitive and removes the entry
* forget on a missing / empty subject is a no-op
* ``_should_mark_speculative`` triggers after remember and stops
triggering after forget
Audit findings referenced:
https://github.com/AssetOverflow/core/pull/76 (Finding 5, "Unbounded
``_speculative_subjects``")
* fix(cognition): add explicit surface resolution policy
* test(cognition): cover explicit surface resolution policy
* fix(cognition): route pipeline surfaces through resolver
* fix(cognition): address PR #76 review comments
- hoist `_is_useful_surface` import from inside `run()` to module top
- call `_render_walk_surface` / `_render_compose_surface` via the class
name (both are @staticmethod) for consistency with the existing
`_fold_*_into_surface` helpers
- drop redundant `realized_surface` truthiness check in
`resolve_surface` — `realizer_useful` already excludes empty /
placeholder surfaces via `_is_useful_surface`
Tests: tests/test_surface_resolution.py + tests/test_cognitive_turn_pipeline.py
green (16 passed); cognition suite 120/1s, smoke suite 67/0.
The original "Why does light exist?" complaint that motivated ADR-0084
was specifically about CAUSE-intent surfaces. ADR-0084 (substrate) +
PR #65 (content) already moved DEFINITION/RECALL to gloss-grounded
surfaces ("Light is visible medium that reveal truth."). But CAUSE
still dispatched through the chain-walk path:
Before: light — teaching-grounded (cognition_chains_v1):
cognition.illumination; logos.core.
light reveals truth (cognition.truth).
No session evidence yet.
After: Light exists as visible medium that reveal truth.
pack-grounded (en_core_cognition_v1).
The chain-walk is structurally correct but the wrong SHAPE for a why-
question — it's a graph traversal, not an explanation. ADR-0085 fixes
the shape using the same gloss material that DEFINITION/RECALL already
consume, with no new content authoring.
Additive composer
chat/pack_grounding.py:gloss_aware_cause_surface()
- Resolves gloss via lexicon-residency-checked resolve_gloss().
- Frames POS-aware:
NOUN -> "{Lemma} exists as {gloss}."
VERB -> "To {lemma} is to {gloss}."
ADJ -> "To be {lemma} is to {gloss}."
* -> falls back to _frame_gloss (predicate-identity).
- Threads anchor lens via the existing helper (ADR-0073c parity).
- Returns None when no gloss exists — runtime falls through to the
existing chain-walk path. Additive: no CAUSE case loses its surface.
Runtime dispatch
chat/runtime.py — IntentTag.CAUSE tries gloss path FIRST under the
flag; falls through to teaching_grounded_surface* on None.
Unconditional fallback — never silent.
Opt-in flag
core/config.py — RuntimeConfig.gloss_aware_cause: bool = False
Default off preserves pre-ADR-0085 chain-walk surfaces byte-
identically (null-drop invariant, CI-pinned).
Prompt-diversity classifier update
evals/prompt_diversity/runner.py — _CAUSE_MARKERS widened with the
explanation-frame markers ("exists as", "is to", "to be", "is for",
"purpose of") plus bare-form predicates ("reveal" alongside
"reveals"). Neither composer path is penalised on shape_fit just on
inflection grounds.
v1/public lift (flag OFF vs ON, 26 cases)
intent_accuracy : 65.4% -> 65.4% ( — )
versor_closure_rate : 100.0% -> 100.0% ( — )
response_shape_fit : 57.7% -> 57.7% ( — , both frames recognized)
audit_in_surface_rate : 42.3% -> 42.3% ( — , envelope ADR's job)
gloss_quote_rate : 11.5% -> 23.1% (+11.5pp, structural lift)
Tests (15)
- 5 pure composer (NOUN/VERB frame, unknown/empty None, no chain-
walk artifacts in surface)
- 5 runtime dispatch (flag-off chain-walk, flag-on gloss, parametrized
across glossed subjects, VERIFICATION unchanged under flag, no-
gloss fallback engages)
- 5 cognition lane invariance (aggregate metrics byte-identical
under both flag states; surfaces deliberately shift on the 2 CAUSE
cases with glossed subjects — the structural-change-vs-metric-
invariance both-sides invariant)
Lanes
smoke 67/0, cognition 120/0/1 skipped, packs 6/0, teaching 17/0,
runtime 19/0. core eval cognition byte-identical 100/91.7/100/100
under both flag states.
Scope limits (per ADR §Scope limits)
- CAUSE only; VERIFICATION still chain-walks (different shape).
- English pilot only; Greek/Hebrew packs not opted into definitional
layer yet (ADR-0084 scope limit).
- Single-lemma subjects; compound/anaphoric fall through.
- Opt-in until cognition holdout confirms the lift transfers off-
fixture. Future PR flips default on.
Out of scope
- Surface-vs-envelope cleanup ("pack-grounded (...)" still leaks).
- Predicate licensing (ADR-0086).
- Content style pass (bare lemma forms in glosses — separate brief).
The v1 gloss-quote detector used a 4-token contiguous window of
≥4-char tokens. That heuristic was too strict for the actual ADR-0084
brief gloss style, which is deliberately short and primitive-only:
light "visible medium that reveal truth" 5 tokens ≥4 chars
parent "person with a child" 3 tokens ≥4 chars ← can't window
recall "get memory from before" 3 tokens ≥4 chars ← can't window
wisdom "good use of knowledge" 2 tokens ≥4 chars ← can't window
Result: post-PR #65 baseline showed gloss_quote_rate=0.0% even though
the pack-grounded composer was visibly emitting glosses verbatim:
surface: "Parent is person with a child. pack-grounded (en_core_relations_v1)."
gloss: "person with a child"
window: could not even form
Replace with substring match against the gloss text. The composer
emits the gloss verbatim (no paraphrasing — that's the no-LLM
discipline), so substring is exact, high-confidence, and trivially
correct:
gloss_quoted ⟺ gloss.lower().strip() in surface.lower()
Re-baselined v1/public (26 cases):
gloss_quote_rate: 7.7% (false-positive 4-token window noise)
→ 0.0% (post-#65, broken metric)
→ 11.5% (this PR, real signal)
The other four metrics unchanged. 3/26 cases (DEFINITION on
``evidence``/``recall``/``parent``) are detected as gloss-quoted now,
which matches reality — the pack-grounded composer at
chat/pack_grounding.py:398 has been gloss-aware all along; it just
had no glosses to quote pre-#65.
Why this is just a heuristic refinement, not a contract change:
The contract.md still says v1 has NO pass thresholds beyond
versor_closure_rate==1.00. The lane's job is to establish baseline
distribution. The heuristic was *measuring the wrong thing* — fixing
the measurement is a contract clarification, not a contract change.
Tests added (TestGlossQuote, 4 cases):
- short brief-style gloss detected via substring
- chain-walk surface for same lemma NOT counted as gloss-quoted
- unknown term returns False
- empty terms returns False
Updated the function docstring with the post-#65 context so future
readers understand why v1's contract predicted 0% but reality is ~12%.
After PR #64 (substrate) and PR #65 (content) both landed on main, this
test is the promised follow-up that exercises the substrate-callable
verify_definitional_closure against the real ratified content rather
than fixture packs. It pins three contracts:
1. Substrate-vs-content handshake. The standalone
scripts/verify_definitional_closure.py is the agent's dev-loop
tool; this test is the gate-callable equivalent the ratification
pipeline can invoke. Both must agree on what passes — divergence
is a contract bug.
2. Content drift catcher. Any future content edit that adds an
unresolved token / non-mounted dependency / silent staging leak
fails this test before the edit lands on main.
3. Staging exclusion. en_minimal_v1 is staging per the ADR-0084
pack-content brief and must not be load-bearing for the closure
rule. Test-pinned via a production-pool subtest.
Substrate fix: allow empty definitional_atoms
The substrate's strict parser previously rejected empty
definitional_atoms. That stance was wrong: per the ADR-0084 pack-
content brief, the per-entry atom list excludes articles, prepositions,
and copulas. A gloss whose every content word is a function word
(e.g. en_core_temporal_v1/prior → "before") has zero content atoms by
construction. The closure rule passes vacuously when atoms is empty
— there is nothing to close. The gloss-vs-atoms mismatch check in
the standalone verifier is the second-layer gate that distinguishes
by-construction emptiness (legitimate) from by-omission emptiness
(laziness). Substrate parser shouldn't double-gate the same concern.
The corresponding substrate test flipped from
test_empty_definitional_atoms_rejected to
test_empty_definitional_atoms_accepted, with comment explaining the
reasoning.
Primitives expansion: can + action
Two content entries (en_core_cognition_v1/person → "who can know and
do" and en_core_meta_v1/intend → "decide before an action") leaned on
'can' and 'action' as atom references. Today those lemmas resolve
ONLY via en_minimal_v1/lexicon.jsonl — the staging pack. That's a
production-vs-staging leak: production content should not be load-
bearing on staging.
Two clean alternatives:
(a) rewrite the two glosses to avoid 'can' and 'action'
(b) promote 'can' and 'action' to primitives
Chose (b): both lemmas are genuinely terminal-feeling (can is a basic
capability modal; action is an irreducible "what is done"); the
content reads more naturally with them present than with paraphrased
substitutes; and the floor was always going to need both eventually.
The cost is two primitives.jsonl rows + checksum + count bump.
Verification:
scripts/verify_definitional_closure.py exit 0
tests/test_adr_0084_integration_closure.py 30/30 pass
tests/test_adr_0084_definitional_substrate.py 39/39 pass
core test --suite smoke -q 67/67
core test --suite packs -q 6/6
core eval cognition byte-identical
(100/91.7/100/100)
Two-layer gate now in place:
- standalone verifier (dev loop, gloss/atom mismatch check)
- substrate verifier (ratification gate, parametrized over every
opted-in pack, staging-exclusion test, primitives floor coverage)
* docs(adr-0084): propose definitional layer + prompt-diversity suite
Three companion artifacts proposing the next substantive design step
after ADR-0083:
1. ADR-0084 (Proposed) — Definitional Layer for Lexicon Packs
Optional `definition` block on pack entries: gloss,
definitional_atoms, predicates_invited, definition_version,
provenance. Pack-level opt-in. Closure rule: every word in a
gloss must resolve to a same-pack lemma, another mounted pack's
lemma, or a primitive in a new `packs/primitives/` pack.
NO composer change in this ADR (sequenced for ADR-0085) —
ratify substrate before any consumer depends on it.
2. evals/prompt_diversity/ (Proposed) — companion eval lane
~50 cases across question-shape × sophistication × domain,
measuring three new metrics: response_shape_fit,
audit_in_surface_rate (quantifies the trust-boundary leak into
user surfaces), gloss_quote_rate (zero today; rises with future
gloss-aware composer). No v1 pass thresholds — the lane
establishes a baseline distribution so future work has
something to move. 26 seed cases authored covering all 21
categories.
3. docs/handoff/ADR-0084-pack-content-brief.md — paste-ready brief
for a cheaper/faster dev agent to produce the pack content in
parallel. Self-contained, 5 sequenced phases (primitives pack
→ extend 9 existing glosses → add to relations/anchors → write
closure verifier → run safety lanes), explicit don't-touch list
(no composer / runtime / algebra / Greek+Hebrew packs / schema
parser), no-LLM-glosses discipline, per-phase acceptance.
Discovery while drafting: 9 packs already carry glosses.jsonl
under language_packs/data/ with a flat schema (78 entries in
en_core_cognition_v1 alone). The brief reflects that — most
work is extending existing entries, not authoring from scratch.
Strategic context: ADR-0083 raised the *depth* ceiling on chain
composition; ADR-0084 raises the *fidelity* ceiling. The φ
separation probe (memory: phi-separation-falsified) established
that semantic capability lives in chain composition, not in φ
geometry, so deepening the composer's substrate is the natural
next step. ADR-0084 → 0085 (gloss-aware composer) → 0086
(predicate licensing at ratification) is the planned sequence.
* feat(adr-0084): substrate — schema parser, primitives loader, closure verifier
Substrate-only code-side for ADR-0084 (Definitional Layer for Lexicon Packs).
No composer touches the new fields yet; consumer integration is ADR-0085.
Schema (additive, default preserves byte-identity)
- LanguagePackManifest.definitional_layer: bool = False
- compiler loader propagates the flag from manifest.json
language_packs/definitions.py (new)
- GlossEntry dataclass: lemma, gloss, pos, definitional_atoms,
predicates_invited, definition_version, provenance_ids
- parse_gloss_entry(payload, *, strict) — strict mode enforces ADR-0084
§Schema validation row-by-row: required keys, typed lists, no
unknown keys, positive definition_version; lax mode preserves the
legacy two-field shape for back-compat
- load_pack_glosses(pack_id, *, strict) with cache + clear hook
- verify_definitional_closure(pack_id, *, mounted_pack_lemmas,
primitive_lemmas, strict) returning tuple[ClosureViolation, ...];
case-insensitive resolution; cycles permitted per ADR
packs/primitives/loader.py (new)
- Sister loader to packs/safety/ and packs/identity/
- PrimitivesPack frozen dataclass with .lemmas frozenset
- Gates: checksum match, kind=='primitives', definitional_layer:true,
never_auto_mutable:true, pack_id matches dir, primitive_count
cross-check, duplicate-lemma rejection, path-traversal rejection,
strict per-entry schema with allow-list
- DEFAULT_PRIMITIVES_PACK = 'en_semantic_primitives_v1'
tests/test_adr_0084_definitional_substrate.py
- 38 tests covering strict parser (each required key rejection, unknown
key rejection, empty predicates_invited allowed, empty
definitional_atoms rejected, invalid definition_version), lax
parser back-compat, load_pack_glosses (missing/strict raise/lax
skip/malformed JSON), closure verifier (same-pack/primitive/mounted/
unresolved/case-insensitive), primitives loader (every gate), and
a back-compat check that every shipped pack still ratifies with
definitional_layer=False
Lanes: smoke 67/0, cognition 120/0/1, teaching 17/0, runtime 19/0,
packs 6/0. Cognition eval byte-identical 100/91.7/100/100.
When the content PR lands (primitives.jsonl + extended glosses.jsonl
under ADR-0084-pack-content-brief.md), the gate catches any closure-rule
violation without further code change.
* feat(evals): prompt_diversity lane runner — measurement instrument for ADR-0084+
Implements the runner against the existing contract.md + 26-case v1
public split. Lane auto-discovered by evals.framework via the standard
contract + runner convention.
Runner (evals/prompt_diversity/runner.py)
- run_lane(cases, *, config, workers) -> LaneReport
- 5 metrics: intent_accuracy, versor_closure_rate (carried over from
cognition), plus the three new lane-specific metrics —
response_shape_fit, audit_in_surface_rate, gloss_quote_rate
- breakdown dict groups by (question_shape, sophistication, domain)
per contract §How to read the output
- mirrors evals.cognition.runner's parallel worker pattern
Per-shape classifier (deliberately substring/regex-simple at v1)
- predicate_identity, explanation, sequence, two_subject_contrast,
narrative, honest_disclosure
- Unknown shape => neutral pass (don't penalise new categories)
Audit-leak detector
- trust-boundary preamble markers (teaching-grounded (, pack-grounded
(, No session evidence yet.)
- dotted semantic-domain tag regex (cognition.illumination, etc.)
Gloss-quote detector
- resolves expected_terms via chat.pack_resolver.resolve_gloss
- 4-token contiguous-window match against surface (high-confidence
"gloss actually quoted", not "shared one common word")
Tests (tests/test_prompt_diversity_runner.py — 23)
- shape classifier parametrized over the six expected_shape values
- audit-leak detector parametrized over preamble + tag + clean cases
- end-to-end on v1 public:
* versor_closure_rate == 1.0 (only v1 pass threshold per contract)
* every metric in [0, 1]
* breakdown groups present with the four per-cell metrics
* diversity gate: >=5 question shapes, >=3 domains
(defends against future regressions that collapse the suite
back to a cognition-shaped fixture)
v1/public baseline (26 cases)
intent_accuracy : 65.4% (contract predicted 70-85%)
versor_closure_rate : 100.0% (only v1 pass threshold) PASS
response_shape_fit : 53.8% (contract predicted low)
audit_in_surface_rate: 42.3% (contract predicted ~100%)
gloss_quote_rate : 7.7% (contract predicted 0%)
Three baseline surprises worth noting in the report (NOT failures —
the v1 lane is explicitly there to establish the distribution):
- audit_in_surface_rate at 42% (not 100%) means the chain-walk leak
fires on ~11/26; the other 15 are honest-disclosure cases that
emit no audit envelope. Sharpens the future surface-vs-envelope
ADR's actual target: grounded surfaces specifically.
- response_shape_fit at 54% (not "low") — classifier likely has
false positives on the ", which " cause-marker. Worth tightening
once we have an ADR-0085 baseline to compare against.
- intent_accuracy at 65% (below predicted 70-85%) — classifier dips
harder on adversarial/cross-pack than expected. Real gap.
All five smoke/cognition/teaching/runtime/packs lanes still green;
core eval cognition byte-identical 100/91.7/100/100.
* feat(packs): ADR-0084 pack content (primitives + extend glosses + closure verifier) (#65)
* feat(packs): ADR-0084 pack content
* feat(packs): repair ADR-0084 definitional content
* test(adr-0084): adjust substrate manifest tests for post-#65 content reality
PR #65 flipped definitional_layer:true on 13 English packs (9 core +
4 relations + collapse-anchors). The substrate's previous test
test_existing_packs_unchanged asserted that en_core_cognition_v1 +
en_core_relations_v1 still had definitional_layer:False — which was
the right pre-content invariant but is wrong post-content.
Replace it with two complementary tests that hold against real content:
- test_non_opted_packs_default_false:
pins that packs that DIDN'T flip the flag (en_minimal_v1,
he_core_cognition_v1, grc_logos_cognition_v1) still surface
definitional_layer=False through the loader. Defends against
a future change accidentally flipping the flag on a non-opted
pack.
- test_opted_packs_carry_flag:
pins that packs that DID flip the flag (en_core_cognition_v1,
en_core_relations_v1) surface definitional_layer=True through
the loader. Proves the substrate's manifest-field propagation
works against real ratified content, not just fixture packs.
Net: +1 test, same intent (substrate ratifies the manifest field
correctly), now with real-content coverage on both sides of the gate.
All 62 ADR-0084 substrate + prompt-diversity tests pass.
test_frontier_compare_report_viewer_exists was failing on main against
the current report_viewer.html because two verbatim substring checks
no longer matched the viewer's UI copy:
- "Drop report JSON" → viewer now says "Drop JSON report" (order swapped)
- "No network calls" → viewer now says "no network calls" (lowercase)
Both copy refreshes were behavior-preserving — drop-zone affordance
and network-free trust boundary are both intact in the viewer. The
test was coupling to verbatim phrasing rather than to the load-bearing
affordances.
Switch to case-insensitive substring checks that pin what actually
matters:
- "frontier compare" — viewer identity
- "drop" AND "json" together — drop-zone affordance, order-independent
- "no network calls" — trust boundary (case-insensitive)
- fetch(/XMLHttpRequest still hard-banned (case-sensitive — these
are JS API surface, not human-readable copy)
Pre-existing failure flagged in PR #66's body as out-of-scope cleanup;
this is that cleanup.
Strict superset of ADR-0062's depth-1 composer. `max_depth` is the
number of follow-up hops appended beyond the initial chain:
max_depth=0 → byte-identical to single-chain surface
max_depth=1 → byte-identical to ADR-0062 composed
max_depth=2 → byte-identical to ADR-0062 when no second hop
survives, strict superset when one does
The composer surfaces content the realizer was silently dropping
from chains already ratified in `cognition_chains_v1`. Example
live lift on `"Why does light exist?"`:
composed: "light reveals truth, which grounds knowledge."
transitive(2): "...which grounds knowledge, which requires evidence."
Cycle-safe at every depth via a single visited-set; single-corpus
traversal in v1 (cross-corpus transitive deferred to a follow-up
ADR alongside ADR-0064's cross-pack model).
Both flags default False — every existing surface is preserved
byte-identically. When both `composed_surface` and
`transitive_surface` are True, transitive wins.
Implementation:
- `core/config.py`: `transitive_surface: bool = False`,
`transitive_max_depth: int = 2`.
- `chat/teaching_grounding.py`: `_resolve_followup` shared helper
refactored out of the depth-1 composer (no behavioural change),
plus new `teaching_grounded_surface_transitive(subject,
intent_tag, *, max_depth)`.
- `chat/runtime.py`: dispatch order — transitive > composed > single.
Verification:
- tests/test_transitive_surface.py: 16 new tests covering pure-fn
contract, visited-set cycle guard at every depth, runtime
integration, and the cognition-lane null-drop invariant at
`max_depth=2` (public + holdout splits).
- tests/test_composed_surface.py: 11/11 pass after the helper
refactor (ADR-0062 behaviour preserved).
- `core test --suite smoke`: 67 pass.
- `core test --suite cognition`: 120 pass, 1 skipped.
- `core test --suite teaching`: 17 pass.
- `core eval cognition`: 100 / 91.7 / 100 / 100 (byte-identical).
#58 shipped providers.py + model_registry.py for cross-provider
benchmarking but never connected them to runner.py — the adapters
sat unused. This PR wires them through with a clear lane split.
Why a new suite instead of refactoring existing ones
-----------------------------------------------------
The three existing suites (determinism / truth_lock / axis_orthogonality)
pull CORE-only telemetry: trace_hash, versor_condition, register_id,
register_variant_id, anchor_lens_id, register_canonical_surface.
None of those fields can come from OpenAI / Anthropic / Ollama.
Forcing those suites cross-provider would silently produce reports
where the cross-provider rows have empty telemetry — a worse failure
mode than not running them at all. So the routing is explicit:
CORE-only suites → --provider must be 'core'
Cross-provider suites → any provider; CORE is one adapter among many
Operator asks for the wrong combo → loud error with the right alternative.
New module: evals/frontier_compare/cross_provider.py
-----------------------------------------------------
- ProviderObservation dataclass — provider-agnostic observation shape
(prompt, surface, provider, model, elapsed_ms, error fields). No
CORE-internal telemetry expected.
- run_prompt_battery(adapter, *, cfg) → SuiteReport reusing existing
CaseResult / SuiteReport shapes so the report viewer renders both
lanes without schema branching.
- _PROMPT_BATTERY: 7 fixed cases spanning definition / cause /
verification / comparison / procedure / unknown intent shapes.
Stable case_ids so future re-runs against the same provider produce
diffable JSON.
- Per-case 'passed' is loose by design (non-empty surface, no
exception). Cross-provider quality is for human review — not for
the runner to silently score.
Updated CLI: evals/frontier_compare/__main__.py
-----------------------------------------------
- --provider {core, openai, anthropic, ollama} (default: core)
- --model <id> (validated via require_model_card)
- --env-file <path> (default: ./.env)
- Auto-persist non-CORE runs to
evals/frontier_compare/results/<provider>_<model>_<utc>.json
even when --report is omitted. API calls are rate-limited / paid;
losing the artifact is costly.
- Existing CORE-native behavior unchanged when --provider not set.
Results directory: evals/frontier_compare/results/
--------------------------------------------------
Created with .gitkeep — matches the convention used by other lanes
(evals/long_context_cost/results/, evals/koine_greek_fluency/results/,
etc.). Distinct from reports/ which .gitignore excludes for
transient debug output.
Tests: tests/test_frontier_compare_cross_provider.py (9 cases)
--------------------------------------------------------------
- prompt_battery runs with CORE adapter (no API needed)
- adapter exceptions recorded as failed observations, never propagated
- empty surfaces flagged distinctly from adapter errors
- CLI default runs CORE-native (no breaking change)
- CLI prompt_battery with --provider core routes through cross-provider path
- CLI rejects CORE-only suite + non-CORE provider with operator-helpful error
- --help surfaces both suite families
- unregistered model is rejected before any benchmark cycles burn
- ProviderObservation.succeeded handles error / empty / whitespace cases
Live evidence
-------------
$ core test --suite smoke -q
67 passed in 26.55s (no regression)
$ python -m evals.frontier_compare --provider core --suite prompt_battery --json
model=core-native mode=core suite=prompt_battery passed=True score=1.000
[definition_truth ] PASS Truth is a claim or state grounded by evidence...
[definition_knowledge ] PASS Knowledge is justified understanding grounded...
[cause_understanding ] PASS understanding — teaching-grounded (cognition_chains_v1)...
[verification_evidence ] PASS evidence — teaching-grounded (cognition_chains_v1)...
[comparison_knowledge_wisdom ] PASS knowledge contrasts with wisdom...
[procedure_recall ] PASS To recall means to retrieve a stored state from memory...
[unknown_term ] PASS I haven't learned 'xylomorphic' yet...
$ python -m evals.frontier_compare --provider openai --suite determinism
error: suite 'determinism' is CORE-only; pass --suite prompt_battery
(the cross-provider suite) when --provider='openai'.
.gitignore: adds frontier_wave1.json (stray report file repeatedly
written by ad-hoc test invocations).
Two follow-up fixes from end-of-session verification of recent merges:
1. core/cli.py — wire `core contemplation` subcommand
PR #55 + #58 added the contemplation CLI at python -m core.contemplation
but never registered it under the `core` umbrella command, so
`core --help` didn't show it. Adds a subparser mirroring the existing
pattern (chat/test/check/.../doctor) that delegates to the existing
core.contemplation.__main__:main() — no duplication of arg parsing.
Surface preserved verbatim: reports (positional, 1+), --lane
{frontier_compare, contradiction_detection}, --pack-id, --note,
--report, --sink-root.
2. tests/test_architectural_invariants.py — restore INV-02 allowlist
PR #57's evals/lab/phi_separation_probe.py imports normalize_to_versor
for construction-time experimental rotor + embedding work, which
triggered INV-02's AST-scan failure (the test enforces that
normalize_to_versor is only called from a small allowed file set).
evals/lab/ is research-only, never imported by runtime — adding the
probe to allowed_files doesn't weaken the runtime invariant the
test enforces.
Verification
------------
$ core test --suite smoke -q
67 passed in 26.63s (was 66 passed / 1 failed before)
$ core contemplation --help
... shows the new subcommand surface
$ core contemplation evals/contradiction_detection/results/v1_public_*.json \
--lane contradiction_detection \
--sink-root /tmp/sink \
--report /tmp/run.json
... 4 SPECULATIVE findings; sink writes to /tmp/sink/2026/2026-05.jsonl
Connects ADR-0080's read-only contemplation loop to the existing
teaching-pipeline plumbing without forcing a type collapse. The
SPECULATIVE-only invariant from #55 is preserved verbatim; what
changes is *where the findings flow*.
What was wrong with the prior shape
-----------------------------------
PR #55 shipped a parallel core/contemplation/ package whose findings
were written as one JSON blob per CLI invocation, with no consumer.
The SPECULATIVE-only invariant protected a write path that didn't
exist. My closed PR #56 (second miner) would have entrenched the
duplication.
What this PR changes
--------------------
1. Schema (core/contemplation/schema.py)
- Adds a BOUNDARY note documenting why EvidencePointer (teaching)
and ContemplationEvidenceRef (core) intentionally stay separate:
EvidencePointer.source is constrained to {corpus, pack,
vault_coherent} — pointers into reviewed in-process memory the
runtime trusts. ContemplationEvidenceRef points to external
report files that have NOT been reviewed. Converging them would
either widen the runtime-grounding enum (losing the "reviewed
memory only" guarantee) or force benchmark reports to masquerade
as vault_coherent. Both are worse than keeping them separate.
- Adds format_contemplation_finding_jsonl(finding) — the canonical
JSONL formatter mirroring teaching.discovery.format_candidate_jsonl.
2. Runner (core/contemplation/runner.py)
- Both runners gain an optional sink: DiscoveryCandidateSink | None
parameter. When supplied, each finding is emitted as one
canonical JSONL line via the SHARED protocol — same protocol
that backs DiscoveryBufferSink and DiscoveryMonthlyFileSink.
- Sink path is additive: the ContemplationRun blob is byte-identical
whether or not a sink is supplied (pinned by test).
- No sink supplied → existing in-memory behavior preserved exactly.
3. CLI (core/contemplation/__main__.py)
- Adds --lane {frontier_compare, contradiction_detection} flag.
Default unchanged.
- Adds --sink-root <path> flag. When set, instantiates a
DiscoveryMonthlyFileSink and findings land at
<root>/<YYYY>/<YYYY-MM>.jsonl — the SAME layout discovery
candidates use, so operators can grep one stream.
4. Miner (core/contemplation/miners/contradiction_detection.py)
- Restored from closed PR #56 under the unified pipeline.
- Failure-mode split preserved (missed_contradiction /
false_contradiction_flag) with asymmetric repair actions.
What this PR does NOT do
------------------------
- Does NOT unify ContemplationFinding with DiscoveryCandidate.
DiscoveryCandidate.trigger is Literal[would_have_grounded,
successful_comparison, hedge_acknowledged, oov_resolved_via_decomp]
— all turn-loop flavored. None describe "I parsed a benchmark
report." Forcing a 5th trigger that no turn-loop extractor
produces would pollute the turn-loop type for the schema's sake.
- Does NOT extend teaching/gaps.py. Gap aggregates DiscoveryCandidate
cells by (subject, intent) — domain nouns. ContemplationFinding
subjects are namespaced ("contradiction_detection/CON-PUB-002").
Different operator views. A sibling aggregator can come later
when an operator actually asks for it.
Why this is the right unification point
---------------------------------------
The honest convergence is at the *sink* (so all SPECULATIVE evidence
lives in one rooted append-only stream), not the *aggregator* (which
appropriately produces typed views per evidence family). The boundary
doctrine from #55 is preserved; it now connects to existing plumbing
instead of writing JSON to disk with no consumer.
Tests (tests/test_contemplation_pipeline_convergence.py, 10 cases)
------------------------------------------------------------------
- DiscoveryBufferSink satisfies DiscoveryCandidateSink (shared protocol)
- frontier runner emits findings to shared sink
- contradiction runner emits findings to shared sink
- sink is optional — no-op when absent
- emission is canonical JSONL (sorted keys, no newline, deterministic)
- DiscoveryMonthlyFileSink persists findings at <root>/<YYYY>/<YYYY-MM>.jsonl
- sink emission does not alter the ContemplationRun blob (additive)
- contradiction miner predicate split + repair-action asymmetry
- config_hash differs between lanes (replay can distinguish)
- BOUNDARY doc is present in schema.py (regression guard)
- ContemplationEvidenceRef field invariants
- format_contemplation_finding_jsonl is deterministic + canonical
All 18 tests pass (5 original ADR-0080 + 13 new convergence).
Live evidence
-------------
$ uv run python -m core.contemplation \
evals/contradiction_detection/results/v1_public_*.json \
--lane contradiction_detection \
--sink-root /tmp/sink_demo
/tmp/sink_demo/2026/2026-05.jsonl ← same layout as discovery candidates
predicate=missed_contradiction subject=contradiction_detection/CON-PUB-002
predicate=missed_contradiction subject=contradiction_detection/CON-PUB-004
predicate=false_contradiction_flag subject=contradiction_detection/CON-PUB-005
predicate=false_contradiction_flag subject=contradiction_detection/CON-PUB-006
