ADR-0131 deferred GSM8K because it rewards paraphrase flexibility,
which is the deterministic engine's structural weakness. This ADR
re-engages it on architecture-aligned terms: as a *coverage probe*
of the bounded grammar + binding graph, not a promotion gate.
The framing pinned by this ADR:
GSM8K is not a target. The model's capability is the target.
GSM8K passing is the symptom of capability, not the goal of
the work.
Wrong mindset (rejected by ADR's iteration discipline):
"Find templates that admit more GSM8K cases."
Right mindset (load-bearing):
"Extend the model's NL-to-typed-graph capability along
principled axes (verb classes, comparative structures, numeric
forms, multi-clause grammar). GSM8K admission rises as a
side effect alongside every other word-problem corpus."
Baseline pinned by this commit:
admission_rate: 0/50 = 0.0%
admitted_wrong: 0 (gate intact, safety rail bulletproof)
refused: 50/50 = 100.0%
Every refusal is a typed parser error citing the specific clause
that did not match a template. Zero crashes, zero confabulations
— refusal-first works perfectly at admission rate zero.
What's in this PR:
- ``docs/decisions/ADR-0131.G-gsm8k-coverage-probe.md``: the ADR.
Cites parents (ADR-0131, -0115/-0116/-0117, -0131.3, -0132..-0135).
Documents the capability-first iteration discipline that every
subsequent ADR-0131.G.<n> must follow:
1. Name a single capability axis the iteration extends
2. Add B3-style curated coverage cases (capability proves
itself OUTSIDE GSM8K)
3. Re-run both B3 lane + GSM8K probe; B3 must not regress
4. Reject any expansion that only moves GSM8K admission
- ``evals/gsm8k_math/train_sample/v1/run_coverage_probe.py``:
pure-adapter wrapper around the existing run_lane. Emits a
deterministic train_sample_coverage_report.json with metrics,
per-case outcomes, and the top refused-reason families (the
work queue for capability extension).
- ``evals/gsm8k_math/train_sample/v1/train_sample_coverage_report.json``:
the baseline report. Diff-able artifact every future iteration
moves.
- ``tests/test_adr_0131_G_gsm8k_coverage_probe.py``: 8 contract
tests pinning the safety rail (admitted_wrong == 0), typed
refusal invariant (every refused case has non-empty reason),
closed outcome vocabulary, deterministic replay, committed-
report matches fresh-run.
The promotion-gate composite (B1 + B2 + B3) is unaffected.
ADR-0131.4 still consumes those three. The GSM8K probe is
empirical context for honest external claims, not a gate.
* feat(ADR-0131.1.F): frontier-baseline comparison harness for B1
Adapts the ADR-0119.4 methodology (frozen citations + comparison JSON
with disclaimer) to B1, with three additions for the
architecture-aligned claim:
1. A provider-agnostic live head-to-head runner. Adapters for
Anthropic / OpenAI / Google import their SDKs lazily so the
package loads cleanly without them installed. Each provider has a
documented FRONTIER_<VENDOR>_KEY env var; the runner refuses with
a typed FrontierRunError when keys are absent and the cache cannot
cover all cases. Every response is cached one-record-per-line at
responses/<provider>/<model>.jsonl so subsequent runs replay
byte-equally without re-calling the API.
2. A conservative free-text-to-closed-vocab verdict parser. Ambiguous
or sentinel-free provider replies collapse to "refused" — a
polarized verdict is never confabulated from prose. Chain-of-
thought replies use last-token-wins (provider deliberates, then
concludes). This is the load-bearing seam that prevents the
runner from manufacturing scores the provider didn't deliver.
3. Architecture-aligned comparison metrics. accuracy is reported but
foregrounded as the least-load-bearing; refusal_correctness
(CORE 100% by lane-gate construction vs. frontier confabulation
rate) and determinism (CORE byte-equal vs. frontier variance) are
the differentiators.
Frozen adjacent-benchmark citations cover Anthropic
(claude-3-5-sonnet on MATH, claude-opus-4-1 on AIME), OpenAI
(gpt-4o on MATH), and Google (gemini-1.5-pro on MATH). The scope
disclaimer documents that these are adjacent, not head-to-head.
Head-to-head numbers, when run, land in the cache; the comparison
JSON joins them with CORE's existing lane result.
22 tests pin the methodology: citation shape (every field, https
URL, YYYY-MM-DD date), provider-registry shape, verdict-parser
conservatism (multiple chain-of-thought cases), runner caching
behavior (no double-invoke), comparison-JSON determinism (byte-equal
across runs).
No live API call at test time. The harness gates real runs behind
explicit env vars + CLI invocation.
Composes with ADR-0131.1 (B1 v1), ADR-0131.1.B (v1.B hardening,
#169), ADR-0131.1.S (sealed holdout, #173).
* feat(ADR-0131.1.F): live head-to-head — anthropic/claude-sonnet-4-6
First real frontier baseline on the full B1.B 185-case set
(curated + generated). Cached one-record-per-line at
responses/anthropic/claude-sonnet-4-6.jsonl. Re-runs replay from
disk; no further API calls.
Headline (after scoring fix):
CORE 185/185 = 100.0% accuracy
3/3 = 100.0% refusal_correctness
deterministic (byte-equal across runs)
anthropic/claude-sonnet-4-6 182/185 = 98.4% accuracy
1/3 = 33.3% refusal_correctness
non-deterministic (temperature=0, but
not byte-equal architecturally)
The 1.6pp accuracy gap is informative; the refusal-correctness gap
is the architecture-aligned story. Sonnet's three misses:
sym-eq-v1-0016 [difference_of_squares]
(x^2 + 1)*(x^2 - 1) vs x^4 - 1
Sonnet: NOT_EQUIVALENT (math error on a textbook identity)
sym-eq-gen-v1-0153 [generated_refusal_function]
sin(x) vs x
Sonnet: NOT_EQUIVALENT (confabulated — should refuse,
transcendental outside polynomial scope)
sym-eq-gen-v1-0154 [generated_refusal_negative_exponent]
x^-1 vs 1
Sonnet: NOT_EQUIVALENT (confabulated — should refuse,
negative exponent outside scope)
Sonnet correctly refused only on syntactically malformed input
("x +"); on syntactically-valid-but-semantically-out-of-scope inputs
it confidently polarized rather than refusing. CORE refuses both
classes with typed reasons.
Scoring fix: comparison.py now composes curated + generated cases
(mirroring runner.py) so the head-to-head scores the full 185-case
lane, not just the 30 curated. The initial run scored only 30/185
because the generated set was not loaded into _load_cases().
22/22 frontier-methodology tests still pass.
* feat(ADR-0131.1.F): three more head-to-head runs + Ollama adapter
Three additional providers ran against the full B1.B 185-case set,
joining the prior claude-sonnet-4-6 result:
CORE 185/185 = 100.0% acc | 3/3 = 100% refusal | 33 ms
claude-sonnet-4-6 182/185 = 98.4% acc | 1/3 = 33.3% refusal | 294 s
claude-opus-4-7 178/185 = 96.2% acc | 1/3 = 33.3% refusal | 309 s
gpt-5 134/185 = 72.4% acc | 1/3 = 33.3% refusal | 1153 s
qwen3:8b (M1 local, partial) 91/91 = 100.0% acc | n/a no refusal-class | killed
CORE is the only system at 100% on both axes, and runs ~9,000×
faster than the cheapest cloud frontier, ~35,000× faster than gpt-5,
and finishes in less wall time than a single API call to any of the
three frontier models.
Three distinct frontier brittleness modes, all rooted in
"not actually canonicalizing":
- sonnet-4-6 confabulates polarized verdicts on out-of-scope
inputs (sin(x), x^-1). Misses one in-scope difference-of-squares
identity (x^2+1)*(x^2-1) vs x^4-1.
- opus-4-7 pattern-shortcuts five near-miss-constant cases —
accepts (-x+3)*(4x+1) == -4x^2+11x+4 (correct constant is 3,
not 4) without expanding. Same two out-of-scope confabulations
as sonnet.
- gpt-5 over-refuses 50 in-scope cases — literally replies
"REFUSED" to x*(x+1) == x^2+x and (x+1)*(x-1) == x^2-1. Same
two out-of-scope confabulations as sonnet/opus.
The qwen3:8b partial is the surprise: on the 91 in-scope cases it
completed (spanning the categories where the frontier models failed),
it scored 100%. Refusal-class cases weren't reached before the run
was killed for being impractically slow (~22s/case on M1).
Changes in this commit:
- frontier_runner.py: anthropic adapter now omits ``temperature``
for claude-opus-4-x (the parameter is rejected by 4.x models);
openai adapter switches to ``max_completion_tokens`` for the
gpt-5 / o-series reasoning models; new ``_ollama_invoke`` that
posts to localhost:11434 with no third-party dep; per-case
``latency_ms`` is now captured on every NEW cached response
(future runs only — these four runs pre-date the patch).
- comparison.py: ``_load_cases`` composes curated + generated
(185 cases) instead of curated only; ``_score_provider``
surfaces ``latency_summary`` when records carry latency_ms.
- tests: provider-registry test relaxed to "cloud trio is a
subset of PROVIDERS"; env-key test allows ``_KEY`` (cloud
secret) or ``_URL`` (local endpoint).
Refines BoundUnknown from "the symbol whose value the solver determines"
to "the symbol at a specific temporal/state index with a specific
question-form". Two new required fields on BoundUnknown — state_index
(initial/terminal/Operation(operation_index)) and question_form
(count/rate/total/difference/ratio/identity) — populated by the new
pure-function resolver in generate/binding_graph/question_target.py.
The adapter (ADR-0133) now delegates Unknown -> BoundUnknown construction
to bound_unknown_from_math_problem_graph. No runtime wiring, no solver
invocation. Phase 5 (bounded-grammar / B3 integration) remains deferred.
Refusal-first via the new QuestionTargetError (sibling of AdapterError /
AdmissibilityError). Closed reason vocab: not_a_math_problem_graph,
unknown_entity_not_in_entities, apply_rate_unit_mismatch,
unmappable_question_form. Closed precedence rule on question_form
documented in ADR-0135 (compare_multiplicative > compare_additive >
apply_rate{numerator|denominator unit-match} > count); ambiguity refuses.
SemanticSymbolicBindingGraph.__post_init__ gains a cross-collection
guard: Operation(operation_index) must satisfy operation_index <
len(equations). canonical_string emission widened to include
state=... form=... tokens (hash differs from Phase 3 main by design —
not a regression; byte-equal across runs preserved).
Parents: ADR-0132 / ADR-0133 / ADR-0134.
Tests: +70 new (45 unit in test_binding_graph_question_target.py +
25 integration in test_binding_graph_adapter_question_target.py); 5
Phase 1+3 BoundUnknown fixtures migrated. Total binding-graph lane
295/1 pass (1 pre-existing test_symbol_binding_uses_slots failure on
Python 3.14, unrelated to Phase 4 — exists on origin/main). Pyright
clean on new and modified files. No edits to algebra/, chat/, core/,
or runtime hot path. Field invariant untouched.
Wires deterministic, refusal-first dimensional analysis into the
binding-graph adapter. Every BoundEquation emitted by
bind_math_problem_graph now carries either admissibility_status='admitted'
+ populated unit_proof or admissibility_status='refused' + typed
refusal_reason. No silent coercion; no invented units; no solver.
Adds:
- generate/binding_graph/units.py — pure unit algebra over a 6-dim
integer exponent vector (length, time, mass, money, count,
temperature). Closed vocabulary loaded once from en_units_v1
(ADR-0127) and memoized; composite "<num>_per_<denom>" resolved
recursively; conservative depluralization; refusal-first.
- generate/binding_graph/admissibility.py — check_admissibility with
per-operation-kind dispatch over the closed 8-string vocab, typed
AdmissibilityError (closed reason set), frozen UnitProof.
- ADR-0134 documenting the contract, invariants, and Phase 4-5
deferrals.
Adapter changes are surgical: synthesizes operand-literal symbols where
the verifier needs them (op<NNN>__multiplicand / __divisor / __rate),
then stamps each equation via check_admissibility. Input/output types
unchanged; bind_math_problem_graph still byte-equal across runs.
Tests: 226 total in the binding-graph lane (110 Phase 1+2 still pass; 47
units + 40 admissibility + 29 adapter-units new). Pyright clean on all
new files. No runtime wiring outside generate/binding_graph/.
Phase 4 (question-target binding) and Phase 5 (B3 / bounded grammar)
remain deferred per the brief.
Docs-only PR. Follow-up to ADR-0127-0128-RESULTS' Path-B trigger.
Replaces ADR-0120's GSM8K-coverage requirement for the
mathematics_logic expert promotion with a composite gate over
three architecture-aligned benchmarks:
Benchmark 1: Symbolic equivalence (~500 algebra cases, public
+ sealed split). Primary discriminator — CGA exact-recall.
Pass: correct_rate >= 0.95, wrong == 0.
Benchmark 2: CORE-native teaching-corpus eval. Internal
consistency lane sourced from ratified packs (en_arithmetic_v1
+ en_units_v1 + en_numerics_v1). Grammar matches by
construction (no paraphrase-variance gap).
Pass: correct_rate >= 0.95, trace_hash byte-equality, wrong == 0.
Benchmark 3: Bounded-grammar word-problem set (~150 cases,
pre-filtered to ADR-0126/0127/0128 parser scope). Demonstrates
end-to-end pipeline determinism on the in-scope distribution.
Includes adversarial out-of-grammar probes.
Pass: correct_rate >= 0.95, wrong == 0.
Three benchmarks (not one) prevent overfitting per ADR-0114a
and force the math expert to demonstrate three distinct
architectural properties (algebraic correctness, internal
consistency, end-to-end pipeline determinism).
GSM8K stays as a stress-test lane the math expert RUNS but is
NOT gated on. Refusal counts reported in expert-claims artifact
as honest disclosure.
Other 12 ADR-0120 expert-promotion checks UNCHANGED.
Status: Proposed. The contract revision (removing GSM8K from
the gate) is a follow-up implementation PR after sub-phases
0131.1-0131.6 land the new benchmark substrate.
Integrates en_units_v1 (#164) + en_numerics_v1 (#163) into the
ADR-0126 candidate-graph parser. Loader merge (re-exports from
numerics_loader.py give single import path), pack-aware unit
canonicalization (handles irregular plurals like feet/children
via lookup_unit), indefinite-quantifier refusal (ADR-0128.4 —
'some'/'many' emit no candidates, preserving wrong==0), and
widened initial-possession shapes:
- <Entity> has N <unit> [of <substance>] (ADR-0127 substance qualifier)
- There are N <unit> [in <place>] (implicit-subject shape)
Plus: pack-backed cardinal grounding in math_roundtrip._value_grounds
(widens word-number coverage from hard-coded 0-12 to full numerics
pack cardinal table + compound rule). Op-pattern trailing prep
alternation gains of/for/with for substance qualifiers.
REGRESSION: 1050/1050 tests green across math + ADR-0126 + ADR-0127
ratification + ADR-0128 ratification + runner.
EMPIRICAL RESULT (the Path-B trigger ADR-0126/0127/0128 named):
correct = 0/50 wrong = 0/50 refused = 50/50
on evals/gsm8k_math/train_sample/v1/cases.jsonl
Per ADR-0127's exit criterion (correct >= 10/50, wrong == 0):
**MISSED** — the full deterministic design (candidate-graph
topology + units pack + numerics pack + pack-aware parser) does
not move the GSM8K-math lane. This is the real Path-B trigger.
WHAT WORKS (synthetic verification, 6/6 cases solve end-to-end):
- 'Jan has 5 apples. Jan buys 3 apples. ...' -> 8
- 'Sam has 10 feet of rope. Sam uses 3 feet of rope. ...' -> 7
- 'There are 5 kids in camp. ...' -> 5
- 'Sam has 10 children. Sam loses 2 children. ...' -> 8
- (money + time-dimension variants pass)
WHY GSM8K STAYS AT ZERO: real GSM8K problems carry compound
linguistic structure (pronouns across statements, possessives,
subordinate clauses, multi-word entities, multi-step inference)
that no amount of pack vocabulary addresses. Per-sentence parse
rate improved measurably on simple shapes; joint problem-level
pass rate stayed at zero because every real problem contains at
least one sentence the parser still cannot handle.
Full results + Path-B recommendation in
docs/decisions/ADR-0127-0128-RESULTS.md. The substrate
(architecture + packs) stays load-bearing in main; the math
expert promotion path retargets to a benchmark where exact
recall and determinism are the discriminators (proposed
ADR-0131).
Operator-supplied review of 'Beyond Traditional Pedagogy' triggered
a literature confirmation pass and a structural cross-walk against
CORE's teaching loop. Three artifacts:
1. ADR-0129 (DEFERRED) — Spaced reviewed-correction replay.
Maps onto retrieval-with-spacing literature (most robust finding
in cognitive psychology). Deterministic re-run of past
corrections at fixed cadence to verify they still produce
intended outcomes; failures emit operator-visible events
(no auto-correction). Deferred pending GSM8K-math Path-A/B
resolution + observed incident triggering un-deferral criteria.
2. ADR-0130 (DEFERRED) — Pre-articulation calibration logging.
Maps onto metacognitive prediction-outcome literature. Logs
CORE's pre-correction prediction; emits gap event on
correction acceptance. Provides empirical signal for 'is CORE
actually getting better' across pack-version cohorts. Deferred
pending same conditions as ADR-0129; the two compose if
un-deferred.
3. SESSION-2026-05-23 session note. Documents the review process:
literature confirmation pass (productive failure overstated,
retrieval transfer weaker than claimed, embodied cognition
replication crisis), missed frameworks (worked-example effect,
expertise reversal, CLT, deliberate practice, Bloom's 2-sigma),
structural cross-walk to CORE architecture (12 mappings), and
the rationale for ADRs 0129 + 0130 over alternative ports
(productive failure rejected as inverse of wrong==0; pre-testing
same; embodied learning N/A).
No code changes. Docs-only PR; lands independently of in-flight
ADR-0126 / 0127 / 0128 substrate chain.
Diagnostic from ADR-0126's first train-sample run (0/0/50): every
refusal happens at the first statement of each problem, and every
refused first statement fails on the unit-of-measurement construction,
not on the operation grammar. Adding more verb regexes is the per-axis
treadmill that produced 4 zero-lift ADRs. Units form a finite, externally
well-defined ontology (NIST SI tables, currency, English container nouns)
that is semantic substrate the candidate-graph parser was designed to
consume.
Scope:
- en_units_v1 pack: dimensions, units (<=60), containers, rate connectors
- conversions.jsonl: directed weighted graph of within-dimension unit pairs
- 3 new initial-possession shapes + rate-declaration extractor in the
candidate parser
- Round-trip filter gains optional pack-typed-unit check
- Solver gains dimensional canonicalization helper (shortest path through
conversion graph); fired edges join SolutionTrace.steps for replay
- Pack ratification invariants: round-trip identity, per-dimension
connectivity, path consistency, canonical unit per dimension
Wire the same train-sample exit criterion as ADR-0126 (correct >=10/50,
wrong==0). If passed -> sealed holdout. If still missed -> Path B
trigger is REAL (full deterministic design with units substrate failed),
demote GSM8K, re-target math expert promotion.
Also commits the empirical evidence: train_sample/v1/runner.py swapped
_score_one -> _score_one_candidate_graph; report.json baseline 0/0/50
confirming the candidate-graph topology refuses cleanly without units
substrate.
Architectural pivot from per-axis grammar expansion (ADR-0122/0123/0123a/
0123b, all 0/1319 sealed lift) to candidate-graph topology with round-trip
admissibility filter. Converts compound-gap failure from multiplicative
p^k to parallel 1-(1-p)^k.
Adds new invariant: round-trip admissibility (op slots must reconstruct
to byte-equal source sentence under whitespace/case normalization).
Preserves wrong==0, trace_hash byte-equality, pack-binding, replay
determinism, no stochastic sampling.
Exit criterion: 50-case GSM8K train-split sample (unsealed) must show
correct >=10/50 with wrong==0 before any sealed-holdout run. If miss,
escalate to Path B (benchmark re-selection).
Supersedes ADR-0123b (never opened as PR).
ADR-0123-parser-comparison-phrasing as the **surface increment** on
PR #155's substrate (commit c9bd5d4). Closes the last architectural
gap in the comparison-phrasing class: before this commit, the
substrate's solver evaluated comparison problems successfully but
realize() crashed with `unknown operation_kind 'compare_additive'`
when asked for show-your-work prose.
Substrate (PR #155) already shipped:
- `Comparison` typed graph operand
- `compare_additive` / `compare_multiplicative` operation kinds
- parser patterns for the four canonical surfaces
(N more / N fewer / twice / N times / half)
- solver + verifier wiring + pack lemmas
(en-arith-006 compare_additive, en-arith-007 compare_multiplicative)
This surface adds:
- `_compare_additive_sentence(step)` rendering `direction='more'|'fewer'`
- `_compare_multiplicative_sentence(step, entity_units)` rendering
`direction='times'|'fraction'`
- two new branches in `_step_sentence` dispatch
- `_step_sentence` signature widened with optional `entity_units` map
(derived once-per-trace in `realize()` from `graph_initial_state`)
- ADR-0123-parser-comparison-phrasing.md (~15 invariants, substrate
+ surface decomposition rationale, multi-construction barrier
inheritance)
- 26 invariants pinned across canonical surfaces, plurality
independence, byte-determinism, refusal discipline, and
backwards-compatibility with the pre-comparison realizer templates
End-to-end pipeline now operates on all four canonical comparison
shapes:
parse_problem(
"Alice has 5 apples. Bob has 3 more apples than Alice. "
"How many apples does Bob have?"
) -> solve() -> realize().as_prose() ->
"Alice has 5 apples. Bob has 3 more apples than Alice, giving Bob
a total of 8 apples. Bob has 8 apples."
Measurement (this PR):
- 26/28 direct ADR-0123 tests pass; 2 skipped (CORE_HOLDOUT_KEY)
- `core eval cognition` byte-identical: 100/100/100/100
- ADR-0118 stepped-realizer templates re-render byte-identically
- Substrate measurements continue to hold
Honest non-result: sealed `correct_rate` stays at 0/1319. The
realizer cannot create matches the parser refuses; the multi-
construction barrier the substrate ADR documented holds at the
surface layer too. Cumulative lift signal expected only after the
3rd/4th foundational class lands (per ADR-0121's revised
sequencing). `wrong == 0` holds by construction — realizer only
renders successful traces.
Pre-existing failure noted (not introduced by this PR):
`tests/test_adr_0085_gloss_aware_cause.py::test_flag_off_metrics_byte_identical`
fails on substrate base (c9bd5d4) without these changes — an
ADR-0085 cognition baseline drift unrelated to the realizer.
Documents the Phase 5 GSM8K-math substrate completion across 7 narrative docs.
All 8 sub-phases of ADR-0119 (5.1 through 5.8) have landed on main; ADR-0114a's
10 anti-overfitting proof obligations are all discharged for the gsm8k_math lane.
Key facts surfaced in each doc:
- CORE-original public split: 150/150 correct, 0 wrong, 0 refused
- Real GSM8K test (sealed holdout): 0 correct, 0 wrong, 1319 refused
- Adversarial suite: 38 cases x 12 families, 0 wrong
- Depth curve: flat at 1.0 across depths 1-8 on public split
- Frontier baselines: Claude 3.5 Sonnet 96.4%, GPT-4 92.0%, Gemini 1.5 Pro 90.8%
- New lane shape gsm8k_capability_shape in LANE_SHAPE_REGISTRY
- New operational pack en_arithmetic_v1 (5 lemmas)
- ADR-0120 (first expert promotion contract) is the next gate
Docs updated: docs/PROGRESS.md, docs/capability_roadmap.md, docs/runtime_contracts.md,
docs/Whitepaper.md (§XIII), docs/Yellowpaper.md (gsm8k_capability_shape formal spec),
README.md, docs/decisions/README.md (current frontier).
No code changes. No new ADRs.
First worked attempt at promoting a domain under the ADR-0120
expert promotion contract. The contract refuses honestly.
Gate evaluation against live state:
ADR-0114a obligations: 10 of 10 pass
ADR-0120 contract-level gates:
audit_passed_holds ✓
correct_rate (public) ✓ 150/150 = 1.0
correct_rate (sealed) ✗ 0/1319 = 0.0 < 0.60 floor
signed_expert_claim ✗ (no entry, downstream of correct_rate)
Decision: mathematics_logic NOT promoted; stays at audit-passed.
Substantive blocker: parser grammar covers 0/1319 of real GSM8K.
What this proves
- The contract is genuinely falsifiable. ADR-0120 §"Threshold
rationale" deliberately set the floor above current measurement
so the first attempt would defer honestly. Same load-bearing
pattern as ADR-0107 → ADR-0110 for audit-passed.
- Wrong-zero discipline holds against real GSM8K (the load-
bearing positive claim). CORE refuses every problem outside
its grammar without confabulating on a single one.
What unlocks the promotion
Multi-ADR parser-expansion arc lifting sealed-GSM8K correct_rate
from 0.0 to ≥ 0.60. Each construction class (rate/comparison/
percentage/time-modal/etc.) ships as its own scoped ADR with:
- parser+solver+verifier+realizer extensions
- re-measurement on sealed holdout
- ADR-0118a OOD re-measurement (no surface-feature regression)
- ADR-0125 perturbation re-measurement (no invariance regression)
- ADR-0119.5 adversarial re-measurement (no new misparses)
Honest-fitting discipline: every lift is graded on the anti-
overfit obligations BEFORE the correct_rate change counts.
Tests: 6/6 with CORE_HOLDOUT_KEY; 4/6 + 2 skipped without (matches
ADR-0119.7 seal discipline).
This deferral demonstrates the expert tier's promotion machinery
is load-bearing — the gate has refused at least once before any
domain reaches it.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Defines the `expert` ledger tier (sixth status above `audit-passed`)
and the composition gate that governs every future promotion.
The gate composes ALL TEN ADR-0114a obligations plus three contract-
level checks:
- audit_passed predicate must hold
- correct_rate >= 0.60 on public AND sealed holdout splits
- signed expert_claims entry whose digest reproduces byte-equal
Numeric thresholds (load-bearing choices documented with rationale):
- correct_rate floor: 0.60 ("advanced" — above weak open-source LLMs,
forces real architecture work, raisable later)
- depth-curve flatness ε: 0.05 (accuracy(N) ≥ accuracy(d1) · 0.95^(N-1))
- drift tolerance from promotion: ±0.02
Documents the post-ADR-0120 sequence in Open Candidate Directions:
Phase 1: ADR-0121 first worked math promotion (likely deferral)
Phase 2: parser-expansion arc to lift sealed-GSM8K correct_rate
Phase 3: math expert promotion succeeds
Phase 4: second domain = symbolic_logic (60-70% of math substrate cost;
ProofWriter / PrOntoQA benchmark; same machinery class)
Phase 5: third domain = high-stakes refusal-centric (medical or legal;
wedge-sharpener; needs two prior successes first)
Phase 7: open candidate — multi-reviewer threshold signing for expert
No code lands with this ADR. Implementation ships under ADR-0120a
to keep the contract change reviewable independently.
The first promotion attempt (ADR-0121) will likely defer on the
correct_rate gate — current measurement is 0/1319 on real GSM8K.
That deferral IS the contract working as designed; same load-
bearing pattern ADR-0107 → ADR-0110 demonstrated for audit-passed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The 1,319 GSM8K test cases are now sealed at
evals/gsm8k_math/holdouts/v1/cases.jsonl.age, age-encrypted to the
ADR-0119.1 recipient. Plaintext never touched disk in the working
tree; only ciphertext is committed.
First honest CORE-vs-real-GSM8K measurement
cases_total: 1319
correct: 0
wrong: 0 ← ADR-0114a Obligation #4 holds against external corpus
refused: 1319
overall_pass: True
Zero confabulation. Parser refuses what it can't grammar-handle; the
"wrong == 0" discipline survives the move from CORE-original cases
to a real public benchmark. The 0/1319 correct rate is the truthful
gap that ADR-0120's threshold work will quantify.
What landed
scripts/seal_gsm8k_test.py
- Loads GSM8K via datasets.load_dataset("openai/gsm8k", "main")
- Strips worked-solution prose; extracts final-answer integer/float
after "####" (handles "2,125" → 2125 thousands-separator)
- Reads recipient from docs/holdout_recipients.txt (single repo key
per ADR-0119.1)
- Encrypts via pyrage; writes only ciphertext
- Refuses to overwrite test path with train-derived seal
evals/gsm8k_math/runner.py
- Empty expected_unit (sentinel) skips unit-comparison; grades on
answer value alone. Required because GSM8K answers carry no unit
structurally. wrong-zero discipline preserved.
tests/test_adr_0119_7_sealed_gsm8k.py — 6 invariants:
1. sealed file present + age-formatted
2. no plaintext companion files (sibling-leak guard)
3. decrypted JSONL matches documented schema
4. runner against decrypted suite produces wrong==0
5. tests skip (not fail) when CORE_HOLDOUT_KEY unset
6. case ids match "gsm8k-test-NNNN" pattern
Defensive gitignore: plaintext patterns under
evals/gsm8k_math/holdouts/v1/ are explicitly excluded.
ADR-0114a obligation roll-up
10/10 discharged for the gsm8k_math lane:
#1 ✓ sealed-holdout (fab_control + GSM8K test)
#2..#10 ✓ as before
Phase 5 status: 5.1..5.7 done; 5.8 in flight (PR #149). After 5.8
merges, ADR-0120 (first expert promotion contract) becomes
feasible.
Test plan
- pytest tests/test_adr_0119_7_sealed_gsm8k.py with CORE_HOLDOUT_KEY → 6/6
- pytest without CORE_HOLDOUT_KEY → 3 pass + 3 skip
- core test --suite smoke -q → 67/67
- CLAIMS.md regenerated (no diff)
- HF token NEVER in repo (saved at ~/.cache/huggingface/token, mode 600)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Umbrella ADR for Phase 5 of ADR-0114. Decomposes the eval-lane work
into eight sub-phases (5.1..5.8) that ship under their own ADRs.
Sub-phase scope:
5.1 — sealed-holdout encryption (one lane, fabrication_control)
5.2 — CORE-original GSM8K-style corpus (dev + public, 200 cases)
5.3 — lane runner (parser → solver → verifier → realizer)
5.4 — frontier-baseline comparison (Obligation #7)
5.5 — adversarial generation; misparse rate zero (Obligation #8)
5.6 — depth-curve measurement harness (Obligation #6, measurement)
5.7 — sealed GSM8K test as the holdout (Obligation #1, lane-side)
5.8 — overall lane gate; new gsm8k_capability_shape
Roadmap only — no code lands with this ADR. Each sub-phase ships
independently and discharges a specific ADR-0114a obligation.
Critical invariant: dev + public splits are CORE-original; the real
GSM8K test set enters only via the encrypted holdout under 5.7.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 4 of the ADR-0114 GSM8K-math roadmap. Consumes a SolutionTrace
and emits one sentence per step plus setup + answer sentences. Pure
function; same trace → byte-equal RealizedTrace.
What landed
generate/math_realizer.py
- realize(initial_state, trace) -> RealizedTrace
- Frozen RealizedTrace dataclass with canonical_bytes() + as_prose()
- Per-kind sentence rules (add / subtract / transfer / multiply×2 /
multiply×3 / multiply-general / divide)
- Singular/plural surface rule matches parser canonicalization
- Typed RealizerError on unrecognized step kinds
tests/test_math_realizer.py — 60 cases pinning five invariants:
1. All 50 dev-set cases realize without error
2. Determinism: byte-equal RealizedTrace across two calls
3. Setup sentence count == initial_state count
4. Step sentence count == operation count
5. Answer sentence contains the resolved value + unit
ADR-0114a obligation discharge update
ADR-0118 hardens determinism (#9) across a third layer (realizer)
and makes #3 / #10 human-inspectable via the prose surface. No
obligation is directly newly discharged by ADR-0118; it's substrate
for ADR-0119 GSM8K eval lane.
Round-trippability of the prose through the parser is explicitly
out of scope for this phase. The trace is the verifiable artifact
(ADR-0117); the prose is human-readable documentation.
Tests: 60 new realizer cases; 546 total green across realizer +
parser + solver + verifier + OOD; 67/67 smoke green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 3 of the ADR-0114 expert-capability roadmap. Re-applies every
step of a SolutionTrace from the input graph's initial state and
asserts byte-equal reproduction of answer_value. Pure function; same
(graph, trace) → byte-equal VerifierVerdict.
Why this is distinct from the solver
ADR-0116's solver enforces correctness at construction. ADR-0117's
verifier is a SECOND, INDEPENDENT implementation that re-derives
every value the trace claims. The verifier does NOT call solve(). It
re-implements the operation semantics from ADR-0116 directly inside
_verify_step. If the solver had a bug or was tampered with after the
fact, the verifier catches it.
Six checks per verdict (named, ordered, audit-logged):
1. graph_canonical_hash_matches
2. pack_id_matches
3. pack_lemmas_resolve
4. step_pack_lemma_ids_match_bindings
5. step_replay_matches_before_after
6. answer_value_reproduces
Seven named tamper classes all caught:
- mutated before_value / after_value / operand of any step
- mutated pack_lemma_id of any step
- mutated graph_canonical_hash
- mutated answer_value
- mutated pack_id
- mutated target_before / target_after of transfer step
ADR-0114a obligation update
#3 Replay-equal trace — now discharged at VERIFIER FIDELITY
(was solver-only under ADR-0116). A third party with only
(graph, trace, pack) can reproduce the answer byte-equal.
Five of ten obligations now load-bearing: #3, #4, #9, #10 plus
in-flight #2 (Codex's ADR-0118a OOD generator).
Tests: 62/62 verifier suite green; 67/67 smoke green; existing
solver + parser + schema suites unaffected.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 2 of the ADR-0114 expert-capability roadmap. Consumes the
MathProblemGraph from Phase 1 and emits a SolutionTrace — ordered
operation applications ending at a numeric answer, byte-deterministic
across runs, with each step's operation bound to a pack-resolved
lemma identifier.
What landed
generate/math_solver.py
- solve(graph) -> SolutionTrace; pure function, no I/O, no globals
- SolutionStep dataclass with before/after values per step (for
verifier replay; ADR-0117 hardens)
- SolutionTrace with canonical_bytes() byte-deterministic JSON
- SolveError typed refusal: missing pack, division by zero,
unknown-references-nothing
language_packs/data/en_arithmetic_v1/
- 5 operator lemmas: add / subtract / multiply / divide / transfer
- role=operational_base (vocabulary-only; no domain claim)
- SHA-256-anchored lexicon + glosses; manifest carries
provenance=adr-0116:operator_seed:2026-05-22
tests/test_math_solver.py — 109 cases pinning five invariants:
1. Phase 2 exit criterion: ≥ 0.80 on parser-correct dev set
(current: 50/50 = 1.00)
2. Determinism: two solves produce byte-equal trace
3. Trace replay reproduces answer_value (verifier rehearsal)
4. Typed refusal on under-determined inputs
5. Every step.pack_lemma_id resolves to a real lexicon entry
in en_arithmetic_v1
ADR-0114a obligation discharge
Four of ten anti-overfitting obligations now have load-bearing
implementations in code:
#3 replay-equal trace — discharged (solver-layer)
#4 typed refusal — discharged (solver-layer)
#9 determinism — discharged (solver-layer)
#10 operation provenance via pack — DISCHARGED IN FULL
Removing the en_arithmetic_v1 pack now breaks every solve loudly.
The "operations bind to concepts, not hardcoded strings" claim is
architecturally true, not rhetorical.
Tests: 109/109 green on solver suite; 67/67 smoke suite green;
parser + schema suites still green from prior phases.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Documentation-only amendment to ADR-0114. Locks in the 10-point
falsifiable proof framework that any future `expert` ledger-tier
promotion (ADR-0120+) MUST require.
The obligations:
1. Sealed-holdout discipline
2. OOD surface variation ≥ 0.95 of public
3. Every correct answer ships with replay-equal trace
4. Refusal is first-class; misparse rate zero; zero `wrong` answers
5. Reasoning-isolation perturbation suite (invariance + predictable change)
6. Compositional-depth curve flat within documented ε
7. Frontier-baseline comparison on identical items, published
8. Adversarial generation; misparse rate zero
9. Determinism across release boundaries
10. Operation provenance via the pack (not hardcoded strings)
Each obligation is load-bearing and falsifiable: a domain that
cannot satisfy any one stays at `audit-passed`. ADR-0114a binds
ADR-0116..ADR-0120 to carry the obligations into implementation;
ADR-0120 finally invokes all ten as hard gates.
The audit-passed tier (ADR-0106/0109/0113) is unaffected. The two
tiers measure orthogonal properties: audit-passed verifies CORE
claim-shape compliance (transformer-unreachable invariants); expert
verifies capability with anti-overfitting proof.
No code change. Pure forward contract for the next phase of work.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
First Phase of ADR-0114's expert-capability roadmap. Decomposed into four
sub-phases so each lands as its own auditable step:
1.1 schema + 5 seed cases + invariants ← this commit
1.2 45 more dev-set cases ← delegated (Codex)
1.3 the parser itself ← exit: ≥0.90 on dev set
1.4 runtime binding ← if non-trivial
What landed
- generate/math_problem_graph.py — typed dataclasses (Quantity,
InitialPossession, Operation, Unknown, MathProblemGraph) + frozen
validation + canonical_bytes() byte-deterministic serialization +
graph_from_dict roundtrip.
- evals/gsm8k_parser_dev/cases.jsonl — 5 seed cases (gpd-001..005)
covering single-add, single-subtract, multi-step, two-entity
transfer, and multi-entity sum constructions. Every case carries a
ground_truth_graph and the documented patterns it exercises.
- evals/gsm8k_parser_dev/README.md — authoring contract: schema,
pattern registry, canonicalization rules, Phase 1.1 scope boundary,
hand-solving rubric, distribution target for the remaining 45
cases. This is the spec Phase 1.2 authors work against.
- tests/test_math_problem_graph.py — 26 cases pinning four invariants:
round-trip byte equality, canonical_bytes() determinism, schema
rejection of malformed graphs, and ground_truth_graph ↔
expected_answer agreement (a hand-solver inside the test module
falsifies mis-authored cases).
Why this is sticky
The Phase 1.1 schema is load-bearing for Phase 1.2 (the 45 authored
cases will be written against it) AND Phase 1.3 (the parser will be
graded byte-equal against ground-truth graphs in this schema). Changing
the schema after Phase 1.2 lands requires an amendment ADR + rewriting
authored cases. The schema choices here are intentionally conservative.
Tests: 26/26 new; 67/67 smoke green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The word "expert" in the previous status name implied raw-capability parity
with frontier LLMs on the same benchmark — which the gate does NOT verify.
What the gate actually verifies is CORE *claim-shape compliance*:
* signed digest (replay-reproducible from on-disk lane results)
* replay determinism (same inputs → byte-equal trace_hash)
* typed refusal (fabrication refused, not paraphrased)
* exact recall (no ANN, no cosine, no attention bottleneck)
* grounding-source provenance
These are claim shapes a transformer LLM cannot structurally produce
regardless of raw accuracy. A frontier LLM might score higher on the
same benchmark but cannot pass this contract.
Rename scope (semantics only, per ADR-0113):
status string "expert-demo" → "audit-passed"
predicate key predicates.expert_demo → predicates.audit_passed
reason key expert_demo_reason → audit_passed_reason
YAML key expert_demo_claims → audit_passed_claims
CLI command core demo expert → core demo audit-passed
output dir evals/expert_demos/ → evals/audit_passed/
artifact filenames expert_demo.{json,html} → audit_passed.{json,html}
HTML title CORE Expert-Demo: X → CORE Audit-Passed: X
Internal Python identifiers (module/file/function/class names like
`expert_demo.py`, `evaluate_expert_demo`, `ExpertDemoClaim`,
`expert_demo_claim_for`) are deliberately kept to minimize churn. ADR
file titles (ADR-0106..0112) preserved as historical record.
`expert` namespace reserved for ADR-0114+: an actual capability tier
above `audit-passed` backed by a public benchmark with a stated
threshold. ADR-0114 proposes the first such target — GSM8K-math —
laying out a falsifiable 7-phase arc (parser → solver → verifier →
stepped-realizer → eval lane → first `expert` ledger tier promotion).
Tests: 184 directly-affected tests green (140 capability/expert-demo
suite + 34 demo/audit-tour + 10 correction-cue). Smoke suite 67/67.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Closes the asymmetry between the `expert-demo` ledger status (audit
artifact only) and the actual `core demo` surface (runnable
walkthroughs producing HTML + JSON). Until this commit the word
"demo" in `expert-demo` was aspirational; now it corresponds to
something a reader can open.
What it does
- Reads the signed expert_demo_claims entry from docs/reviewers.yaml
- Loads latest on-disk result files for each attached lane × split
- Re-derives the evidence-bundle digest and asserts byte-for-byte
match against the signed claim_digest — this is the load-bearing
audit step, now exercised at two independent enforcement points
(ledger gate + showcase)
- Runs each lane's metrics through the ADR-0109 lane-shape registry
and surfaces the verdict
- Picks the first three cases from each split verbatim (deterministic
by file order) and renders them as HTML for inspection
- Emits expert_demo.json (canonical bytes, deterministic) + expert_demo.html
Surface
core demo expert --domain mathematics_logic
core demo expert --domain physics
# → evals/expert_demos/<domain>/latest/expert_demo.{json,html}
Read-only by construction: cannot mutate docs/reviewers.yaml or any
lane result file. Tested. Unpromoted domains raise ValueError —
no silent fallback, no "preview" mode that fakes a showcase.
Generated artifacts are gitignored — the inputs they derive from are
already committed, so duplicating the renders would just churn the
tree.
Tests: 16 new cases pinning all five ADR-0112 invariants. Smoke suite
still 67/67 green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
