BOUNDED AUTONOMOUS SCAFFOLD DRAFT (Fable 5) — not a PR, not merged, no status
flip, no main push. For Opus 4.8 + Shay audit before anything proceeds toward main.
Stacked on the verified §3-primitives + §3.4/3.5-ledger stack (reuse, not re-derive;
all descend from main @ 04d67ca5).
Adds (directive steps 3-4):
+ evals/adr_0246_geometric_suite/ runnable §6.1 synthetic geometric suite
(identity, pi-inversion, 90deg-permutation, mild drift, alien tilt e14,
boost e15, near-singular Gram, malformed F) + §6.2 path/holonomy suite
(lawful sequence, small-rotation session accumulation, interleaved refuse,
pack-change hard break, raw!=lawful forensic). 14/14 cases pass.
+ tests/test_adr_0246_geometric_suite.py pins every case + explicit §6.1 pins
(pi-inversion s=-1, 90deg s=0, near-singular Gram error, malformed-F error)
+ identity_manifold.py: MalformedVersorError + _validate_versor guard on
induced_action / typed_residual_energy (§6.1 fail-closed on malformed F)
+ docs/handoff/adr-0246-slice1-scaffold-notes.md placeholder list, §5-§7
uncertainties, constraint-compliance record, explicit Opus/human TODOs
+ docs/audit/artifacts/adr-0246-slice1-scaffold-runlog.txt actual run output
Constraints honored: H_id={I} only; no soft-projection of unlawful A; path
composes lawful actions only (never raw product); no C_id corrector; chat/runtime,
flags, D4 gate wiring untouched (A-04 quarantine pinned); no discrimination report,
no ADR body, no claims language (deferred TODO: Opus/human); D4 plan not modified.
epsilon_turn/epsilon_session are UNCERTIFIED PLACEHOLDERS, never baked into a
module default (PathBudget is caller-supplied) — flagged in notes + run log.
[Verification]: uv run core test --suite smoke -q => 176 passed;
python -m evals.adr_0246_geometric_suite => 14/14 all_passed;
ADR-0246 suites + adjacent D4 identity surfaces => 114 passed (see run log).
First implementation unit of ADR-0246 Ring-1 per the preflight brief §3/§13.
Promotes the induced-action apparatus from the slice-0 eval prototype into pure,
tested geometry — the single source of truth both the future gate surface and the
§11 grounding-feasibility study will consume.
core/physics/identity_manifold.py:
+ induced_action(versor) A(F) = G^-1 · <a_k, F a_j F~>_0 (§3.1)
+ orthogonality_defect(versor) d_orth = ||A^T G A - G||_F (§3.2)
+ typed_residual_energy(versor) e4/e5/spatial_foreign/unclassified (§3.6)
+ orthogonality_defect_of_action / SPATIAL_GRADE1_INDICES / E4,E5 index pins
core/physics/identity_action.py (new, pure):
+ IdentityStabilizer (locked singleton H_id={I}, §3.3)
+ stabilizer_defect d_stab = min_H ||A - H||_G (§3.2)
G-weighted norm reduces to Frobenius at G=I (default pack)
evals/adr_0246_mismatch_diagnostic: rewired to delegate to the canonical
primitives (removes the duplicate prototype copies; drops unused import).
Scope: off-serving (algebra-only, A-04 quarantine, pinned by test); no gate,
threshold, axis, H_id, corrector, or flag change. identity_wave_gate stays
default-off. Path ledger (§3.4/3.5), gate admit surface (§3.7), and the eval
matrix + §11 feasibility study are subsequent units (brief §0a records the
sequencing). No self-Accept.
[Verification]: uv run core test --suite smoke -q => 176 passed;
tests/test_adr_0246_induced_action.py 12 passed (RED-first);
identity surfaces (mismatch-diagnostic rewired, identity_manifold, identity_gate
wave/runtime/eval, gamma_calibration) 87 passed.
D4 Phase 3 (§2.4). Calibrates the wave-gate leakage bound γ_id via the
bracketed-local Fibonacci section search and emits a content-addressed tuning
certificate (evals/adr_0244_gamma_calibration/). Two results, kept separate:
1. The search certifies γ* = 0.2126624458513829 (cert 0079b5f2…) separating
the geometric attack signal — in-subspace rotors admitted, axis→e4/e5
tilt/boost attacks flagged (inversions excluded: ~0-leakage, handled by the
orientation floor). Pinned as identity._WAVE_LEAKAGE_BOUND; the wave path is
decoupled from alignment_threshold (legacy path + hedge bands keep it).
2. The same bound does NOT separate real live traffic. Measured on the engine,
benign final_state.F versors do not preserve span(e1,e2,e3): leakage
0.14–0.81 (mean 0.55), self-align to −0.52, 12/13 benign false-refused at
γ*, best balanced error 0.346. The calibration certifies
flag_flip_authorized=False → identity_wave_gate stays OFF.
Root cause: the shipped value axes (truthfulness=e1, coherence=e2,
reverence=e3) are nominal basis vectors, not dynamically-preserved eigenmodes;
the field evolution gives the identity subspace no dynamical anchoring. The fix
is the ADR-0246 induced-identity-action programme (preflight brief merged). The
wave gate is validated, correctly-off scaffolding — not a live gate. A slow
drift-guard test canaries any future engine change that starts preserving
identity.
Off-serving (evals/ + tests); flag-off byte-identity preserved. ADR §4a +
plan doc + memory updated with the finding.
[Verification]: smoke 176 passed; fast lane 11897 passed / 109 skipped (-n
auto, not quarantine/slow); 59 targeted gate/calibration tests passed.
evals/adr_0244_identity_gate/ (off-serving, deterministic) measures, honestly:
does the operator-preservation wave gate add identity-attack detection value
the geometry-blind legacy path cannot?
Runs a controlled panel through both paths:
- aligned: small rotations WITHIN the value subspace span(e1,e2,e3) — a
legitimate transformation that preserves the value axes;
- attack: versors that geometrically violate identity — π-rotations that
INVERT a value axis (caught by signed self-alignment ≈ −1), and
rotations/boosts that TILT a value axis toward an alien dimension e4/e5
(caught by the subspace-leakage fraction).
Measured result (pinned by tests/test_adr_0244_identity_gate_eval.py):
- the wave gate separates the panels — every aligned versor admitted, every
attack flagged, min_attack_signal 0.35 > max_aligned_leakage 0.0;
- the two measures are non-redundant (inversion: ~0 leakage but −1 alignment);
- wave adds detection value 6-vs-0 over the legacy path, which never reads
the versor geometry and flags none of these geometric attacks.
Honest scope caveat baked into the artifact + docstring: this is detection
value on the geometric signal the gate is DESIGNED to catch. Whether real
paraphrased injections induce such versor geometry through the live encoder,
and whether the placeholder value axes (e1/e2/e3) are the right identity
directions, is empirical (governance annotation item 6) and is the subject of
D4 Phase 3 (γ_id calibration). The flag stays OFF until that is evidenced.
6 tests + CLI (exit 0 iff separates AND adds detection value).
[Verification]: 6 eval tests passed; CLI exit 0; smoke 176 passed; serve-
quarantine + governance pins 14 passed (eval is off-serving); fast lane
(-m 'not quarantine and not slow' -n auto) 11889 passed, 109 skipped.
Additive-only (new evals/ package + test); no core/chat/algebra changes.
Turns the D0 "likely vacuous" annotation on ADR-0244 §2.3 into a re-runnable proof
that ΔQ_top = 0 is a hollow egress gate (the #19-pseudoscalar failure mode).
evals/adr_0244_qtop_vacuity/ — computes Q_top = <psi I5 ~psi>_0 on the real
Cl(4,1) algebra and finds: (1) Q_top = 0.000e+00 EXACTLY across every unit versor
(spatial rotors AND boosts through e5, and products) — I5 is central so
psi I5 ~psi = I5*(psi ~psi) = I5 for a unit versor, scalar part 0; (2) off the
versor manifold Q_top = -grade5(psi ~psi), nonzero only where the I-05 closure
residual already fires; (3) it is a conserved Spin(4,1) invariant but identically
0 on the valid manifold; (4) the decisive test — an aligned identity and an
adversarially-rotated one (a valid versor, overlap 0.963) both read Q_top = 0, so
ΔQ_top = 0 PASSES the attack the spectral-leakage/closure check actually catches.
Typed JSON artifact + CLI (exit 0 iff proven_vacuous).
tests/test_adr_0244_qtop_vacuity.py — pins each step + off-serving quarantine.
Docs: ADR-0244 §2.3 annotation upgraded "likely vacuous" -> "PROVEN vacuous —
retire from egress"; audit-doc Q4 marked RESOLVED with the evidence.
Verdict: do NOT wire ΔQ_top = 0 as an egress admit condition; keep Q_top, if at
all, as a closure-derived diagnostic. Closes the first of the two ADR-0244 D4
blockers with proof.
[Verification]: qtop eval 8 tests green; CLI proven_vacuous=true; smoke below.
Completes ADR-0243 Phase 4 (plan §5). Adds the metrics benchmark scoring the
cognitive lifecycle against concrete falsifiable comparison classes, joining the
already-committed decisive propositional falsifier (bdf8146a) under one eval package.
evals/adr_0243_cognitive_lifecycle/benchmark.py — five falsifiable metrics, each
grounded in a live lifecycle primitive (no decorative numbers); typed
BenchmarkVerdict / MetricResult; deterministic, JSON-safe:
- fidelity: decode overlap |<psi_steady, target>| after relax_to_ground on a
quadratic well from a perturbed start (min >= 0.999; measured 1.0).
- surprise separation: energy-above-ground of ID (small rotations of the
identity axis) vs OOD (near-orthogonal rotations into distinct Cl(4,1)
planes) against a fixed identity well; strict margined separation
min_ood - max_id > 0.05 (measured 0.834); lam0 verified ~0, not assumed.
- insertion cost: relaxation certificate.steps_taken — all certified, bounded
(<= 256), real work (max_steps > 0; measured 22-23).
- f32 drift over T=1000: unit versor iterated by a fixed rotor with no renorm;
f64 holds versor closure to 1e-9 (~8e-14 measured) while f32 truncates to
~4.7e-5 (ratio ~5.9e8) — the gap motivating ADR-0244 §2.5/§2.6.
- falsifier: run_propositional_falsifier wrong == 0 (1008 ID + 18 refusal-parity).
evals/adr_0243_cognitive_lifecycle/__main__.py — subcommand dispatcher
(benchmark [default] / corridor / falsifier); non-zero exit on falsification.
tests/test_adr_0243_benchmark.py — pins overall pass, each metric's falsifiable
claim, the genuine f32/f64 drift gap, CLI routing + exit codes, and the A-04
off-serving quarantine.
[Verification]: in-worktree smoke gate 176 passed; fast lane
(-m "not quarantine and not slow" -n auto) 11808 passed, 108 skipped;
serve-quarantine + third-door cohesion + dispatch hygiene 22 passed; Phase 4
tests (benchmark + falsifier) 17 passed; benchmark overall_passed True
(deterministic across runs).
The propositional slice of ADR-0243's falsifiability benchmark (plan §5 Phase 4),
the arc's load-bearing evidence. Two mechanism-disjoint engines decide the SAME
enumerated propositional problems:
- Field decoder: propositional_entails (Cl(4,1) wave-field ground energy).
- Gold: generate.proof_chain.entail.evaluate_entailment (exact ROBDD tautology
check, ADR-0201 keystone) — genuinely independent, not the field checking itself.
A clause→string bridge (render_clause: `a | ~b | c`) lets both engines decide
byte-for-byte the same problem. Result over the enumerated panel (atoms a,b,c;
premise sets size ≤2):
- 1008 satisfiable-premise ID comparisons: field.entailed ⇔ gold is ENTAILED —
wrong == 0.
- 18 inconsistent-premise cases: field discloses satisfiable_premises=False
exactly where gold returns REFUSED — matched refusal, no ex-falso leak.
- OOD (>5 atoms): field decoder refuses by construction while gold still decides
— the honest ID/OOD scope boundary in the propositional domain.
Deterministic (enumerated, no randomness/wall-clock); off-serving (evals/ only,
chat/runtime non-import pinned).
[Verification]: tests/test_adr_0243_propositional_falsifier.py — 6 passed;
run_propositional_falsifier() → wrong=0 over 1026 field-vs-gold decisions.
Wire the ADR-0242 §5 pinned primitives into the contemplation runner
(ADR-0243 §2.4): contemplate_surprise_history() gates the newest
caller-timed ADR-0239 dual-operator observation through
is_discovery_eligible, the measured history through
cross_band_discovery_gate (F5/F6/F7 persistence), and only then emits a
DiscoveryCandidate via the existing emit_surprise_discovery ->
DiscoveryCandidateSink. Proposal-plumbing only: the sole side effect is
sink.emit().
Surprise-signal sourcing (the Lane A open question): ContemplationFinding
carries no surprise measurement and the report-mining paths have no
geometry, so nothing is fabricated for them — the canonical carrier is
the ADR-0239 dual audit dict (SelfAuthorshipMiner / ADR-0240 harness),
consumed as explicit SurpriseObservations with caller-supplied logical
time (deterministic, no wall clock).
Give propose_kappa_line_search its live caller (ADR-0242 §5-P1):
evals/analogical_transfer/kappa_calibration.py calibrate_transfer_kappa,
confined to the ADR-0240 calibration pipeline per R-04; kappa_search_event
fires exactly once when a sink is passed. Proposal-only — no
GoldTetherMonitor mutation.
A-04 hardening found during review: chat/runtime.py idle_tick lazily
imports core.contemplation.runner inside the serve process, and the
package __init__ eagerly pulled wave_seam -> holographic_vault (a
serve-banned module) — a PRE-EXISTING process-level quarantine breach.
Closed by making the wave-seam re-exports lazy (PEP 562) and keeping the
multi_scale_energy gate import function-local; pinned by a new dynamic
idle-tick-edge test in test_serve_quarantine_transitive.py.
Pinned primitives untouched: is_discovery_eligible,
cross_band_discovery_gate, candidate_from_surprise_dual signatures
unchanged; no changes to cognitive_lifecycle.py or chat/runtime.py.
New evals/adr_0243_cognitive_lifecycle/ composes AudioCompiler/VisionCompiler ->
sensorium_wave_feed -> relax_to_ground/CognitiveLifecycleEngine.egress ->
generate/realizer readback -> explicit GoldTetherMonitor.decide into one
deterministic fixed-replay corridor, per
docs/handoff/ADR-0243-PHASE3-BRIEF-PACK.md Brief B. Pure composition of
already-tested organs; no new core.physics module; OFF-SERVING (evals/ only,
never imported by chat/runtime.py).
Relaxation targets the full audio+vision integrated percept but starts from the
audio-only partial percept (engine.solve() is not used, since it ties start ==
target) so the relaxer decodes across real steps rather than starting
already-converged; stages are driven explicitly and reassembled into one
LifecycleOutcome keyed off the full percept.
[Verification]: uv run core test --suite smoke -q (176 passed); sensorium suite
(152 passed); algebra suite (82 passed, 50 skipped); new
tests/test_adr_0243_sensorium_corridor.py (6 passed, includes a
steps_taken > 0 pin against silent regression to a no-op relax, a cosine/ANN-free
static pin, and a chat/runtime.py non-import pin).
Complete remaining W5/post-pivot backlog without Docker CI: INV-21 allowlist
dedup, -n auto fast/full defaults, frozen f64 GP Python SOT suite, real
sensorium compiler → ψ → algebraic ρ feed, V2 fib/dyadic/log fixed-replay
evidence without production energy promotion, and project doctrine restored to
local-first + ubuntu-latest:host (no Docker CI for merge).
[Verification]: Smoke suite passed locally (145s, 176 passed); make test-fast
(-n auto) passed twice (11709 passed, 108 skipped each ~11–12 min).
Hunt for the -n auto fast-lane polluters flagged in docs/testing-lanes.md
("Follow-up: xdist by default"). Three root causes, all shared-repo-path
writers with no per-test isolation, following the #782 monkeypatch idiom:
1. tests/test_workbench_replay.py::test_replay_leaves_no_trace hardcoded
`Path("engine_state")` (the real shared dir) instead of reading
`engine_state._DEFAULT_DIR`, which the root-conftest autouse fixture
already redirects per-test. It was a victim, not a polluter: any
concurrent worker writing the real dir made this snapshot-diff flake.
Fixed by reading `engine_state._DEFAULT_DIR` dynamically.
2. evals/gsm8k_math/train_sample/v1/runner.py hardcoded its report.json
output to the committed repo path with no override. Two test files
(test_rat1_end_to_end_admission.py, test_wave_a_multiplicative_
aggregation_injector.py; 4 tests total) spawn it as a subprocess and
read the same file back — a write race under -n auto, and a confirmed
downstream victim (test_gsm8k_sealed_attempt_scout.py::
test_report_json_mtime_unchanged_by_scout_import asserts the file's
mtime is stable). Added an optional CORE_GSM8K_TRAIN_SAMPLE_REPORT_PATH
env override (default unchanged) and pointed the 4 call sites at
tmp_path.
3. core/cli.py's `_DEMO_RESULTS_DIR` (evals/forward_semantic_control/
results/) is written, glob-scanned, and index.json-rebuilt by ~11 tests
across tests/test_cli_demo.py's TestDemoSubcommand and TestDemoPreambles
classes. Added an autouse fixture monkeypatching `cli._DEMO_RESULTS_DIR`
to a per-test tmp dir. This also unmasked a latent order-dependent
coupling: test_demo_list_results_indexes_reports and
test_demo_list_results_json_well_formed never wrote their own report,
relying on a sibling test's leftover file in the shared dir (silently
correct only because pytest ran the file in declaration order). Made
both self-contained.
No assertion weakening, no test deletion, no global autouse fixture masking
real bugs. All three fixes preserve the real (non-test) default behavior
byte-for-byte when unpatched/env-unset.
Verification: 8x targeted -n 8 loop over the 7 affected files (before and
after) did not force-reproduce the underlying race live (narrow timing
window, small-scale run) — confirmation is source-level (hardcoded shared
paths bypassing the established isolation idiom) plus the prior documented
flake for test_replay_leaves_no_trace in docs/testing-lanes.md. Full
fast-lane -n auto run recorded in the PR description.
Root cause of demo_composition pin thrash was not cancelled CI runs:
verify_no_global_state_mutation dumped full env_subset before/after
tuples into the lane report, embedding CORE_ENGINE_STATE_DIR temp
paths (and any ambient CORE_* host values). Each hermetic run then
produced a different report SHA.
Fix for good:
- Report key-level env deltas only (+/- / changed), with path-like
CORE_* values redacted to <path> in the message text.
- Detection still uses raw snapshots (mutations are not ignored).
- Hermetic CORE_ENGINE_STATE_DIR for demo_composition and all
verify_lane_shas invocations.
- Dual-run-stable re-pin: demo_composition e2ba2314…,
public_demo 7d8ba0db… (matches CI-observed soft-budget SHA).
- Drop dead CORE_SHOWCASE_SKIP_BUDGET workflow env (soft is default).
- Tests pin the delta format and hermetic-path stability.
Lane: python scripts/verify_lane_shas.py → 9/9 match.
Finding #20 (Super-Blueprint §3.2). surprise_residual was Euclidean Gram-Schmidt
on flat 32-coefficient vectors — metric-blind (it ignored the (+,+,+,+,-)
signature and the blade grade structure), so "inside the admissible span" was
judged by the wrong geometry.
Operator math (core/physics/surprise.py):
- Exact metric-orthogonal projection: solve the normal equations G c = r
(G_ij = cga_inner(b_i,b_j), r_i = cga_inner(b_i,x)) via lstsq, under cga_inner
(32-vec) / eta (5-vec).
- Fail-closed (typed SurpriseResidualError) on a metric-degenerate span, keyed on
rank(G) < rank(B) — a null direction with no reciprocal (lone n_o). Refines the
literal "rank(G) < k": mere linear dependence among non-null columns is admitted
(lstsq projects onto the span), so a redundant live basis [1, source] and the
non-degenerate pair {n_o, n_inf} are admitted; only a lone n_o is refused. The
disclosure names the Gram null-space direction (not just zero-diagonal columns).
- Reconciled productivity polarity: productive_transfer = low Procrustes AND low
surprise (was `sur_norm >= 0.0`, always true). High surprise routes to discovery
(split follow-up). Corrects the ledger's transfer/discovery conflation.
Adversarial verification (3 independent lenses) found, and this fixes:
- HIGH soundness hole: sur_norm was the reversion pseudo-norm, which VANISHES on
a nonzero metric-null residual (the n_o/n_inf light cone) -> false-zero surprise
-> an out-of-span light-cone probe was wrongly admitted as in-span. Now the
DEFINITE (Euclidean) norm of the residual: the projection stays metric-exact,
the magnitude is 0 iff nothing is unexplained.
- HIGH regression: the analogical-transfer harness called surprise_residual
OUTSIDE its try/except, so a degenerate source crashed the whole run. Now
guarded: records a refused case and continues.
- grade-support `allowed` -> exact-nonzero (removes a spurious-leak edge under
coefficient amplification of sub-tolerance grade dust).
DiscoveryCandidate wiring into the contemplation loop is split to its own
follow-up (a distinct cross-cutting surface). Off-serving (nothing in
serving/runtime imports core.physics.*). Tests: 15 behavioral tests
(metric-vs-Euclidean divergence, null-cone regression, null refusal +
combination-degenerate disclosure, 5-vector branch, polarity); 139-test physics
sweep green; ruff clean.
Adds policy document and sealed manifest records for the first 8
external audit datasets. No data is vendored. Local cache paths are
gitignored. Fetch/verify scripts and smoke fixtures come in PR-2.
Datasets: GSM1K, ASDiv, SVAMP, PARA-MAWPS, ARC-Easy, ARC-Challenge,
OpenBookQA, CLUTRR
- evals/gsm8k_math/train_sample/v1/report.json: rebaselined from
correct=6/refused=44/wrong=0 to correct=30/refused=20/wrong=0.
wrong_ids == []. exit_criterion.passed: false → true.
Written deterministically by the canonical runner (write_report /
json.dumps sort_keys=True).
- docs/sessions/gsm8k-train-proxy-rebaseline-2026-06-20.md: new audit
note documenting why #846's report update was unauthorized, why #847
restored the pin, and why this PR is the explicit authorization
boundary for the committed rebaseline.
- HANDOFF-gpt55-2026-06-20.md: small factual note appended to the
'Latest continuity pointer' section recording this rebaseline status.
Authorization chain:
ADR-0126 (exit criterion) →
#846 kernel changes (produce 30/20/0) →
#847 reconciliation (restores pin, delegates to this PR) →
this PR (explicit rebaseline authorization)
Non-goals: no runtime, serving, teaching, capability, or kernel change.
No quantity-entity work. No new eval artifacts. No teaching proposals.
- Restore evals/gsm8k_math/train_sample/v1/report.json to pre-#846 pin (6/44/0)
The 30/20/0 committed in #846 was generated eval output, not a ratified rebaseline.
Historical pin is required until a dedicated rebaseline PR is explicitly authorized.
- Delete teaching/proposals/comprehension_failures/34ce9254...json
Newly added in #846 without authorization. File did not exist in first parent.
HANDOFF claimed teaching paths were untouched — contradicted by the actual diff.
Proposal must go through the reviewed teaching lifecycle, not committed PRs.
- Add docs/sessions/pr846-reconciliation-audit-2026-06-20.md
Full audit note: what #846 was expected vs. what merged, authorization analysis,
restoration rationale, what was preserved, remaining risk, next correct step.
Tests kept from #846:
- tests/test_problem_frame_builder.py (lint fix: _frame_names -> direct accessor)
- tests/test_problem_frame_skeleton.py (unused ProblemFrame import removed)
Both are valid narrow fixes unrelated to the unauthorized artifacts.
Forbidden paths not touched: algebra, field, vault, recall, identity, policy,
packs, generate/derivation, generate/math_candidate_graph.py, runtime/serving.
- merge_compacted_runs merges key-by-key without re-expanding prior counts
- load default train_sample cases for live operation_class resolution
- hash full scout row evidence in source_report_hash and source_run_id
- replace scripts.gsm8k_frontier_report import with local _extract_category
Introduce deterministic practice-memory infrastructure that adapts sealed
scout output into compact, retention-gated ExperienceRecords with family,
hazard, and promotion-candidate summaries. No serving, corpus, pack, or
report.json mutation — measurement-only adapter for future sprint reuse.
Gate A1 and related cognition-path changes shifted the showcase
determinism fingerprint (86db25b6…) while the lane pin still reflected
the pre-drift artifact (e323adb3…). Re-pin to the CI-verified SHA
2895df08… and refresh v1_dev.json + CLAIMS.md.
Hermetic demo hygiene: use no_load_state for register-tour, learning-
loop, and multi-hop showcase adapters, and isolate CORE_ENGINE_STATE_DIR
in the public_demo lane runner so local engine_state/ pollution cannot
affect eval output or wall-clock.
* feat(derivation): Workstream A inc 1 — lexeme-only reader components for fractions/comparatives + cleaned Phase B exemplars + proxy rebaseline (ratified, post-review fixes)
- generate/derivation/extract.py: lexeme-only passes (surface tokens only; no synthesized values or non-surface source_tokens for "half of", "X/Y of", "X more/less than"). Composition deferred per module contract (ADR-0165). Hygiene and EX-6 preserved. Docstring updated.
- teaching/admissibility_exemplars/*_v1.jsonl: grown + cleaned (removed category-polluted seeds, deduplicated ma IDs, fixed comparative exemplar to surface tokens per lexeme rule).
- evals/gsm8k_math/train_sample/v1/report.json: rebaselined after fixes (runner executed; wrong=0 held).
- docs/analysis/*-2026-06-1[6-7].md + lookback: ratif + honest post-review lookback (local paths removed, scope claims match actual clean diff, extractor/exemplar fixes documented).
Governing: the two 2026-06-16/17 ratifications (included; local session path replaced with stable ref). Scope per ratif: reader (lexeme) + exemplar growth (clean) + rebaseline + tests + lookback. No sealed, no CLOSE (deferred), 0 wrong.
Added tests in test_adr_0179_extract.py for new lexeme paths + grounding.
Branch prepared clean from origin/main; diff limited to scoped files only (no posture ratif, no runtime_contracts drift).
See updated lookback and PR body for full evidence (6/44/0 or current post-fix, tests green, etc.).
(Addresses review: Semantic Rigor restored, branch hygiene, exemplar quality, tests, docs.)
* fix(review): resolve all verifiability blockers
- generate/derivation/extract.py: conflict markers removed (was from stash); lexeme-only implementation confirmed clean.
- tests/test_adr_0179_extract.py: added TestWorkstreamAReaderLexemeOnly class with direct tests for half-of, fraction-of, more/less components, source_token grounding, no-synthesis invariants.
- evals/gsm8k_math/train_sample/v1/report.json: fresh rebaseline via runner with fixed code (6/44/0, wrong=0).
- docs/analysis/gsm8k-workstream-a-increment-1-lookback-2026-06-17.md: updated to exactly match the actual diff and actual report numbers on this head.
- No conflict markers remain in extract, tests, lookback, ratifs, or exemplars (verified by grep).
- Branch head now clean and matches all claims in PR body/lookback.
This is the verifiable head for the lead engineer review.
* fix(test): make lexeme-only half test case-insensitive on source_token to match actual extraction (preserves input casing like other EX tests)
* fix(review): commit current report.json (6/44/0 from runner) and updated lookback so they appear in branch diff; lookback now exactly matches actual files in git diff --name-only and actual test count (29) and report numbers.
* fix(review): precise fractional doc in extract contract (normalizes factors while preserving source_token; comparative deferred); remove stale comparative sum/diff comments; lookback exact reality (29/29 tests, no deep-dive in diff, no sum/diff claim); mark remaining exemplars REVIEW-REQUIRED/ambiguous with explicit notes. Small final patch.
* fix(exemplars): remove three REVIEW-REQUIRED ambiguous-but-machine-admissible seeds (rwc-v1-0022, rwc-v1-0023, ma-v1-0024) from ingested JSONL to eliminate synthesis-poisoning risk per final review. Update lookback wording for actual seed count and rationale. Small final patch.
* docs(lookback): align exemplar category list with actual diff (only dcs/rwc/ma touched in this increment; no ca/ta/dsnq changes)
- Ratified first (docs/analysis/close-flywheel-proposal-review-visibility-ratification-2026-06-16.md) before any impl edits.
- Additive proposal_review_posture in evals/close_derived_climb (from already-captured emission bodies; asserts born proposal_only/SPECULATIVE/requires_review posture, review_eligible, none_accepted_or_promoted inside the yardstick).
- Additive proposal_review_summary in evals/anti_regression DemoReport + run_demo (teaching gate review_states + ProposalLog transition counts from the temp log events exercised by S1-S3; close_derived subsection from the embedded climb).
- Updated RESULT text, contract test docstring + soft presence assert, climb __init__/contract, testing-lanes (new subsection), anti-regression docs, runtime_contracts.
- All strictly within scope: measurement/visibility only; no changes to review logic, FrameVerdict, promotion, teaching mutation, or core behavior.
- Verified: 90 contract tests green, core demo anti-regression JSON + text show the signals, climb posture present with correct values, wrong_total=0, all_gates_held + corpus_identical preserved, smoke lanes unaffected.
- Builds on #793 dedicated surface + #792/#791 embedding. Third Door / Mechanical Sympathy / Semantic Rigor alignment per ratif.
- No brief adjustments needed; workflow, scope, and success criteria followed exactly.
- Makefile: new 'test-close-flywheel' target (clearly named heavy surface; runs full Claim-B yardstick + anti embedding; explicitly not in fast/slow/full).
- docs/testing-lanes.md: new prominent 'Dedicated CLOSE Flywheel Regression Surface (Claim-B Level)' section with purpose, invocation (make primary), full Claim-B capabilities list, runtime (~60s+ heavy), hermeticity guarantees, Engineering Pillars alignment (Mechanical Sympathy / Semantic Rigor / Third Door per Whitepaper), and all cross-refs (ratifs, contract, anti demo, etc.).
- Cleanly embed / polish (additive, hermetic, building on #792): anti_regression/run_demo.py (comments + RESULT label the surface + make target), test_anti_regression_demo.py (updated docstring), docs/evals/anti_regression_demo.md (how-to + complementary section updated for surface participation).
- Supporting: contract.md (dedicated surface note + ratif links), runtime_contracts.md (tightened ref).
Ratified first (docs/analysis/close-flywheel-dedicated-regression-surface-ratification-2026-06-16.md). Strictly in scope. Preserves all invariants (wrong_total=0, determinism, proposal-only, etc.). No core, no CLI additions, no fast/CI inclusion, no teaching logic changes.
See ratification for pillars justification and 'why only correct path'.
- Semantic: determine() asserts (rule=direct) on positives post-FP in climbs
- Lived flag: _proposal_flag_effect now uses real idle_tick + IdleTickResult.derived_close_proposals_emitted (temp sink isolation)
- Checksum: added content_replay_checksum (canonical closures + proposal bodies)
- Docs: contract.md + module docstrings updated for accuracy
Ratification artifact created before impl (see docs/analysis/...-ratification-...md).
All original metrics, wrong_total=0, invariants preserved.
Closes the 3 gaps from post-merge audit for Claim B.
* feat(eval): add yardstick lane for CLOSE derived climb (is-a + relational + flag + wrong=0)
* fix(eval): correct proposal flag measurement in yardstick (direct emit for isolation; runner now reports True)
* fix(eval): measure CLOSE climb from live queries
* fix(eval): require strict CLOSE climb growth
* wip(b4): lift _basis to generate/epistemic_basis.py (behavior-preserving)
B4 PR-1 foundation checkpoint — the shared epistemic-basis helper extracted
byte-identically from determine.py so the closed-world frame_verdict evaluator
(coming) can compute standing without importing determine (ADR-0222 §8 A3).
determine.py imports it aliased; its 3 Determined sites + INV-30 unchanged.
PAUSED: rest of B4 (FrameVerdict type, evaluator, INV-31, PR-2/3/4) resumes
after the suite-speed arc.
* feat(frame-verdict): closed-world envelope + isolated evaluator + INV-31 (B4 PR-1)
The sealed, off-serving FrameVerdict type + the isolated text-frame evaluator + the
INV-31 firewall (ADR-0222). Composes proof_chain.entail (no second prover); a
closed-world False comes ONLY from an ROBDD refutation; absence is never false.
- generate/frame_verdict/{types,evaluate,__init__}: FrameVerdict / ClosedFrame /
ClosedWorldProof + PositiveRefutationKind enum. evaluate_frame_verdict(frame, query)
maps entail ENTAILED/REFUTED/UNKNOWN/REFUSED -> entailed_true/false/undetermined/
contradiction/scope_boundary; OPEN / undeclared-closure / non-TEXT -> scope_boundary.
__post_init__ admissibility: entailed_false needs a NAMED positive refutation; a
generic FALSIFIED raises. trace_hash is order-invariant + replay-stable.
- INV-31 (test_architectural_invariants.py): A1 determine.py clean+visible; A2 exact
construction allowlist (evals-inclusive, tests-excluded) + non-vacuity anchor; A3
transitive spine -/-> frame_verdict (reuses INV-27 walker) + resolve anchor; B1
determine() refuses a ClosedFrame. B2 deferred to the governance slice (no
non-vacuous surface yet; ADR §8 permits).
No serving wire. determine.py unchanged except the _basis import. No Determined(answer=False).
INV-30 green. Shapes follow the B4 operator master brief (PositiveRefutationKind enum,
closure_declared/source/provenance) — a consistent refinement of ADR-0222 §3.
Verified: 97 (full INV incl. INV-31 + FrameVerdict + OWA floor) + 20 frame_verdict + 35 determine.
* feat(frame-verdict): text closed-world (CWA) evaluation lane (B4 PR-2)
A measure-only lane over the text FrameVerdict evaluator (ADR-0222). Proves the sealed
type evaluates propositional closed frames safely — incl. a sound entailed_false —
without touching determine() or runtime serving.
- evals/frame_verdict_text_cwa/: cases.jsonl (12 cases across all 5 verdict kinds +
gating + absence safety), oracle.py (an INDEPENDENT truth-table propositional checker —
own recursive-descent parser + brute-force enumeration, disjoint from the ROBDD; imports
no engine module), score.py, README.
- tests/test_frame_verdict_text_cwa_lane.py: wrong=0; the disjoint oracle confirms every
gold (non-vacuity) AND the engine matches the oracle on every case; entailed_false is
proof-backed (ROBDD_REFUTATION); absence / OPEN / undeclared-closure are never
entailed_false; SHA-pinned fixtures.
Input contract: propositional-formula strings (no prose lowering). Capability-index
deliberately NOT touched (off-serving lane; documented). No Determined, no answer=False,
no serving wire. INV-30 / INV-31 / ProofWriter-OWA stay green.
* feat(frame-verdict): perception changed-slot falsification adapter (B4 PR-3)
Lift an ADR-0211 FalsificationRun into a FrameVerdict SAFELY (ADR-0222 §5.2). Critical
doctrine: "FALSIFIED" is not enough — only a POSITIVELY observed changed-slot contradiction
produces entailed_false (PERCEPTION_CHANGED_SLOT). Missing observation (absence), unexpected
extra (over-observation), and a whole-missing actual frame NEVER become false — they refuse
(undetermined / scope_boundary). SUPPORTED -> entailed_true (frame-conformance proven).
- generate/frame_verdict/perception_adapter.py: frame_verdict_from_perception_falsification;
the proof carries the FULL run trace_hash (binds expected+actual+verdict, ADR §5.4).
- generate/frame_verdict/_construct.py: extracted the single FrameVerdict builder; the text
evaluator + perception adapter both funnel through it, so the literal FrameVerdict(...)
lives in ONE file -> INV-31 ALLOWED_FRAME_VERDICT_SITES = {_construct.py}.
- tests: changed-slot -> entailed_false (+ empty-hash raises); missing/unexpected/whole-missing
never false; supported -> entailed_true; non-perception -> scope_boundary.
Off-serving; perception_adapter is in generate.frame_verdict, so INV-31 A3 already proves it
unreachable from the open-world spine. No determine() change, no answer=False. INV-30 / INV-31
/ ProofWriter-OWA / text-CWA lane stay green.
* feat(response-governance): default-dark FrameVerdict surface mapping (B4 PR-4)
The only lawful surface path for a closed-world verdict (ADR-0222 §7/§14). Lowers a
FrameVerdict to a served disposition through the EXISTING epistemic_disclosure tables (no
parallel object): entailed_true/false -> COMMIT at INFERRED + DisclosureClaim.NONE (a
committed "Yes"/"No"; entailed_false is an answer, NOT a contradiction/refusal/LimitationKind);
contradiction -> REPORT; undetermined -> REFUSE; scope_boundary -> EXPLAIN.
DEFAULT-DARK: a NEW module that changes no existing file — the open-world govern_response /
shape_surface STRICT path is byte-identical, and nothing in the live runtime calls it. The
TYPE is the closed-world tag: a forged dict / untagged object cannot widen serving (raises).
INV-31: A3 still green — core/response_governance/frame_verdict.py is NOT imported by the
package __init__ or the open-world spine, so frame_verdict stays unreachable from
chat/runtime/session/vault. The B2 anchor is now live (the forged-object rejection); all six
INV-31 anchors firm. The open-world render_determination cannot render a FrameVerdict;
determine() still has no answer=False.
Verified: 84 (governance + full INV incl. INV-31 A3/B2) green.
* harden(frame-verdict): fold in adversarial-review findings (B4)
A 4-skeptic adversarial read of PR-1..4 against real source surfaced one
soundness gap (major) and four defensive gaps (minor). All folded in; the
4 B4 files + full INV suite stay green (128 passed).
S1 (major) — perception negation was not frame-gated. The text evaluator
refuses OPEN / undeclared-closure frames (-> SCOPE_BOUNDARY) but the
perception adapter only gated frame_kind + the missing-frame sentinel, so a
PERCEPTION + OPEN + changed-slot residual produced entailed_false with
world_assumption=OPEN — a verdict that self-contradicts the type invariant
(OPEN => negation illegal). Fixed two ways:
* perception_adapter: gate OPEN / not-closure_declared -> SCOPE_BOUNDARY,
mirroring the text evaluator (graceful upstream refusal);
* types.__post_init__ §(0): STRUCTURAL backstop — ENTAILED_FALSE + OPEN
raises, frame-general, so NO producer (text/perception/future) can emit
an OPEN-world negation even if it forgets the gate.
S4-a — entailed_true was admissibility-asymmetric. Only entailed_false was
proof-gated at construction; a committed "Yes." leaned entirely on the
INV-31-A2 allowlist. Added a symmetric §(2) guard: entailed_true requires a
positive entailment/support proof (proof_chain.entail/ENTAILED or
sensorium.falsification/SUPPORTED), non-empty sha, and NO refutation kind —
a mutation test can now trip a forged positive.
S2 — A2 construction detector extended to flag FrameVerdict.<factory>(...)
classmethod constructions and module-qualified mod.FrameVerdict(...), so a
future alternate constructor cannot evade ALLOWED_FRAME_VERDICT_SITES. The
dataclasses.replace gap is RECORDED (a blanket replace() match would
false-positive tree-wide) for the serving-wiring PR, not faked.
S4-b/c — A3 firewall now bars core.response_governance.frame_verdict
DIRECTLY (not only via its re-import of generate.frame_verdict), plus a new
test proving the response_governance package __init__ stays default-dark
(does not re-export the adapter).
S3 — oracle grammar is a strict SUBSET of proof_chain.entail (it mis-parses
`false` as a free atom). Corrected the docstring and added a lane guard:
every DECIDED cases.jsonl formula must stay inside the subset, so a future
out-of-subset case fails loudly at SHA-add review instead of as a confusing
engine-vs-oracle red. SCOPE_BOUNDARY-gold garbage is exempt (both refuse).
S1-minor (construction-time-only admissibility) documented in __post_init__:
a future codec/deserialization path must re-construct via build_frame_verdict.
* docs(analysis): B4 FrameVerdict implementation lookback (PR-1..4 + hardening)
Mandatory multi-slice lookback (CLAUDE.md triggers 2+3). Audits all six B4
commits across the lookback template: documentation drift, test-coverage /
parity gaps, wrong=0 hazard surface, cross-slice consistency, honest LOC.
Categorized findings: solid (INV-30 untouched; entailed_false positive-only;
INV-31 firewall non-vacuous; default-dark; replay determinism) / hazards (the
5 adversarial-review findings, ALL fixed in the hardening commit) / recorded
gaps (dataclasses.replace A2 hole + construction-time-only admissibility,
both deferred with an explicit guard obligation on the serving-wiring PR) /
drift (type shapes are a consistent refinement of ADR-0222 §3; one non-gating
ADR note filed). Global red-line ledger: all green.
Add a closed finite-verb relational surface to comprehend_relational, narrowly:
declarative <A> overlaps <B>, interrogative Does <A> overlap <B>?, predicate
overlaps_event ONLY (closed finite-verb table, default-off).
Kept SEPARATE from the copula-connective grammar (byte-unchanged): the other
connectives (before/after/during/inside/adjacent/…) still REQUIRE the copula, so
"Monday before Friday." stays a refusal — no connective bypass (adversarial hazard #2).
Fail-closed slot gate: each finite-verb argument slot must be EXACTLY ONE content
token (after article stripping). An enumerated adverb/negation blocklist is unbounded
and leaks — the adversarial audit found "Meeting never overlaps lunch." committed as a
POSITIVE overlaps_event(meeting_never, lunch), plus almost/sometimes/trailing-qualifier/
second-verb compound fabrications in BOTH slots and the query path. The single-token
gate closes the whole COMPOUND-fabrication class (any extra token refuses); the
_FINITE_VERB_MODIFIERS and _CONNECTIVE_TOKENS checks add precise reasons for common
bare-modifier / second-verb slots. Multi-word entities in the finite-verb surface are
deferred (they need a positive content lexicon, which OOV entities preclude). A lone
bare token remains an entity (the reader's universal OOV single-token contract — the
copula and general readers behave identically), not a compound fabrication.
Hazards pinned (adversarial audit hazard #1 + #2): adverb absorption, negation-as-
positive ('never'), interrogative double-verb, trailing qualifier, and the copula-bypass
firewall.
Measurement: migrated ref-009 "Sunrise overlaps dawn." (the documented coverage gap) to
a positive rel-018 — comprehension_relational_predicate 17->18, wrong_total 0, breadth
still 11 (coverage added, not a new domain), baseline digest re-frozen. Refusal floor
rose 9 -> 24 (finite-verb confusers replace the one migrated input). The #775 inference
and #781 transitive lane SHA pins are UNTOUCHED. determine.py is untouched (reader-only):
no answer=False, no FrameVerdict, INV-30 unaffected.
Verified in the worktree: finite-verb + reader + reader-lane + #775 inference lane +
#781 transitive lane + #779 OWA floor + capability baseline/index + INV-30/29/21 + full
smoke — all green. Two read-only adversarial passes: the first found 4 real slot-
fabrication blockers, closed by the single-token gate; a black-box probe of 24
adversarial inputs confirms the compound-fabrication class is fully closed.
* feat(determine): add transitive strict-order relational inference
Add sound transitive closure for declared strict-order relational predicates —
less_than, greater_than, before_event, after_event, and ONLY these. A query
p(a, c) determines True when a same-predicate chain p(a, b), p(b, c), ... over the
predicate's OWN realized edges is found (BFS reachability) and the proof_chain ROBDD
verifies the transitive entailment (search-then-verify, through the UNCHANGED
evaluate_entailment via a new lower_transitive_chain). Mirrors _determine_subsumption.
Scope / firewall:
- TRANSITIVE_PREDICATES is closed and default-off; sibling_of, parent_of,
left_of/right_of, inside_of, during_event, overlaps_event are EXCLUDED.
- Same-predicate edges only — no transitive-through-inverse, no cross-predicate
(triple firewall: recall filter + lowering re-reject + ROBDD re-verify).
- Open-world: asserts only answer=True via the shared _relational_determined
surface (rule="transitive"), so INV-30 stays at exactly 3 Determined sites and
no answer=False is added. One-hop inverse/symmetric (#775) is unchanged.
FrameVerdict / closed-world (ADR-0222) is untouched.
wrong=0 bite (unit + lane + independent oracle): non-transitive-predicate chains,
non-admitted spatial chains, mixed-predicate chains, disjoint chains, reflexive
cycles, and inverse+transitive composition all REFUSE.
Measurement: new evals/relational_transitive lane with an INDEPENDENT BFS
transitive-closure oracle (imports no engine module; INV-25/27), cross-checked both
directions (positives oracle-True, confusers oracle-False). Capability-index breadth
10 -> 11, wrong_total 0, deterministic digest re-frozen.
Cross-PR reconcile: migrated the one-hop confuser rinf-ref-001 (a less_than chain,
which now correctly determines transitively) to a B2 positive; repurposed its unit
test to a non-transitive parent_of chain; corrected a pre-existing stale breadth==9
assertion (#775 added a 10th domain but never updated test_capability_index).
Verified in the worktree: determine + transitive/one-hop lanes + capability
baseline/index + ProofWriter-OWA floor + INV-30/29/21/02 + full smoke — all green.
A 3-skeptic read-only adversarial review found no wrong=0 leak.
* test(determine): B2 hardening — TRANSITIVE_PREDICATES pin + lower_transitive defensive + budget
Add the required B2 hardening tests before merge:
- test_transitive_predicates_closed_and_excludes: the table is EXACTLY
{less_than, greater_than, before_event, after_event}, every member is in
RELATIONAL_PREDICATES, and every deliberately-excluded predicate (sibling_of,
spouse_of, parent_of, child_of, left_of, right_of, inside_of, during_event,
overlaps_event) is explicitly absent. (Closes the gap where relational.py's
comment claimed this test existed but it did not.)
- tests/test_composition_lower_transitive.py: direct defensive tests on
lower_transitive_chain — empty path, mislabeled (cross-predicate) edge, wrong
arity, non-contiguous path, and path-not-reaching-target all lower to None
(refuse); plus exact 2-hop and 3-hop theory pins.
- test_edge_budget_exhaustion_refuses: above _TRANSITIVE_EDGE_BUDGET the transitive
search declines (a safe coverage refusal), never proves.
* feat(eval): ProofWriter-OWA refusal-floor lane (B1) — independent oracle, measure-only
Mastery-v2 Step-3 Brief 1. Proves the open-world soundness floor: determine() never
asserts a query True when the open-world truth is Unknown or False. Hardens
"unknown != false" before the transitive-chain (B2) and closed-world (B4) work can
stress it.
- evals/proofwriter_owa/oracle.py: a SEPARATE minimal OWA label oracle (its own parser
+ reasoner + inverse/symmetric tables), importing NONE of determine, comprehend/
MeaningGraph realization, or production predicate-entailment helpers (INV-25/27 — the
gold producer is disjoint from the solver). Grammar: member/subset/is-a closure,
explicit negation (No X is a Y -> disjointness, the source of gold-False), and #775
inverse/symmetric ONE-HOP relational rules. No transitive relational chains.
- evals/proofwriter_owa/fixtures.jsonl: 19 hand-authored ProofWriter-OWA-style items
(9 True / 7 Unknown / 3 False), SHA-pinned. Gold computed by the oracle; the fixture's
hand-authored `expected` pins the oracle (so the oracle is itself verified).
- evals/proofwriter_owa/score.py: runs the production path (comprehend/comprehend_relational
-> realize -> determine) vs the oracle gold; wrong = determine asserted True on a non-True
gold.
- tests/test_proofwriter_owa_lane.py: SHA pin; oracle==expected; wrong==0; every
serving_support gold-True determines True (no coverage gaps); no answer=False (INV-30).
Result: 19 items -> 9 correct (all serving_support True), 10 refused (Unknown/False),
0 wrong. Measure-only: no engine code touched; deliberately NOT a capability-index domain
(a refusal floor has low coverage and would drag coverage_geomean). INV-30 green; smoke 99/99.
Note: no live ProofWriter dataset access in this environment, so items are hand-authored
in OWA style (provenance.md cites ProofWriter V2020.12.3 / arXiv:2012.13048 for semantics,
attribution only) — but the GOLD is oracle-computed and verified, not hand-asserted.
* harden(eval): OWA lane CLI fails on coverage_gaps too (review #779)
score.py::main() exited nonzero only on wrong>0; it now ALSO exits nonzero when
coverage_gaps is non-empty (serving-supported gold-True items that refused), matching
the acceptance rule the tests already enforce and the module docstring. Both failure
modes print to stderr; exit is 1 if either fires. Verified non-vacuously (injected gap
-> exit 1). Lane test 5/5 green.
* feat(determine): one-hop sound relational entailment — inverse/converse + symmetric
Mastery-v2 Step 3 lead capability (core). DETERMINE could perceive relational
structure (16 predicates, breadth-complete) but could not derive the simplest
entailed relational facts — it read only the stored direction. This adds two
SOUND one-hop rules that read a stored edge in its other lawful direction:
INVERSE/converse greater_than(a,b) <= told less_than(b,a)
SYMMETRIC sibling_of(b,a) <= told sibling_of(a,b)
Strictly scoped (per review): OPEN-WORLD (asserts only True, never False — the
answer=False path stays unbuilt, INV-30's firewall holds), ONE hop (NO transitive
chaining), DECLARED rules only. Ontology/metadata-driven, not prose intuition:
- generate/meaning_graph/relational.py: declarative algebra tables — _INVERSE_PAIRS
(the converse edges; the pack carries no inverse metadata) + INVERSE_OF (derived,
an involution) + SYMMETRIC_PREDICATES (mirrors the pack's graph.edge.symmetric tag).
load_relational_pack_symmetric() reads the pack ontology; a test pins the constant
equal to it (no silent divergence).
- generate/determine/determine.py: _relational_one_hop() between direct entailment
and transitive subsumption; new Determined.rule provenance field (direct/inverse/
symmetric/subsumption) — replay-safe (render reads only basis; trace hashes surface).
wrong=0 confuser block proves the rule cannot over-fire: less_than is not
self-inverse, sibling_of does not imply parent_of, greater_than does not imply
equal_to, NO transitive chain (direct or through-inverse), and no answer=False is
ever emitted. An obsolete pin (test_symmetric_converse_is_not_faked, which asserted
the pre-rule direct-only floor) is updated to the new sound behavior and a preserved
asymmetric-converse bite.
Tests: test_determine_relational_inference.py (13, the capability contract) +
updated test_relational_reader.py. Broad relational/determine/grounding +
capability_index baseline sweep: 114 green; baseline digest UNCHANGED (this is a
determination capability; the comprehension lanes are untouched).
REMAINING for the full lead PR (next): the measurement lane — an independent-oracle
evals/relational_inference lane + a capability_index adapter + baseline re-freeze
(breadth 9->10, wrong_total still 0), so the capability registers on the yardstick.
INTEGRATION NOTE (cross-PR): this adds 2 Determined() construction sites (now 4:
direct/inverse/symmetric/subsumption, ALL answer=True). When rebased onto main with
INV-30 (PR #770), update INV-30's test_determine_construction_sites_are_visible count
2 -> 4 and confirm all four assert True. Merge order: #770 -> this.
* feat(capability-index): relational-inference lane — breadth 9->10, wrong=0
Mastery-v2 Step 3 lead, measurement half. Puts the one-hop relational-inference
capability ON the capability index with an independent-oracle lane:
- evals/relational_inference/v1/: cases.jsonl (13 positive: 8 inverse + 5 symmetric,
authored from the relation algebra INDEPENDENTLY of determine/relational per
INV-25/27), refusals.jsonl (8 confusers that MUST refuse), provenance.md (incl. the
honest overlaps_event reader-surface coverage gap).
- evals/comprehension/relational_inference_runner.py: told fact(s) -> determine(query)
vs gold; refusal = coverage miss, disagreement = wrong (structurally 0 on positives).
- evals/capability_index/adapters.py: comprehension_relational_inference_result added
to ADAPTERS.
- baseline.json re-frozen: breadth 9 -> 10, capability_score 0.94403 -> 0.949483,
wrong_total still 0, digest deliberately changed (35dea2b2...).
- tests/test_relational_inference_lane.py: SHA-pinned gold, positive wrong=0 +
coverage>0, and the confuser wrong=0 BITE (every confuser must refuse).
Also reconciles INV-30 (now in main via #770): determine.py grew from 2 to 3
Determined() construction sites (direct, the shared relational one-hop constructor,
transitive subsumption) — ALL answer=True. test_determine_construction_sites_are_visible
updated 2 -> 3 (correcting the prior commit note's overcount of 4; inverse and symmetric
share the single _relational_determined constructor, so it is ONE new site).
Also fixes a .gitignore gap: ADR-0219 gen-dir checkpoints (engine_state/current,
engine_state/gen-*/) were not ignored (only the old flat engine_state/*.json patterns
were), so runtime checkpoints could be committed by accident.
Verification: new lane 13/0/0 coverage 1.0; capability_baseline digest matches the
re-freeze; INV-30 + architectural invariants green; smoke 99/99.
* docs+test: reconcile stale relational-inference docstrings; assert rule provenance (review #775)
Addresses two review-requested cleanups on #775:
1. Stale docstrings. generate/meaning_graph/relational.py said the symmetric converse
is "a sound-but-incomplete refusal at DETERMINE" and there is "NO transitive/
symmetric/rule inference"; generate/determine/determine.py said "symmetric-converse
questions are Undetermined", "no transitive/symmetric/rule inference", and "asserts
only answer=True on a direct hit". All false after the one-hop relational algebra
landed. Updated: the reader does direct-reading only; DETERMINE applies declared
one-hop inverse/converse + pack-declared symmetric (plus the existing member/subset
subsumption), still open-world / never answer=False; transitive relational closure,
negation, and closed-world falsehood remain out of scope.
2. Rule-provenance gold now meaningful. The relational_inference runner compared only
(answer, predicate, subject, object), ignoring res.rule though the gold carries
"rule": "inverse"/"symmetric". Added res.rule to the got/gold tuple so a "right
answer via the wrong rule path" can no longer pass silently. Lane still 13/0/0;
baseline counts (and digest) unchanged.
Verification: runner 13/0/0; lane + capability_baseline + unit + INV-30 green (22).
Removes code_revision from the engine-identity hash: engine_identity is now the
sha256 of the 5 ratified packs ONLY. The build revision is provenance (the
manifest's written_at_revision), not identity — so a behavior-neutral rebuild is
the SAME identity and the always-on daemon no longer flag-day strict-breaks on
every commit (the ADR-0220 defect).
Core:
- core/engine_identity.py: ratified_substrate/compute_engine_identity drop the
git_revision arg (packs-only); add compute_legacy_engine_identity (reproduces
the pre-split packs+rev hash, for migration verification), ENGINE_IDENTITY_SCHEME=2,
IdentityReconciliation enum, and reconcile_loaded_identity — the single source of
truth for the runtime guard AND the workbench reader.
- chat/runtime.py: the load guard reconciles via scheme. Current scheme -> direct
packs-only compare. Legacy (code_revision-folded) stamp -> a VERIFYING migration:
reconstruct the legacy hash from the persisted written_at_revision; a match proves
packs unchanged (warn + re-stamp, resume, no break) — a mismatch means the packs
genuinely changed (DIVERGED -> strict-refuse). Preserves 'distinct packs => refuse'.
- engine_state/save_manifest: stamp identity_scheme alongside engine_identity
(additive-optional; no schema_version bump).
- workbench/readers.py: continuity reader uses the same reconcile (no phantom break
on legacy checkpoints). cli.py break message reworded (packs, not 'build revision').
Callsite cleanup: drop the now-unused git_revision arg + imports across
always_on.py, evals/l10_always_on/runner.py, workbench/readers.py.
Tests:
- test_identity_provenance_split.py (new): 5 reconcile unit proofs + 3 runtime
integration proofs incl. the wrong=identity defense (legacy stamp of DIFFERENT
packs still strict-refuses) and the re-stamp migration.
- test_engine_identity.py: invert the code-revision test (rev no longer changes
identity); assert the substrate is packs-only.
- Restore 3 lineage tests silently red since ADR-0219 (flat-path _manifest helper
now resolves the gen-dir).
- Update remaining callsites/monkeypatches for the new signature.
Hygiene: .gitignore now covers the ADR-0219 gen-dir runtime files
(current, gen-*/, session_state.json, proposals.jsonl).
Verified: 59 identity/migration/lineage/workbench + 32 L10 + 76 invariants/cli +
34 smoke pass; serving lane SHAs unchanged (no derivation/reliability_gate touch).
Adds empirical coverage of the three ADR-0219 checkpoint sub-steps where
a process kill can occur: PARTIAL_GEN (gen dir exists but partially written),
FULL_GEN_BEFORE_SWAP (all four files written but current pointer not yet swapped),
and AFTER_SWAP (clean-commit control case, already covered by rec_a/rec_b).
For each of the two non-trivial cut-points the harness:
1. runs a probe soak to commit reboot_turn turns,
2. injects the corresponding orphan shape,
3. reads the on-disk manifest to confirm the loader follows `current` and reads
the committed generation (not the orphan),
4. runs two independent recovery soaks (each injecting the same orphan at the
reboot boundary) and verifies their post-reboot trace_hash tails converge,
5. checks versor_condition < 1e-6 throughout every recovery soak.
New predicate evaluate_p4_arbitrary_interruption (predicates.py) expresses
this as a PredicateOutcome with a CutPointEvidence carrier. It has:
- holds gate: recovered_turn_count == expected (orphan was ignored)
- holds gate: tail_hashes_a == tail_hashes_b (two recoveries converge)
- holds gate: all versor_conditions < ceiling (closure throughout)
- bites gate: tested by four mutation tests (wrong count, diverging tails,
vc violation, empty tails)
contract.md: all spec predicates P1-P5c are now covered (NOT_COVERED is empty).
Per CLAUDE.md schema-as-proof: every predicate gate is verified by a *_bites
mutation test — 9 tests total (3 injection unit, 2 real-soak holds, 4 bites)
all pass in 109s.
Adds the P5a recall-precision predicate that was listed as NOT_COVERED in
the L10 continuity lane. Closes the gap in the schema-as-proof discipline
(CLAUDE.md): every predicate must have both a *_holds test (real soak) and
a *_bites mutation test so a passing lane cannot silently miss the violation
it nominally catches.
Changes:
- runner.py: ProbeRecord dataclass + probe_at/verify_probes_at params on
run_soak(). Registers a field state (float32 bytes, matching vault's
_exact_index dtype) at a named turn and recalls it against the vault at a
later turn — including after a reboot, which is the cross-reboot claim.
- predicates.py: evaluate_p5a_recall_precision — fails if any ProbeRecord
has rank=None or rank>top_k, or if no cross-reboot probe was recorded.
- report.py: wires P5a into build_report(); probe registered at turn 1,
verified at reboot_turn+2 (intentionally before the vault's
null_project auto-reproject cycle at store_count=20, which would destroy
all CGA inner-product scores — documented as a real finding, deferred to
a follow-up increment); NOT_COVERED is now empty ().
- contract.md: P5a row in the predicate table + reprojection-boundary
scope note.
- test_l10_continuity.py: 4 tests — holds (real 6-turn soak across reboot)
+ 3 bites (rank=None, no cross-reboot probe, empty probe_records).
Key finding: vault.null_project() fires every vault_reproject_interval=20
stores and produces CGA-orthogonal versors (inner product → 0.0 with the
original), completely destroying both exact-match and ranked recall. This
is a long-horizon vault stability issue, recorded here rather than silently
avoided. The P5a probe window is constrained to the pre-reproject interval
to keep wrong=0 intact while documenting the gap.
Closes the cross-file checkpoint-atomicity gap in ADR-0156. The four
checkpoint files now live in a committed gen-NNNN/ directory; the single
atomic os.replace of a 'current' pointer file is the commit boundary.
A kill before the pointer swap leaves the prior committed generation
intact; a kill after commits the new generation. Unreferenced gen dirs
are ignored. ADR-0156's deferred parent-dir fsync is also closed.
Key changes:
- engine_state/__init__.py: begin_generation() + commit_generation() +
_resolve_dir() for all load_* methods. Flat-layout legacy checkpoints
migrate into gen-0000 on first begin_generation call. GC retains K=2
committed generations.
- chat/runtime.py: checkpoint_engine_state uses the two-phase commit;
finalize_turn_trace_hash no longer writes discovery_candidates outside
the generation sequence (the second unguarded write path is closed).
- evals/l10_continuity/runner.py: _inject_orphan_tmp updated to inject
the two orphan shapes of the generation model.
- tests/test_adr_0219_generation_checkpoint.py: 18 tests, one per
acceptance-gate bullet + biting mutation variant each.
L10 lane: all_gates_pass=true; versor_condition<1e-6 throughout.
Smoke: 95 passed. Runtime: 20 passed.
The always-on PROCESS shipped (run_continuous + daemon + CLI), but the claim "it holds
over uptime" was UNFALSIFIED: the existing evals/l10_continuity soak drives the TURN loop
and is disjoint from run_continuous; the daemon path was covered only by ≤5-beat unit
tests. This lane converts "built" → "proven over horizon" for the idle path.
evals/l10_always_on/ mirrors the l10_continuity predicate/mutation harness but drives the
IDLE heartbeat: it seeds a real continuous life (held self + a cognitive turn to excite the
field), runs N beats with NO user turn, and gates four falsifiable predicates, each with a
*_holds (real soak) AND a *_bites (mutation) test per CLAUDE.md schema-as-proof:
- H1 closure — every observed idle beat versor_condition < 1e-6 (held over uptime, READ
never repaired); bites on a breached beat AND on a vacuous no-field soak.
- H2 bounded idle — a no-work beat adds NOTHING to the vault (no idle leak — invisible at
5 beats, fatal at 100k); bites on idle vault growth.
- H3 convergence — a saturated idle life SETTLES and stays settled (no re-awakening), at
rest at the end, closure intact on the tail; bites on never-settling + on re-awakening.
- H4 reboot resume — a reboot mid-soak resumes the SAME life (strict identity guard passes,
pre-reboot DERIVED learning survives, post-reboot closure holds); bites on failed
resume + on lost learning.
Measured (5000 idle beats, reboot@2500, ~20s): ALL gates pass — versor_condition flat at
1.389e-07 (no drift, no repair), vault bounded at 6 (no idle leak), converged at beat 1
with a 4999-beat at-rest tail, reboot resumed the same life with learning intact. This is
the empirical resolution of the L10 riskiest-unknown FOR THE IDLE PATH (the closure-by-
construction ruling covered the field-transition walk; this covers indefinite idle uptime).
Honestly NOT covered (recorded in the report, no silent skip): the continuously-LEARNING
life's resource cost under a sustained new-fact stream (O(n²) snapshot; per-run lived_life)
— out of scope until an afferent/intake feed + incremental persistence exist (the next
frontier).
Run on demand: PYTHONPATH=. python -m evals.l10_always_on [n_beats] [reboot_beat]. Path-
scoped out of CI smoke (like l10_continuity); runs in post-merge full-pytest. Additive — a
new eval lane, no existing files touched. 12/12 lane tests + architectural invariants green.
Completes the Wave M B3.5 consolidation slice (b–e), built on #728.
B3.5-b — calibration as a first-class evidence subject (`calibration_class`,
address `calibration:<class_name>`): RightInspector projection + Evidence
Chain Rail semantics (serving-discipline evidence, not runtime truth).
B3.5-c / B4a — nullable `LeewayEvidence` read model threaded through turn,
replay, cognition-proposal, and math-proposal surfaces, with a shared
absence-honest card. B4 is gated correctly: the tuple exists in typed data but
no producer populates it, so the card renders absence (verified: no non-null
producer in workbench/core/chat).
B3.5-d/e — UI-UX-GUIDE.md, b4-leeway-feasibility-gate.md, phase-a-residue-ledger.md.
Practice artifact — earn-it-for-real (runner-reproducible). The committed
`report.json` (additive earns PROPOSE @0.861, 95/5/50) is now emitted by a
deterministic runner rather than copied from the queue. `propose_runner`
gains `regenerate_practice_artifacts()`, which runs ONE sealed `resolve_pooled`
practice pass and writes BOTH report.json (the per-class ledger the calibration
reader consumes) and ratification_queue.json — two projections of one ledger,
coherent by construction and byte-reproducible. `runner.main()` delegates to
it (lazy import, no cycle), so both entry points produce the identical pair.
This closes the gap where a hand-copied report.json agreed with the queue but
no runner produced it. `resolve_pooled` is the aggressive sealed PROPOSE-regime
scorer (proposal-only/HITL, unsafe for serving, legitimate for
attempt-and-eliminate); wrong=5 is the sealed-practice learning signal, NOT the
serving wrong=0. No serving/derivation/reliability_gate source touched; the
practice lane is not in the serving-frozen SHA gate.
Validated:
- python -m pytest tests/test_workbench_{calibration,journal,replay,schemas}.py -> 31 passed
- python -m pytest tests/ -k "workbench or propose or learning_arena or practice"
-> 190 passed (3 failing tests in test_adr_0175_phase2_practice_lane.py are
PRE-EXISTING reds on clean origin/main: stale 4/0/46 assertions on build_report,
which this change does not touch)
- report.json + ratification_queue.json: deterministic (run1==run2) and
reproduced byte-identically by both `python -m ...runner` and `...propose_runner`
- pnpm build green; 144 UI tests across calibration/leeway/evidence/replay/
doctrine-gates/routes-docs-drift all pass
ADR-0184 S4b — the replay/provenance equivalence boundary, the
precondition for any semantic-primary path. After #685 the in-repo
legacy path delegates to the semantic boundary, so old-vs-new is
vacuous; the durable reference is a pinned canonical trace artifact
(evals/gsm8k_math/equivalence/v1/, 937 problems) whose authority is
the #684/#685 cross-tree differentials (documented provenance chain).
- state/provenance.py: replay_is_faithful — the replay bridge's law as
a checkable structural property (no verify/pool import; covered by
the #685 authority scan).
- equivalence/trace.py: canonical traces (worlds, candidates, order,
multiplicity, classifications, resolutions) + authority_violations
re-deriving the pool commit law from trace content.
- scripts/verify_semantic_equivalence.py: check / --update gate
mirroring verify_lane_shas.py.
- tests/test_adr_0184_s4b_replay_equivalence.py: full-corpus live-vs-
pinned equality, corpus-wide authority + faithfulness, and 17
single-mutation non-vacuity proofs.
Pure additions; verify.py/pool.py untouched; no serving, CLAIMS,
metrics, or lane pins moved.
Adds bound_slots_digest as a distinct VerificationProof obligation before any serving wiring. Keeps the VERIFIED lane off-serving while separating derivation, stated-slot binding, and back-substitution.
The first consumer of the P1-A VERIFIED contract. Given an R2 problem (prose) +
the INDEPENDENTLY hand-authored gold SETUP, it builds a VerificationProof and runs it
through the contract. NOT a serving-time VERIFIED producer: it depends on the eval-lane
gold setup (off-serving only), cannot move any serving metric, cannot emit served
[verified]. Serving-time independence (a non-gold second reader) is a separate, larger
rung, explicitly deferred.
evals/constraint_oracle/verified_producer.py (placed in evals/, not core/, to avoid a
backwards core->evals import: it needs the reader+solver (generate), the canonical
signature + gold (evals), and the contract (core) — clean direction evals->core+generate):
- verify_r2(text, gold_setup) -> R2VerificationOutcome.
- Two independent reads: primary = read_constraint_problem(text) (lineage
constraint_reader_v1); independent = the gold-authored setup (lineage
constraint_gold_author_v1). Convergence = equal constraint_setup_signature digests
(the SAME function for both reads). The gold ANSWER never enters — only the gold
STRUCTURE, through the signature.
- A wrong reader structure diverges from the gold-authored structure -> reads_disagree
-> NOT_VERIFIED (the poison). Reader refusal -> no proof + a limitation. Solver
boundary -> boundary_clear False -> NOT_VERIFIED. VERIFIED state/claim reached ONLY
via disclosure_for_verification.
On the real r2_gold lane (13 records): VERIFIED=7 (exactly the 7 'solved'),
NOT_VERIFIED=6 (3 solver_refuses + 3 reader_refuses) — wrong=0.
11 tests: integration over real gold (verifies >=1, never on divergence); all 7 solved
verify; the poison (divergent independent read -> reads_disagree); reader-refusal -> no
proof + limitation; solver-boundary -> not verified; distinct/named lineages; the
sanctioned route; all digests populated; gold answer structurally cannot enter; off-serving.
NOTE (bound_slots_digest): the P1-A VerificationProof (merged, unchanged per scope) has
no bound_slots_digest field; the "bound to stated quantities" obligation is carried by
derivation_digest (hashed over the canonical constraints + solution). Flag to extend.
Off-serving; smoke 90/0.
The mandatory lookback before CMB-d (3+ PRs on one surface; next-phase boundary). A 5-dimension
read-only fan-out audit over the 10-item checklist found the CMB-a/b/c substrate composes cleanly:
25 solid / 4 drift / 8 gap / 1 reader hazard. Doc:
docs/analysis/cmb-lookback-review-2026-06-08.md.
Fixes the one real hazard (H3): a query attributing the answer to a SINGLE agent ('how many words
does Alice type' with a distractor second rate) wrongly claimed the substantive
combine_mode_ambiguous — a hygiene over-claim that would corrupt the shared router in CMB-d. The
combined-query gate now excludes single-agent 'does <Agent>' attribution (_SINGLE_AGENT_QUERY) and
steps aside as not_combined_rate_shaped; the genuinely-combined forms ('do they' / 'are produced' /
'does it') still yield combine_mode_ambiguous. Pinned by gold cmb-16 + a reader test.
The review also resolves the CMB-d failure-family classification (input_shape:
not_combined_rate_shaped; must_remain_refused incl. rate_unit_mismatch — DECISIVELY, the reader has
no dimension representation so it cannot tell convertible from incompatible units; proposal_allowed:
combine_mode_ambiguous, missing_second_rate) and enumerates the live CMB-d registry preconditions
(not_combined_rate_shaped mapping, rate_unit_mismatch/non_integer_solution string collisions,
non_positive_net_rate family, Organ/ALL_REASONS extension). Provenance false-alarm (stale local main
checkout) noted + verified against origin/main.
gold 19/19 (6/5/8); reader 11/0/0 + 8 refused-correct; solver 6/0 + 5/0; 0 hygiene breaches; 123
tests; serving + R1/R2/R3 unchanged.
Third rung of the CMB ladder. Reads explicit combined-rate prose into a canonical
CombinedRateProblem or a closed reader refusal; graded by the CMB-a ruler + CMB-b solver via a new
run_reader lane. Off-serving; no router/contemplation/serving/ledger. Claim: 'English prose reads
into a canonical CMB setup or a closed reader refusal.'
Enforces the CMB-a 2x2 domain-entry grid + the deferral taxonomy (rate_unit_mismatch,
combine_mode_ambiguous, missing_second_rate, three_or_more_rates, reciprocal_work_rate_deferred,
clock_interval_deferred, not_combined_rate_shaped). Non-positive-net / non-integer stay the SOLVER's
refusals (the reader parses them). Router-organ hygiene: foreign R1/R2/R3 text steps aside as
not_combined_rate_shaped (the input_shape family), never a substantive boundary (0 breaches on all
36 R1/R2/R3 gold).
THREE adversarial verification rounds found 11 real wrong=0/hygiene hazards a curated-gold lane
cannot surface; all fixed + regression-tested. Root-cause fixes (not point-patches):
- cues are whole-word; difference = a CLEAN fill-vs-drain OPPOSITION classified per rate clause by
the verb in its own lead-in; rate_a=fill/rate_b=drain BY ROLE (a drain listed first still
subtracts); incidental drain nouns / drain-verbs-between-clauses no longer flip the mode;
- query duration + time-query quantity are read only from AFTER the 'how many' question, so neither
a premise number nor a transitional time marker is mistaken for them;
- the rate regex rejects decimals; substantive refusals are gated behind a real combination cue;
combine_mode_ambiguous requires a genuine COMBINED query (a compared pair steps aside);
- sequential segments (each rate carrying its own duration, tolerant of 'also for') step aside.
reader lane 11/0/0 + 7 refused-correct; gold 18/18; 19+17+3 adversarial triggers across 3 rounds;
126 tests (CMB + router/hygiene + R1/R2/R3 readers); serving + siblings unchanged.
Next (ladder): CMB-d router/contemplation wiring -> CMB-e ledger.
solve_combined_rate(CombinedRateProblem) -> int | Refusal over effective_rate = rate_a +/- rate_b.
effective_rate query returns the net (even <=0); quantity/time queries need a positive net rate
(non_positive_net_rate) and exact division (non_integer_solution). Pure integer arithmetic, no
float/Fraction (CMB v1 crosses no units). Off-serving. Graded by a new run_solver lane in
evals/combined_rate_oracle against the committed gold.
Adversarial 5-lens(+adjudicator) verification returned fix_first; both real hazards fixed:
- wrong=0 breach: negative known-slot inputs (time<0/quantity<0) bypassed the eff<=0 guard and
produced negative answers. FIXED UPSTREAM in model.py __post_init__ — rate_a/rate_b and the known
time/quantity are now positive ints, so the illegal state is unrepresentable and the solver can
never receive a negative-yielding path. Added model + gold (cmb-07d eff<0/time) + lane coverage.
- doc overclaim: the solver lane does NOT grade 'two independent paths' (both solver and the
oracle's _canonical_outcome delegate to model.effective_rate). Corrected both docstrings to name
the true anchor (committed gold + inline-computed literal tests) and added a difference-mode
inline-computed test.
R3-vac came back SOLID (separate PR). gold 18/18 (6/5/7); solver lane 6/0 + 5/0; 36 CMB tests;
router-hygiene + serving unchanged.
Closes the sibling vacuity CMB-a surfaced: R3's evals/rate_oracle solver_refuses branch
accepted any fixture labelled non_integer_solution without checking the arithmetic, and the
solved branch never recomputed the gold. A curator could certify an arithmetically-impossible
refusal or a wrong gold (with a self-consistent answer key) as valid — the 'obligation that
can't meaningfully fail = decoration' case CLAUDE.md forbids.
Adds _canonical_outcome (independent reimplementation of the single-rate algebra WITH R3.2
conversion — NOT a call to generate.rate_comprehension.solver, which would be circular). The
validator now cross-checks every solved gold (== computed) and every solver_refuses reason
(genuine) against it, with dedicated meaningful-fail tests anchored on hand-written literals.
Also removes the dead determinism self-comparison.
R3 gold stays 13/13 valid (7/2/4); reader 9/0/4 unchanged; 26 oracle/reader tests pass.
Off-serving; serving + other organs untouched.
First rung of the combined-rate ladder, off-serving (imports no generate.derivation /
core.reliability_gate). Claims exactly: the combined-rate setup ruler is defined and
gold-valid. No reader/solver/wiring yet — no capability claim.
- generate/combined_rate_comprehension/: local RateUnit (decoupled from R3), CombinedRateProblem
(two-explicit-rates + per-query-slot guard; effective_rate = rate_a +/- rate_b; non-positive net
left to the solver, not the model).
- evals/combined_rate_oracle/: span-free signature (sum commutative / difference ordered) + a
NON-VACUOUS validator. 17 gold fixtures: 6 solved (full combine_mode x query grid), 4
solver_refuses, 7 reader_refuses (the complete refusal taxonomy + the 2x2 domain-entry grid).
Adversarial 5-lens verification returned fix_first; the validator now cross-checks every solved
gold and solver_refuses reason against the canonical arithmetic (_canonical_outcome) so a
mislabelled / arithmetically-impossible fixture is rejected (meaningful-fail per CLAUDE.md), with
dedicated tests. Added eff<0 and eff=0/time coverage fixtures; removed a dead determinism guard.
gold 17/17 valid; 25 oracle tests; R1/R2/R3 + router-hygiene + serving all unchanged.
Doc: docs/analysis/cmb-a-combined-rate-ruler-2026-06-08.md
Option A (text-faithful): RateProblem gains time_unit (the duration's ORIGINAL unit; time stays int; defaults to the rate denominator). The SOLVER converts via convert_time -> exact Fraction -> int-or-refuse(non_integer); Fraction never leaves the solver, no floats. The reader accepts a convertible (minute<->hour) duration mismatch (keeps original time_unit); a non-convertible one still refuses rate_unit_mismatch. Signature includes time_unit ('30 minutes' != '30 hours').
Gold: r3-09 flips reader_refuses -> SOLVED (60 mph for 30 min = 30 mile, exact 1/2 hour; distractor 1800 rejected); new r3-13 non-convertible (3 gallons) stays rate_unit_mismatch. R3 gold 13/13; reader 9/0/0 -> answers 7/0/6 (setup_correct 8->9). Failure-family: rate_unit_mismatch now non-convertible-only -> still unsupported_rate_duration proposal surface; router-hygiene invariant stays green; R1 7/0/3, R2 10/0/0 unchanged. 119-test smoke green incl. invariants. Off-serving, no float path.
evals/constraint_oracle: span-free canonical setup signature (unknowns/facts/constraints/query; terms merged+sorted, lhs constant folded into rhs), a 13-fixture gold (7 solved / 3 solver_refuses / 3 reader_refuses) with a closed expect+refusal taxonomy, and a gold-validation runner. No reader yet — this proves the ruler is coherent (13/13 valid), not capability.
ADR-0211 ratifies the R2 finite-integer two-category constraint compiler: the IR, the canonical signature, the closed taxonomy, the four wrong=0 gates (setup oracle / reader refusal / exact solver / answer-choice contradiction flag), off-serving disjointness, and the C1-C9 ladder. Cross-refs ADR-0207/0175/0083/0055-0057; reuses the R1 setup_oracle pattern. Off-serving (no generate.derivation / reliability_gate). 13 oracle tests incl. per-branch meaningful-fail.
r1-07 now reads setup-correct and answers 6 — 'Nia has 9 more beads than Omar. Nia has 15. -> Omar = 6'. The reader binds the unknown base's unit FROM the relation when its subject is a known fact and its referent is the otherwise-ungrounded query target, so the equation is admissible; the answer oracle reverse-solves it (PR-7a). Bounded: single base == query target (no chains), known subject value, base not otherwise grounded, <=1 inverse (multiple_inverse_bases else), never over times/divide.
R1 setup 6/0/4 -> 7/0/3; R1 answers -> 7 correct / 0 wrong; 15-case 15/0/0; setup_wrong stays 0. Off-serving. Refreshes the R1 ledger to 7/0/3 (R1 closed; the 3 remaining refusals are wrong=0 boundaries).
Pin the EXACT reverse-solve semantics in the independent answer oracle before the
reader learns the inverse frame (PR-7b). A single more_than/fewer_than whose entity
is grounded and whose ref is the (otherwise ungrounded) query target inverts to the
base: base = entity -/+ delta (always integer-exact for +/-).
Refuses everything wider (each a meaningful-fail oracle test): more than one inverse
constraint, an inverse base that is not the query target (a chain), an already-grounded
base, a negative count result, or an inverse over times/divide. The duplicate-definition
guard moved into the per-kind branches so a grounded entity on a more/fewer_than is read
as an inverse constraint, not a redefinition; forward more/fewer/times/divide/sum paths
are unchanged (regression-guarded).
Contract only — the reader is NOT touched: r1-07 still refuses (admissibility_refused),
so R1 stays 6 / 0 / 4 and 15-case 15 / 0 / 0. r1-07 gold drops the non-consumed ask_base
(aligning with the reader's future projection) and gains gold=6, dormant until PR-7b.
Off-serving (only evals/tests import the oracle). SHA gate 8/9 (public_demo env flake).
Widen the aggregate-query recognizer so a multi-part total may be asked with a
trailing qualifier after 'have': 'How many X do A and B have altogether?' and
'... in total?'. The qualifier is stripped and honored ONLY for the multi-part
(sumquery) form — a single-entity query carrying it refuses, guarded by
'not aggregate'.
Phrasing-only: no new arithmetic, no new relation kind, no inverse solving, no
distractor/pronoun handling. The parts still flow through the existing sum_of;
an ungrounded part (unit unbound) or a unit-incompatible part (unit mismatch) is
refused downstream by the REAL admissibility check, so the recognizer cannot
over-read. Off-serving organ only (no generate.derivation / reliability_gate).
Flips r1-03 (more+altogether -> 25) and r1-04 (fewer+in total -> 34):
R1 setup 6 / 0 / 4 R1 answers 6 / 0 / 4 (setup_wrong 0, gold_error 0)
15-case 15 / 0 / 0 29 quant tests, 102 affected-file tests green
Tests are meaningful-fail: the single-entity-qualifier, ungrounded-part, and
unit-incompatible-part refusals each fail loudly if their guard is removed.
PR-6d adds the partition frame: combine all parts into a total, then split that
total equally into N containers. r1-06-subtotal-reused moves refused → correct —
the FIRST case where the divisor applies to a DERIVED symbol (the total), not a
directly given fact. That is real progress toward GSM8K setup comprehension,
where intermediate quantities are the norm.
Scope (kept narrow on purpose):
No new relation kind.
No new arithmetic operation.
No rational support.
No rounding/flooring.
No serving path touched.
The frame reuses the already-ratified pieces — SumOf(parts) + Div(Symbol(total),
Literal(N)) → divide_by — so this PR is reader-only (no IR / admissibility /
oracle / signature change).
Frame grammar:
"They combine their <unit> and split them equally into N <containers>."
+ "How many <unit> are in each <container>?"
-> total = sum(all facts); per_<container> = total / N; ask per_<container>.
wrong=0 boundaries:
- Exact-divisibility still gates the ANSWER, now over a derived total: 5+6=11,
11/3 is non-exact -> the setup reads correctly but the answer REFUSES (never
floors). Setup comprehension and answer exactness are cleanly separated.
- Partition/query coherence: a partition is read ONLY together with its
"in each <container>" query (and vice versa); container mismatch (box vs jar)
refuses. Prevents over-reading a story detail into an unused derived value.
Meaningful-fail verified: disabling the guard makes a dangling partition
wrongly comprehend.
Gates:
R1 setup: 4 correct / 0 wrong / 6 refused
R1 answers: 4 correct / 0 wrong / 6 refused / setup_wrong 0 / gold_error 0
15-case setup: 15 / 0 / 0
97 PR-6d tests + 99 relational/invariant tests green. Reader is off-serving
(no generate.derivation / core.reliability_gate import).
PR-6c adds the divisive comparative frame: "half as many" read as EXACT INTEGER
DIVISION. It is the divisor twin of PR-5c's multiplicative frame, and moves the
independent R1 gold's r1-02-half from refused → correct.
No serving path touched. No rational/fractional answer support added. Non-exact
division refuses.
Design (ADR-0134 amended — divide made symmetric with multiply):
- `_check_divide` now admits a SINGLE-DEP divide-by-dimensionless-literal
(item / dimensionless = item), the exact twin of single-dep multiply. The
2-dep rate-divide path is untouched. This keeps the IR's "literal operands
are not deps" invariant (proven in PR-6a) uniform across Mul AND Div, so the
reader builds both without a per-op special case and WITHOUT synthesizing a
divisor symbol that would pollute the setup-oracle's unit signature.
- `Div(Symbol, Literal)` IR node: "ref / divisor", operation_kind "divide",
projects to `divide_by`. Divisor-only contract mirrors the scalar-only one.
- Reader: `_DIVISOR_WORDS={half:2}` slots into the same 8-token "<WORD> as many"
template as the factor words; graph carries only the two entities.
- Gold reconciliation: r1-02 placeholder `times_as_many factor 0.5` → exact
`divide_by divisor 2` (gold 4). Makes the INDEPENDENT gold integer-faithful.
The wrong=0 boundary — exact divisibility:
the oracle admits `divide_by` only when `base % divisor == 0`. An odd base
halved REFUSES (gold_error), never floors to a wrong integer. Divisor must be
a nonzero int (0, 0.5, 1.5, bool all refuse); divisor=1 is intentionally the
identity (pinned). admissibility proves DIMENSION; the oracle proves EXACT VALUE.
Meaningful-fail (CLAUDE.md Schema-Defined Proof Obligations), both verified red:
- drop the `% divisor` guard → test_oracle_refuses_non_exact_division fails (returns 3).
- disable the single-dep divide branch → the admissibility test AND the reader's
`half` test fail (admissibility refuses → reader refuses → half stays refused).
Gates:
R1 setup: 3 correct / 0 wrong / 7 refused
R1 answers: 3 correct / 0 wrong / 7 refused / setup_wrong 0 / gold_error 0
15-case setup: 15 / 0 / 0
91 PR-6c tests + 60 relational lanes + 56 architectural invariants + 502
binding-graph/proof-chain/adapter tests green. All 8 SHA-content lanes match
(serving unmoved; admissibility has no generate.derivation/reliability_gate consumer).
The first capability slice on the R1 arc, gated by the setup-oracle: turn the
"twice / N times as many" reading from REFUSED into a correct setup, without a single
misread. Builds on the typed IR (PR-4) and the R1 gold (PR-5b).
- IR: a Mul(symbol, literal-factor) node — to_canonical_string "ref * factor",
operation_kind "multiply", dependencies {ref}, to_relation -> times_as_many. The
product keeps the symbol's unit (count * scalar = count), admitted by the REAL
check_admissibility multiply path (the literal factor is dimensionless, not a dep).
- Reader: a multiplicative template "Y has <factor> as many <unit> as X" (factor word:
twice/double/triple/quadruple) and "Y has <N> times as many <unit> as X", checked
BEFORE the digit gate (the factor may be a word). 'half' (a /2) is deliberately
deferred — divide-by-literal is a separate admissibility path.
- setup-oracle: relation_signature now canonicalizes times_as_many.
Setup-oracle R1 result: 2 setup_correct (r1-01 twice; r1-05 the multi-step chain
ivy/jon=3*ivy/kim=jon+2), 0 setup_WRONG, 8 setup_refused. Every hard negative stays a
safe refusal: missing-base (Rosa ungrounded), ambiguous referent, distractor, inverse,
partition, 'altogether'/'in total' phrasings, and 'half' (divide). wrong=0 held through
the first capability addition.
Gates green: setup-oracle R1 setup_wrong=0; 15-case setup gate 15/15 setup_wrong=0;
relational_metric answer lane 15/15 wrong=0; binding-graph admissibility + realize +
architectural invariants + chat-runtime + pipeline (122+). No serving path touched (this
reader feeds the relational_metric / setup-oracle lanes, not the candidate-graph serving).
Pure-gold. Defines what "correct" means for the R1 (comparative / derived-symbol /
multi-step) shapes the reader cannot read yet, and verifies the current reader REFUSES
them rather than misreading them. NO frame extraction, no parser change, no serving.
- evals/setup_oracle/r1_gold.jsonl — 10 INDEPENDENT, self-contained fixtures (id, text,
relations, expected_units, query, notes) covering: twice/half-as-many (multiplicative),
more/fewer-than -> total, a multi-step derived chain (jon=3*ivy; kim=jon+2), a derived
subtotal reused downstream (total=lee+mae; per_box=total/3), an inverse target, and the
hard negatives (ambiguous referent, missing base quantity, distractor quantity). The
gold is the authority — each fixture is reviewable without joining files.
- run_r1() scores the CURRENT reader against this gold.
Investigative result (the deliverable): **10/10 setup_refused, setup_wrong=0, setup_correct=0.**
The reader refuses every R1 shape with a TYPED reason (non_digit_quantity /
no_quantity_template / unreadable_quantity_clause / admissibility_refused) — it NEVER
misreads an R1 case as a simpler form. So the wrong=0 posture holds all the way down, and
PR-5c (the R1 frame) starts from a clean foundation: every R1 case is currently a safe
refusal, and the frame's job is refusal -> correct, with the oracle ensuring no
setup_wrong is ever introduced.
`python -m evals.setup_oracle r1` prints the auditable R1 report (exit 0 iff
setup_wrong==0). The 15-case setup gate + relational_metric answer lane are untouched.
Strengthen the ruler before it judges real GSM8K frames — NOT a capability change, no
new parser behavior, no serving path. The setup-oracle now grades UNITS, read from the
binding-graph itself, against an independent expected-unit gold.
- New per-symbol unit axis (symbol_unit_signature) — read from the GRAPH's symbols
(reader_symbol_units), not the answer projection; covers fact + relation-operand units
(operands are symbols). A None unit canonicalizes to "unset" so an unmodelled symbol
never silently matches a declared gold unit.
- The question-target signature gains the target's expected_unit (reader from
BoundUnknown.expected_unit; gold from the fixture).
- expected_units.json — independent, hand-authored MODELING unit per entity for the 15
cases (a discrete sortal count -> the generic count unit "item"; money -> "dollars").
Authored from the PROBLEM, not copied from the reader.
- The reader's per-symbol units match the independent gold on all 15 (setup_wrong=0) —
agreement, not circularity (both apply the same correct count->item / money modelling).
- Meaningful-fail: a reading whose STRUCTURE matches but whose UNIT diverges now FAILS
(test_unit_mismatch_is_caught, test_target_unit_mismatch_is_caught). The ruler is not
decoration on the unit axis either.
Gate held: setup-oracle 15/15 setup_wrong=0 (now structure + units + target). The
relational_metric answer lane is untouched (oracle-only change). No generate/ change.
This is the unit-aware ruler PR-5b (independent R1 gold) and PR-5c (the first R1
comparative/derived-symbol frame) must pass before serving.
Two coupled, additive, off-serving changes toward the typed math-comprehension organ.
No serving path touched; the relational_metric answer lane stays 15/15 wrong=0.
PR-1 — QuantQuery → BoundUnknown. comprehend_quantitative now emits the question
target as a BoundUnknown INSIDE the binding-graph (symbol_id, state_index="terminal",
question_form "count"|"total", expected_unit), so the graph is a real question-bearing
mathematical object and its canonical serialization carries the target. The external
QuantQuery is RETAINED, consistent-by-construction, so the two consumers
(to_relational_metric, realize/quantitative) are byte-identical; a follow-up rewires
them onto graph.unknowns and drops the duplicate field.
Setup-oracle lane (evals/setup_oracle) — grade the READING, not the answer. The
relational_metric lane scores answers, which can bless a semantically-wrong derivation
that coincidentally lands on the right number (the exact hazard the held-out
measurements + the 2/87 resolve_pooled probe exposed). The setup-oracle compares the
reader's comprehended STRUCTURE — a span-free signature of facts + typed equations +
the BoundUnknown target — against the INDEPENDENT gold structure (the relational_metric
cases' own relations/query, authored separately from the binding-graph reader). A
structural mismatch is setup_wrong, the wrong=0-critical count, even when the answer
would be right. v1 grades structure (units deferred — covered by admissibility). The
reader reads all 15 cases with the gold structure (setup_wrong=0); a meaningful-fail
test proves the oracle catches a right-answer/wrong-structure reading (it is not
decoration). `python -m evals.setup_oracle` exits nonzero iff setup_wrong > 0.
This is the measurement rig BEFORE investing in frame families: setup_wrong=0 is the
gate; serving must not move while setup_wrong > 0. It is the first milestone of the
math-comprehension organ, not a path to "solve GSM8K".
Verified: setup-oracle 15/15 setup_correct wrong=0; quantitative + setup-oracle unit
tests (17); realize-binding-graph + binding-graph + architectural invariants (183).
The final AGI-spine step (A INSTRUMENT → B WIRE → C DEEPEN → D CLOSE → E ESTIMATION).
The engine may now SERVE a DISCLOSED estimate for a query it would otherwise refuse —
but only for a predicate-class that has measured itself reliable, and never as fact.
This executes the ADR-0206 §5 cognition-path widening: the bridge's LICENSE node
(reliability_gate.license_for), previously "built — not yet called from serving", is now
called. govern_response returns APPROXIMATE iff a genuine licensed Action.SERVE
LicenseDecision is passed (STRICT for every other input — so every existing serving call
site is byte-identical); shape_surface DISCLOSES the estimate as "[approximate] …".
Mechanism:
- generate/determine/estimate.py — a BLIND converse-guesser: told p(a,b), asked p(b,a),
it commits the converse. It never reads the pack's symmetry metadata; whether the guess
is right is MEASURED, not assumed.
- evals/determination_estimation/ — the gold lane: run_practice (sealed, ADR-0199) folds
the converse-guesser over symmetric (sibling_of) vs directed (parent_of) cases, scored
against the pack's graph.edge.symmetric truth (gold independent of the solver). The gate
DISCRIMINATES: sibling_of earns SERVE (660 correct → Wilson floor 0.990046 ≥ θ_SERVE),
parent_of does not (660 wrong → 0.0). The license is earned by VOLUME — 657 perfect
commits is the exact θ_SERVE=0.99 threshold (656 is below).
- generate/determine/data/estimation_ledger.json — the ratified committed ledger,
hash-verified on load (a hand-edited ledger raises RatifiedLedgerError); it IS the
deterministic sealed-practice output (a GSM8K-style --check test pins this).
- chat/runtime.py — when a converse query is refused and the class holds a SERVE license,
the disclosed estimate is surfaced through the bridge (gated by config.estimation_enabled,
default OFF; only meaningful with accrue_realized_knowledge).
Invariants:
- wrong=0 by construction — an estimate is ALWAYS disclosed ([approximate]), never a silent
commit (UNVERIFIED_POSSIBLE is never in APPROXIMATE's admissible set), and only a genuine
ratified license widens (a forged {"licensed":True} dict / a PROPOSE license / an
unlicensed SERVE all stay STRICT). Defense-in-depth: type-gate ∧ admissible-set ∧
hardcoded disclosed state.
- never self-authored — ceilings stay at safe defaults (θ_SERVE=0.99); the engine cannot
raise its own bar. The ledger is sealed practice, hash-verified.
- session/serving only — no corpus/pack/identity/proposal/vault mutation; the HITL teaching
path is untouched. Deterministic; no clock/random.
- byte-identical for every non-E turn (the 2643 govern_response call passes no license).
Out of scope (separate ADR-0206 §5 PRs): the math-serving seam (select_self_verified,
touches the sealed metric), SITUATE (stakes), and the live FEED-BACK loop.
Verified green: smoke (90), architectural invariants (56), response_governance (321,
incl. the new license-gated widening test), the determination-estimation lane (12), and
the B/D/determine regression net. Four-lens adversarial review (disclosure/wrong=0,
calibration integrity, byte-identity, boundary/determinism): all held. Design:
docs/analysis/E-estimation-design-2026-06-06.md.
