Implements ADR-0181 PR-6 (eval-plan §4): teachers label or align; they never
define the substrate and never fold embeddings into the versor path.
- sensorium/audio/teachers.py:
- TeacherHint: typed, versioned, checksummed annotation (no raw embeddings).
- AudioTeacher protocol (pure on the signal).
- attach_teacher_hints: the ONLY admission path — appends content.* anchors to
the IR's content_anchors (immutable, recomputes ir_sha256). content.* is not
an operator key, so compile_events skips it: versor + projection_sha256 stay
byte-identical; only the ir leg of the merge_key moves (evidence recorded).
- KNOWN_TEACHER_LANES (whisper/nemo/clap/encodec): declared + gated behind
optional extras; load_teacher import-guards and fails loudly (never a silent
fallback). StubTranscriptTeacher is the deterministic reference instance.
- parser.py: extract _ir_payload + ir_sha256_of (DRY single source of truth for
ir_sha256; byte-identical to parse() output — regression-guarded).
- pyproject.toml: audio-whisper/nemo/clap/encodec optional extras (never
runtime-required).
16 failable proof tests in tests/test_audio_teachers.py. Load-bearing:
test_teacher_hint_does_not_change_versor. Mutation-verified — giving a teacher
anchor an operator event_type (folding it into the versor) fails the
versor-invariance proof; reverted, all pass.
Additive only (ADR-0013): no core layer touched. Audio suite 57/57; eval-gate
ir_sha256 pins unchanged by the parser refactor; architectural invariants 40/40.
Real model adapters are deferred until extras+weights are present; this PR ships
the policy, the typed-hint contract, and the shadow-only guarantee.
Implements ADR-0182's first win on top of EX-6 (#473). The distractor-quantity
confuser 0014 misfired (20x3x5=300) because a blunt product-of-all was the *unique*
self-verifying reading: the completeness clause forces the distractor into it, and
no rival reading existed to trigger the wrong=0 disagreement rule. No tight cue rule
separates it from the legitimate cross-unit products (`for` licenses both the 0014
distractor AND the correct train-0021 product) -- that is the deferred cue-precision
problem. So instead of a reactive patch, let the disagreement rule do the refusing.
Mechanism (sealed lane; chat/ does not import these -> serving 3/47/0 frozen):
- verify.classify_derivation: a derivation is `complete` (commit-eligible),
`exempt` (verified but for an isolated-foreign unused quantity -> commit-
INELIGIBLE), or None. Refactored self_verifies into _base_reasons +
_unused_quantities so the two share logic and cannot drift (behavior identical;
385 derivation tests + smoke 67 green).
- accumulate.accumulation_candidates: exposes the ungated readings, incl. a
distractor-skip reading that drops an isolated-foreign quantity from a multi-
quantity change clause (20+5=25). compose_accumulation is byte-identical
(drop_isolated_foreign=False + the same gate).
- search.multiplicative_candidates / multistep.candidate_chains: ungated candidates.
- pool.resolve_pooled: pools every composer's readings; disagreement -> refuse; a
single answer commits only if a `complete` candidate produced it (exempt-only ->
refuse, so the commit-path completeness guarantee from ADR-0175 is untouched).
- confuser runner: _engine_answer now delegates to resolve_pooled (the prior
first-composer-wins order could not notice it held two incompatible readings).
Result (the microscope):
- confuser wrong 5 -> 2. distractor 0014 refuses (product 300 vs additive 25);
BONUS: both disguised-polarity cases (0001/0003) refuse -- the spurious
"buys X for N coins" product disagrees with the accumulation reading. Remaining
wrong: 0016 (distractor in the anchor clause -> needs anchor-skip, separate step)
and 0020 (temporal-scope). pair-tells 4 -> 1.
- genuine positives still 7 solved, 0 wrong.
- train_sample 3/47/0 and practice 3/47/0 byte-identical (they call
compose_accumulation/search directly -- unchanged -- not the pool).
- smoke 67, architectural invariants 40, lane-SHA freeze 8/8.
Tests:
- test_adr_0182_pool.py: classify (complete/exempt/None incl. narrow-exemption
edges) + resolve_pooled, with the wrong=0 obligation test_exempt_only_never_commits
(a distractor with no multiplicative cue must refuse, not commit 25 -- fails loudly
if the exemption is made commit-eligible).
- test_adr_0163_f2_confusers.py: baseline tightened wrong 5->2, pair_tells 4->1;
new test_distractor_0014_refuses_via_pooling + test_disguised_polarity_does_not_misfire.
Stacked on #473 (EX-6); merge #473 first. 0016 + the remaining wrongs are follow-ups.
The confuser probe's two pseudo-accumulation misfires (0005 ->796, 0007 ->996)
both traced to the same extraction blind spot: a number bonded to its unit by a
hyphen (`25-foot sections`, `20-inch pieces`) was invisible to the base
`number + space + word` pattern, so the self-verification completeness clause
never saw the divisor and the bare `buys ... gives` accumulation read as
"complete". This is the highest-leverage lever the 2026-05-29 session named
("the gate must see the fractions/25-foot it currently misses").
EX-6 adds a tight, ADR-0165-safe lexeme pass: a digit run, a single hyphen, an
alphabetic unit word. The alphabetic-only unit group keeps numeric ranges (`3-5`)
out; taking only the first hyphen segment keeps the postmodifier tail
(`25-year-old`) from inflating the unit — so it stays clear of the deferred EX-3
multi-word-unit traps. Over-extraction here is strictly refuse-preferring: making
the divisor visible drives 0005/0007 to refuse via the polarity-None
`cuts`/`splits` clause, never to a wrong answer.
Evidence (deterministic, the microscope):
- confuser probe: wrong 7 -> 5; pseudo-accumulation 0 wrong / 4 refused;
genuine positives still 7 solved; pair-tells unchanged (4).
- train_sample (capability): 3/47/0 byte-identical.
- practice accumulation: 3/47/0 (wrong=0) byte-identical.
- smoke 67 passed; lane-SHA freeze 8/8 (serving frozen).
Tests:
- TestEX6HyphenatedUnitNumbers pins the new lexeme (value+unit, decimal, no
double-count, word-compound unaffected, numeric range not read as a unit).
- TestProbeBaseline tightened wrong 7->5; new test_pseudo_accumulation_does_not_misfire
is the failing-under-violation obligation (fails loudly if the pass regresses).
- TestSlashFractionLeakHazard pins the deferred `1/4`->`4` denominator leak: not
fixed here because suppressing the leaked operand *removes* a quantity and can
unblock the completeness clause (not unambiguously refuse-preferring), so it
needs its own train_sample + probe validation.
PR-4 of ADR-0181. The acceptance-gate lane that decides whether audio_core_v1's
gate may open. Deterministic synthesis-spec fixtures (no .wav blobs) with
predicted parses, so the gates grade parser semantics as well as determinism.
evals/audio_sensorium/:
- synth.py deterministic fixture synthesis (PCG64 + float32-at-boundary)
- fixtures.json 5 specs: silence, rising-pitch question, falling statement,
noise burst, speech-then-pause
- generate_expected.py reproducible pin generator (uv run -m ...)
- expected_ir.jsonl frozen canonical_sha256 + ir_sha256 + event_type_counts
- expected_projection.json frozen projection_sha256 + reference versor
tests/test_audio_eval_gates.py (12): the gate table per fixture —
shape/dtype, versor_condition<1e-6, within-run replay, canonical-checksum
stability (hard int/cast-stable pin), IR-replay + frozen ir_sha256, semantic
event_type_counts (parser-accuracy gate), and cross-platform versor stability
within atol=1e-6 of the reference (float-safe per eval plan); plus trace
hygiene and gate-closure refusal.
Verified semantics: rise→prosody.rise, fall→prosody.fall, silence→pause.long+
turn.boundary, noise→nonspeech.noise, speech_then_pause→all three.
Cross-platform note: int/quantized-derived hashes are pinned hard; the float
versor is compared within tolerance rather than hash-pinned, since cos/sin/
geometric_product can differ by a ULP across arches. This is the eval-plan's
"equal within declared numeric tolerance" reading — keeps CI stable.
All audio 44 + arch-invariants 40 + smoke 67 green. No core mutation.
Builds the corpus from the ADR-0163-F2 spec: 30 hand-curated, real-sourced cases
across the proven misfire categories (disguised-polarity, pseudo-accumulation/
fractions, multi-referent H1, multi-actor-pronoun ADR-0174, distractor-quantity,
temporal-scope H3, comparative-referent H2, unit-confuser) + genuine-positive
minimal-pair twins. Schema carries category/surface_trap/expected/pair_id/source.
The runner scores OPPOSITE to a coverage lane: the bar is `wrong` -> 0 (a confuser
*answered* is a defect regardless of value) plus pair-consistency (solving a twin
but answering its confuser = a surface-matching tell). It runs the realistic sealed
attempt (accumulation -> multiplicative -> chain, first to resolve).
Honest measured baseline (the probe's whole point — these are the defects the
templated corpus hid): 30 cases -> 7 solved / 15 refused / 7 WRONG / 1 spurious;
4 pair-tells (0001/0003/0014/0020). Wrong by category: disguised-polarity 2
(buys-a-toy-for-30 -> +30), pseudo-accumulation 2 (the 0002 cable/fraction),
distractor-quantity 2, temporal-scope 1 (before-giving -> gave the now-value).
Per the overfitting lesson, the composers are NOT reactively patched to pass the
probe (that is the trap). The baseline is pinned as a no-regression gate (wrong
<= 7, pair-tells <= 4, positives keep solving); future fixes must be GENERAL
mechanisms validated on train_sample, driving wrong down. Sealed: serving 3/47/0
byte-identical (lane-SHA 8/8, claims OK); architectural invariants green.
The first cross-clause comprehension reading: one actor's quantity changes over
successive clauses ("Sam has 14 apples. He buys 9 more." -> 14 + 9). It is the
safe specialisation of the cross-clause sum that GB-3a refuses wholesale (the
Alice/Tom hazard) — we chain only when (same referent) AND (a licensed change
cue of unambiguous polarity), else refuse.
generate/derivation/accumulate.py — compose_accumulation:
- anchor on clause 1's single quantity; apply +M (gain) / -M (loss) per later
change clause, operand taken in the anchor's unit (accumulation is same-dimension);
routed through the unchanged self-verification gate.
- polarity (ordered, so ambiguous "gives" is resolved not guessed): "more" -> gain;
else unambiguous loss verb -> loss; else gives/gave + to/away -> loss; else
unambiguous gain verb -> gain; else REFUSE.
- referent guard (the ADR-0174 multi-actor hazard's defensive fix, built minimally
in the clean lane — NOT the retired gender-blind resolver): a later clause's
subject token must be a pronoun or the anchor's name; a NEW named subject (Tom)
-> refuse. Pronoun gender/number is not matched; a new name is the only signal.
evals/.../accumulation_runner.py — practice scorer: on a base refusal, attempt
compose_accumulation and gold-check (mirrors search_runner). Sealed: fires only
on already-refused cases, never alters serving.
Measured (sealed practice additive lane): 0 -> 55 correct, wrong unchanged at 1
(the base scorer's pre-existing one; accumulation added 55 correct, 0 wrong). The
36 still-refused are multi-change (GB-3b.2) or unrecognised verbs (vocab growth) —
conservative, never wrong.
Proof obligations (tests fail under the violation): new-named-actor refuses (H1),
no/ambiguous change cue refuses, list anchor refuses, multi-change refuses,
determinism. 136 targeted tests + architectural invariants green; serving 3/47/0
byte-identical (lane-SHA 8/8, claims --check OK).
ADR-0180 §1.5.4 + CLAUDE.md work-sequencing item 5 require these four
properties green on main before any core-rs/src/vault.rs change. They are
also the foundation ADR-0181 PR-5 (audio Delta-CRDT wiring) rides on.
T-1 set-equality of vault writes under shuffled ingest (+ idempotent
re-ingest at the content-addressed layer)
T-2 trace-hash invariance to vault order, + recall result-set invariance
to insertion order (the genuinely-failable half)
T-3 versor_apply non-commutativity (negative guard)
T-4 ProjectionHead.project purity across calls and threads
Findings (docs/audit/ADR-0180-t1-t4-findings.md):
- compute_trace_hash folds only vault_hits (a count), NOT vault contents, so
ADR-0180 §1.5.3's "re-sort vault state in content-addressed order" is
currently vacuous at the trace-hash layer; the live order-invariance
obligation is at recall() (result-set + count). Recommend amending §1.5.3.
- equal-score recall ties are index-sensitive; the Merge Kernel needs a
content-addressed tiebreak (mirrors ADR-0181 §2.2 merge key). Recommend
amending §2.2.
- append is genuinely semilattice-eligible; versor_apply is non-commutative.
7 passed; smoke suite green. No runtime/core mutation — tests + audit only.
CP-2a populates the CP-1 ledger from gold-labelled candidate readings and reports
per-pattern reliability — the measurement the cue-precision thesis rests on. Plus
the function-word unit filter, whose value this measurement makes concrete (clean
unit_shape labelling).
What landed (all sealed; serving 3/47/0 byte-identical):
- generate/cue_precision/trainer.py — train_from_cases(cases, enumerators): folds
gold-labelled candidate chains into the ledger via record_case. Decoupled (the
candidate enumerators are injected, so the package still imports nothing from
search). candidates_for dedupes a reading shared by two enumerators.
- generate/derivation/multistep.py — extracted the enumeration half of search_chain
into public candidate_chains(problem_text); search_chain now delegates (verified
byte-identical: ms3 tests + practice counts unchanged). CP-2 needs the readings
the search weighs, not just the one it resolves.
- generate/derivation/extract.py — function-word unit filter (_NON_UNIT_WORDS):
blanks spurious function-word units ($0.75 each -> "", 3/4 of -> "") that
corrupt same-unit detection and unit_shape. Closed lexeme set, ADR-0165-safe.
- evals/gsm8k_math/practice/v1/cue_precision_report.py — trains over 200 sealed
cases (50 train_sample + 150 ADR-0163-F additive) with the real enumerators and
prints the per-pattern reliability table.
- tests/test_adr_0177_cp2a_training.py — trainer obligations (credit/dedupe/
determinism/empty) via synthetic enumerators; real-measurement well-formedness;
search_chain parity.
Load-bearing finding (recorded in ADR-0177): over 200 cases EVERY (cue,op,unit_shape)
pattern floors at ~0.0 reliability (best: for-multiply-cross_unit 0.0116 at 2/34).
The blunt product/sum-of-all readings are almost always wrong vs gold, so the
conservative floor correctly trusts nothing. => CP-2b (trust reliable cues) is
blocked on candidate GENERATION, not the ledger: candidate readings must get less
crude (clause/referent structure, ADR-0178 GB-3b) before any cue earns reliability.
Cue-precision and compositional structure are coupled; structure comes first.
Verification: 107 targeted tests green (CP-2a/CP-1/extract/ms3/GB-1/2/3/MS-1/2) +
architectural invariants; serving CLAIMS.md sha unchanged; practice 4/1/45 and
0/1/149 unchanged. Inert: trains/reports only, consulted by no search/gate.
Track C of docs/handoff/PARALLEL-WORK-PLAN-2026-05-29.md asked for a tight
EX-3 multi-word-unit redo satisfying (a) "12 jumping jacks." -> "jumping
jacks", (b) "6 apples and 4 apples." -> two apples, (c) all GB-1/2/3 tests
green. The cleanest tight rule that satisfies all three —
(?<![\w.])(\d+(?:\.\d+)?)\s+([a-z]+\s+[a-z]+)(?=\s*[.?!,]|\s*$)
— was implemented and passed the four pinned test files. Full-suite
verification then surfaced a second trap the audit at
docs/handoff/AUDIT-ADR-0179-EX-RECONCILE.md did not anticipate:
postmodifier-adjective tails. "25 years old?" fires the tight rule and
produces unit "years old" rather than "years", regressing
test_adr_0176_ms1_question_target.py::TestQuestionQuantities::
test_extracts_quantity_stated_in_question and the "X years old" pattern
in tests/test_adr_0176_ms2_chain.py. The pattern is endemic in GSM8K
(cases 0006 and 0033 both use "X years old"); closing it would need a
second closed lexeme set ({old, tall, long, wide, deep, away, ago, ...})
which the brief judged too open-ended to enumerate responsibly.
Per the brief's escape hatch ("If no rule satisfies all of (a)-(c) without
a grammar template, write a note and ship no code — a refusal is fine")
this commit:
* updates extract.py's module docstring to name BOTH known traps
(connective-crossing AND postmodifier-adjective tails);
* adds tests/test_adr_0179_extract.py::TestEX3StillDeferred with two pins
asserting the postmodifier-adjective shape stays at unit "years" alone,
so no future redo silently re-introduces the regression;
* ships NO extractor code change — the regex remains exactly as on main.
Scope/safety:
* Files touched are within Track C's allowed set (extract.py + its test).
* Zero functional change: extract_quantities byte-identical to main.
* Serving lane untouched (chat/ does not import this module).
* Safe alongside GB-3b on compose.py / clauses.py.
CP-1 of ADR-0177: the per-(cue, op, unit_shape) reliability ledger + credit
assignment mechanism. Mirrors the ADR-0175 per-class ledger discipline
(core/reliability_gate/ledger.py): counts-only integers, reliability via the
pinned conservative_floor, refusals never counted as commitments.
- generate/cue_precision/ledger.py
- CuePattern: (cue, op, unit_shape) key; op in VALID_OPS, unit_shape closed-set.
- pattern_for_step / patterns_in_chain: per-step extraction. unit_shape compares
the operand unit to the model's running (primary/start) unit; a dimensionless
comparative scalar scales within the dimension -> same_unit.
- PatternTally: counts-only (correct/wrong, no refused axis); reliability =
conservative_floor(correct, committed); 0.0 while cold/below N_MIN.
- CuePrecisionLedger: immutable pattern->tally map (canonical sorted tuple);
record_chain / record_case credit candidate chains by gold label, independent
of whether the search resolved or refused.
Inert substrate: not wired into the gate, any scorer, or the search (CP-2/CP-3).
Imported by nothing outside its own tests (asserted by a source-tree scan).
Tests (tests/test_adr_0177_cp1_ledger.py, 27 passing): pattern validation;
unit_shape classification; cold ledger -> 0 reliability; credit assignment;
refusals-not-counted; reliability earned by volume; determinism/replay;
immutability; inertness scan. Smoke suite green (67 passed).
The mandated lookback review before GB-3 (CLAUDE.md §Lookback Review Discipline)
confirmed the audit's hazards H1/H2/H3 were LIVE: compose_sequential summed
same-unit quantities from the whole problem, merging unrelated referents/scopes
and admitting wrong structures whose value happened to ground:
H1 (second actor's apples) -> 6+4+2 = 12
H2 (comparative on other actor) -> (6+4)*2 = 20
H3 (later depletion event) -> 6+4+3 = 13
Root cause is the audit's G1/D2 drift: GB-2a re-extracts from the whole text and
ignores GB-1's clause structure. The fix is the GB-3 increment — make the composer
clause-scoped (consume segment_clauses), refusing when the licensed structure spans
clauses, because this slice cannot model referents:
- quantities must live in exactly ONE clause (0 or >1 -> refuse);
- a comparative outside that clause -> refuse (unmodelled referent binding).
All three hazards now refuse; all 7 GB-2 single-clause structures preserved
(list-sum, three-item, sum-then-scale, and the mixed-unit/disagreement/too-few
refusals). tests/test_adr_0178_gb3_referent_guard.py would fail against the
pre-guard code (12/20/13), so the obligation is proven, not decorative.
Scope/safety:
- compose_sequential is sealed substrate, not wired to a scorer -> serving
byte-identical 3/47/0 (lane-SHA 8/8, generate_claims --check OK); practice
unchanged 4/1/45. No new test failures (2 pre-existing on main).
- ADR-0178 amended: GB-2 relabelled GB-2a (list slice, drift G1 recorded);
GB-3 split into GB-3a (this referent guard, landed) and GB-3b (constructive
cross-clause chaining, next).
Reconciles ChatGPT's four independently-branched extraction PRs (#451/#452/
#453/#454) into one coherent generate/derivation/extract.py. They each rewrote
the same file + same new test off main, so they conflicted pairwise and needed
integration, not a merge.
Integrated (span-tracked, most-specific-pass-first so numbers are never double
counted):
- EX-1 word-numbers (#452): reuses WORD_NUMBERS; tens-one hyphen compounds;
factor-bearing half/third/quarter excluded.
- EX-4 list-unit inheritance (#451): bare numeric list with one trailing unit.
- EX-5 sentence-final numbers (#454): bare final number with empty unit.
Deferred: EX-3 multi-word units (#453). Its greedy lowercase span reads
"6 apples and 4 apples" as unit "apples and", regressing GB-2's
test_same_unit_list_sums, and still can't recover real multi-word units from
0024-class text ("jumping jacks on"). Needs a tighter rule; see
docs/handoff/AUDIT-ADR-0179-EX-RECONCILE.md.
Verification (sealed lane only; chat/ does not import this module):
- Serving frozen: lane-SHA 8/8 match, generate_claims --check OK -> 3/47/0
byte-identical, wrong=0 held.
- Sealed practice improved 4/2/44 -> 4/1/45: case 0025 flips wrong->refused.
EX-1 reads "three", so completeness sees a quantity the 6x50 chain omits and
refuses the spurious 300 (gold 1200) instead of committing it.
- No new test failures (3 pre-existing on main).
Also fixes stale test drift from EX-2 (#447): TestDecimalGroundingGapIsDeferred
asserted decimals still refuse, but #447 made $0.75-class resolve to 864.
Renamed to TestDecimalGroundingResolves and updated to assert the flip.
Honest scope note: EX-4 does NOT unblock real case 0024 (its PR test used a
fabricated bare-list paraphrase). TestRealCase0024StillBlocked pins the true
boundary.
EX-2 — the ONE shared-primitive (serving-path) touch of extraction richness.
_value_grounds already grounds the symbol form $N.NN (currency) and N/M (fraction);
a decimal written WITHOUT a symbol ('0.75') is never a single token (the tokenizer
splits on '.') so it failed to ground — refusing correct products like 0003
(48*24*0.75=864). Now a bare decimal 'N.M' grounds when both digit-runs appear,
symmetric with the $N.NN and N/M branches. Only returns True on a match; non-matching
decimals fall through (refuse).
wrong=0 (the load-bearing obligation, this being a shared/serving-path change):
- serving 3/47/0 BYTE-IDENTICAL (verified).
- round-trip + candidate-graph tests 51/51.
- lane-SHA gate (pinned eval lanes unchanged) — verified before merge.
Payoff: 0003 unblocked in sealed practice (search_chain -> 864 = gold; +1 flip).
Decimal cases that now ground but mis-compose become sealed eliminations (learning
signal); serving untouched. 6 EX-2 tests (decimal grounds/refuses, integer
unchanged, serving byte-identical, 0003 resolves).
GB-2 first increment (ADR-0178). compose_sequential() adds the structure the blunt
MS-3 shapes couldn't reach: a same-unit quantity LIST sums (additive cue), and any
stated comparative scales the sum (sum-then-scale, 0024-family). Op-per-step from
text structure (list => add; comparative => scale); operands are text quantities
(grounded) + comparative steps (cue-grounded) on the flat left-fold — no derived-
intermediate model needed (running value is the intermediate).
Deliberately narrow: same-unit lists only. A stated comparative is ALWAYS applied
(no bare-vs-scaled self-disagreement). A product base over the same list is added
WITHOUT a comparative tail purely as a disagreement-safety candidate -> a same-unit
list that also carries a mult cue (ambiguous) REFUSES. Product-of-all/cross-unit
products stay MS-3's job (avoids the product x comparative blowups a blunt all-bases
composer produced: 0024 -> 4.3M).
Clean-case capability proven: 8 tests (list-sum, sum-then-double/triple, mixed-units
refuse, ambiguous-disagreement refuse, determinism). Honest practice result: 3/2/45
— NO new flips (extraction wall: real cases like 0024 extract non-uniform units
'36 on' so they aren't seen as same-unit lists), 2 sealed eliminations (0037/0039:
list-sum was the wrong structure -> learning signal). Coverage gated by extraction
richness + cue precision, as predicted.
Sealed; serving untouched. Full derivation surface 53/53; ruff clean; smoke 67.
Continuation: richer relational ops (per/each->multiply, more/older->add), branch/
DAG (0033), and the extraction richness (uniform-unit extraction) that unblocks this
on real cases.
GB-1 — first slice of the comprehension-guided composer (ADR-0178). Reads the
problem one clause at a time and derives each clause's LOCAL contribution; GB-2
combines them across clauses.
generate/derivation/clauses.py:
- segment_clauses(text): sentence-level orthographic split (ADR-0165; not grammar).
- clause_local_results(text) -> tuple[ClauseResult]: per clause, 0 quantities =
context (hold), 1 = leaf (its value), >=2 = bounded local search (reuses MS-3
search_chain). Refuse-preferring: ambiguous multi-quantity clause -> unresolved
hold, not guessed.
Locality is the guidance that bounds the search + steers grouping. 9 GB-1 tests
(segmentation, leaf/context/local-product, ambiguous-holds, determinism,
per-clause structure of a multi-sentence problem). Full derivation surface 86/86;
ruff clean; smoke 67. Sealed; not wired into serving (ClauseResults ready for
GB-2 sequential combination).
MS-3 composes MS-1 (Target) + MS-2 (comparative chains + completeness) + the gate.
generate/derivation/multistep.py: search_chain(problem_text, target=None).
Shape-based, NOT blind enumeration: enumerates a small principled candidate set
(product-of-all if a multiplicative cue is present; sum-of-all if an aggregation
hint is present; each optionally + comparative scalars), each using all
quantities, routed through select_self_verified (grounding ∧ cue ∧ unit ∧
completeness ∧ uniqueness). Bounded (MAX_QUANTITIES, refuse-on-overflow) +
deterministic. Target supplies the aggregation cue + question quantities; target-
UNIT matching is deferred (answer_unit=start.unit is wrong for cross-unit products
-> a unit gate would over-refuse; documented).
Honest practice measurement (sealed lane): 4 correct / 9 wrong / 37 refused
(baseline 3/0/47). +1 flip is the unambiguous whole-problem product (0021); the 9
wrongs are product-of-all eliminations on multi-step problems (caught by gold,
the learning signal). Whole-problem shapes add no coverage beyond the unambiguous
product WITHOUT cue precision: when product and sum both self-verify they disagree
-> uniqueness refuses (safe-but-low-coverage by design). The lever remains cue
precision (the ADR-0175 learning loop).
Microscope finding: 0003-class flips (48*24*0.75=864=gold) are blocked by a
DECIMAL/currency grounding gap -- '$0.75' tokenizes to 0/75 so '0.75' is not
grounded by the shared round-trip primitive. Not a search bug; deferred
extraction-richness work (won't casually change the serving round-trip primitive).
A test documents the current refusal so the fix is detectable.
wrong=0: serving untouched (sealed); ambiguity + no-licensed-cue refuse; routes
through the proven gate. 8 MS-3 tests; full derivation surface 77/77; ruff clean;
smoke 67.
MS-2 of multi-step composition. Extends the derivation model so a chain mixes
text-quantity operands and COMPARATIVE-scalar operands (twice->x2, 'N times'->xN,
half->x0.5), self-verifying the whole chain with completeness over body+question
and question-target matching.
- model.py: Step gains comparative flag.
- comparatives.py: ComparativeScalar gains number_token (the '<N> times' number,
so completeness counts the consumed body quantity); comparative_step(cs) bridges
a scalar into a Step (operand grounded by cue, not a text value token).
- verify.py: self_verifies exempts comparative operands from value-grounding
(clause 1) — they are cue-grounded (clause 2); completeness (Counter) counts a
digit comparative's number_token as consuming the body quantity. Adds target_units
to select_self_verified: a chain whose answer_unit isn't the asked unit is dropped
(question-target match; empty target_units imposes no constraint).
Proves the multi-step shapes from the gold structures: 0024 (text sum then 'three
times' scale -> 438), 0033 father-chain (digit-comparative '7 times' + fixed 'half'
+ text add -> 47). Full 0033 DAG (quantity reuse + the question's 25) deferred.
25 MS-2 tests; full derivation surface 69/69 (3a/3b/comparatives/ms1/ms2); ruff
clean; smoke 67. Not wired into serving (model ready for MS-3 target-guided search).
MS-1 of multi-step composition. Turns the question into a Target = what the
problem asks for, the search's pruning signal + stopping criterion (MS-3).
Lexeme-level only (ADR-0165): the existing question parser returns nothing on
these GSM8K questions, and 0165 forbids new question-shape grammar regex. Three
robust signals:
- quantities: numbers stated IN the question (0033's 'when she is 25') via the
body's lexeme extractor — they participate in the derivation.
- aggregation: presence of an aggregation lexeme (total/altogether/combined/sum/
'in all'/'in total') — soft hint the final step is a sum.
- units: asked units resolved by INTERSECTION with the body's known units
(precise lexeme match, e.g. 'jumping'). Superordinates (weight<->pounds) are
NOT faked — deferred to a curated superordinate-units pack; until then the unit
signal is precise-but-incomplete and the search leans on completeness.
Refuse-preferring: empty target field is not an error, just a weaker prune.
generate/derivation/target.py: Target + extract_target(question, known_units=()).
12 MS-1 tests (question-quantity, aggregation, body-unit intersection,
superordinate-not-faked, determinism, frozen). Verified: derivation suite 57/57;
ruff clean; smoke 67. Not wired into serving (Target ready for MS-2/MS-3).
The curated, irreducible world-fact primitives multi-step composition needs
(ADR-0175 section 10: the engine can't derive 'twice = 2' from arithmetic). The
microscope flagged these via the 0015/0025/0024/0033 wrongs.
language_packs/data/en_core_comparatives_v1/: 9 closed-set multiplicative
comparatives (twice/double/triple/quadruple/half/quarter + inflections) -> scalar
ops. manifest.json with sha256 of the bytes on disk (CLAUDE.md pack rule).
Refusal-preferring: non-terminating/ambiguous comparatives (a third, several)
deliberately excluded; expansion via HITL corridor.
generate/derivation/comparatives.py: extract_comparative_scalars() ->
ComparativeScalar(op, scalar, span, cue). Fixed lexemes + the '<number> times'
pattern (digit or word-number via WORD_NUMBERS). Lexeme-level (ADR-0165);
deterministic (text-order); supplies only the SCALAR primitive — referent
binding is the multi-step search's job (ADR-0176).
14 tests incl. refusal-preferring discipline + pack integrity (manifest checksum
matches bytes on disk). Verified: derivation suite 45/45; ruff clean; smoke 67;
packs 141. Not wired into serving (data + extractor ready for ADR-0176 MS phases).
Self-verification strengthening (microscope-driven). The Phase 3b measurement
showed self-verification was necessary-but-not-sufficient: 9/13 self-verified
attempts were wrong. Inspecting them deterministically revealed most were
correct FIRST STEPS of multi-step problems that ignored numbers stated elsewhere.
Adds clause 5 to self_verifies: a derivation must account for every quantity the
problem states (problem quantities subset of used). Refuse-preferring: unused
quantities -> not self-verified. This catches the multi-step-incomplete attempts
the grounding/cue/unit clauses cannot (their operands ARE grounded).
Practice measurement: wrongs 9 -> 2 (4 correct / 2 wrong / 44 refused). The 2
survivors (0015, 0025) are COMPLETE but wrong due to missed WORD-quantities
('twice', 'her friends') -> the microscope points the next change at extraction.
Updated the disagreement test to use two complete derivations; added an
incomplete-refusal test. 32 tests pass; smoke green; serving untouched (sealed).
ADR-0175 Phase 3b — the first live attempt generator. Runs only in the sealed
practice lane, only on cases the engine refused; every proposal is gated by the
Phase 3a self-verification gate.
generate/derivation/:
- extract.py: extract_quantities() — lexeme-level (number + unit word; ADR-0165).
- search.py: search_multiplicative() — one in-clause product candidate per
sentence with >=2 quantities + a present multiplicative cue; gated by
select_self_verified. Per-sentence scope + multi-candidate disagreement give
the uniqueness gate real teeth (two qualifying sentences -> refuse). The cue
set {each,every,for,per,times} is an explicit PROVISIONAL hypothesis the
practice loop refines, not a claimed-correct grammar.
evals/gsm8k_math/practice/v1/search_runner.py: search_augmented_scorer +
build_search_report — base scorer, then a practice-only attempt on refusals.
MEASUREMENT (the deliverable, per the breadth-of-impact test):
practice with search: correct=4 wrong=9 refused=37 (baseline 3/0/47)
- Flips +1 (0021, the clean in-clause aggregate) and its renumbered/reworded
variants (ADR-0114a perturbation guard) -> a real capability, not memorisation.
- 9 wrong attempts -> elimination records (§9), the learning signal. The naive
full-product cue model over-attempts; the eliminations are exactly the signal
that refines it.
HONEST FINDING: self-verification (grounding ∧ cue ∧ unit ∧ uniqueness) is
NECESSARY but NOT SUFFICIENT — 9/13 self-verified attempts were wrong vs gold.
The gap is cue PRECISION / which-quantities-compose (the knowledge axis), not
'can we multiply' (skill). This is why the search runs sealed: gold catches the
9, and case 0050 (canary) attempted-and-failed IN PRACTICE without touching
serving -> validates the seal.
Invariants: #1 seal (serving still 3/47/0; 0050 refuses in serving; no
generate/chat import of the lane), #3 determinism. Serving wrong=0 untouched.
Verified: 3a+3b 31/31; ruff clean; serving lane 4/4; smoke 67/67.
ADR-0175 Phase 3 splits wrong=0-first: build the gate (3a) and PROVE invariant #2
before the bounded search (3b) that could exploit gaps.
generate/derivation/:
- model.py: Quantity / Step / GroundedDerivation. A derivation is a left-fold over
text-sourced quantities; each Step carries its licensing cue (the lexeme the
search claims licenses the op).
- verify.py: self_verifies() — grounded operands ∧ grounded operation cues ∧ unit
consistency ∧ no divide-by-zero. Grounding REUSES the canonical primitives from
math_roundtrip (_tokens/_token_in/_value_grounds) so the gate cannot drift from
the round-trip contract. select_self_verified() adds the uniqueness rule:
unique self-verifying answer resolves; zero or disagreeing refuse (wrong=0).
INVARIANT #2 proven (TestInvariant2_NoSpuriousSelfVerification): the gate refuses
to self-verify a derivation that is not grounded+unit-consistent+unique even when
its value coincides with gold — the 20/5==4 class:
- invented operand not in text -> refused
- operation cue not in text -> refused (division not licensed by any present cue)
- value coincidence (20/5=4) with ungrounded op -> still refused
- add across units (pounds + reps) -> refused
- divide-by-zero -> refused
Plus uniqueness: disagreeing grounded derivations -> refuse; agreeing -> resolve.
Phase 3a is inert (nothing wires generate.derivation into serving). 3b is the
bounded search that produces derivations for this gate + measures the flip-curve
in the practice lane under perturbation.
Verified: 16/16; ruff clean; smoke 67/67; no serving import.
ADR-0175 Phase 2 — a NEW lane (evals/gsm8k_math/practice/v1/), separate from the
wrong=0-pinned serving runner which is NOT modified. Runs the 50 cases in
practice mode: scores correct/wrong/refused as practice metrics, feeds per-class
counts into the Phase 1 ledger, diagnoses every refusal (§8), emits an
elimination record per wrong.
- classify_operation: gold-derived primary op class {multiplicative,divisive,
additive} from <<a*b=c>> calc annotations (Tier-1 checkable in practice).
- diagnose_refusal (§8): skill_gap / knowledge_gap / genuine_ambiguity router.
- EliminationRecord (§9): wrong attempt gold caught -> pruning signal.
- PracticeReport: counts + per-class ledger + diagnoses + eliminations; as_dict.
- run_practice(cases, scorer=...): injectable scorer for tests; defaults to the
candidate-graph scorer (read-only — never alters serving).
Live result mirrors serving (3 correct / 0 wrong / 47 refused of 50) because the
engine still refuses rather than guesses — attempts/eliminations go live in
Phase 3. But the diagnosis is already actionable: 35 skill_gap / 12 knowledge_gap
/ 0 genuine_ambiguity — 74% of refusals are skill gaps (Phase 3's search target),
quantifying the skill-vs-knowledge split.
Invariants: #1 seal (serving still 3/47/0; no generate/chat import of the lane),
#3 determinism (report byte-identical across runs). Elimination + wrong-tolerance
paths unit-tested via injected scorer (no live wrongs yet).
Verified: Phase 1+2 53/53, serving train_sample tests 4/4 (seal), smoke 67/67,
ruff clean.
ADR-0175 Phase 1 — standalone, deterministic, zero serving change. Nothing in
the serving/eval path imports it.
core/reliability_gate/:
- floor.py: conservative_floor(s,k) — pinned one-sided Wilson lower bound over
COMMITTED trials. z=2.576, N_MIN=10; range [0,1) (never exactly 1.0); float64
rounded half-to-even to 1e-9 for cross-backend replay. Perfect record reduces
to k/(k+z²) (earned by volume).
- ledger.py: ClassTally — immutable per-class counts; reliability = commitment
precision (refusals excluded so coverage never penalizes reliability);
t2_precision over the anchor set; coverage tracked separately.
- ceilings.py: Action{PRACTICE,PROPOSE,SERVE} + Ceilings — human-set θ
(practice=0, propose=.85, serve=.99). Frozen; with_override returns a NEW
instance (no in-place self-authorization).
- gate.py: license_for() — deterministic gate, measured/required≥1 (≡ measured≥
required; required=0 ⟹ always). Pure; never mutates/emits ceilings.
34 tests, each ADR invariant exercised by a test that fails under its violation:
#3 determinism/replay (idempotent, pre-rounded, deterministic decisions),
#4 no self-authorization (frozen ceilings; gate never emits/mutates them),
#1 proxy (zero serving coupling). Plus the §4a worked examples (38 clean
commitments clear propose; one wrong in 40 drops below; serve needs ~657).
Verified: 34/34 pass; architectural invariants 40/40; smoke 67/67; ruff clean;
no serving/eval import of the package.
The recognizer/candidate-graph path is the single canonical reader.
Retires the flag-gated incremental-reader dispatch that admitted 0/50 on
train_sample and only added a dead fall-through:
- remove _try_comprehension_reader, _try_reader_for_question, _tokenize_sentence
and both dispatch blocks from generate/math_candidate_graph.py
- delete generate/comprehension/lifecycle_runtime_adapter.py (402 LOC,
used only by the question-reader dispatch)
- drop the comprehension_reader_questions config flag and the parse_and_solve
/ _score_one_candidate_graph config threading
- remove the --use-reader runner plumbing + flag-ON/OFF delta report from
the train_sample runner; refresh report.json (drops stale use_reader field
and a stale refusal-reason; verdicts unchanged at 3/47/0)
- remove the now-dead use_reader field from teaching/coverage.py
CoverageReport + the core teaching coverage CLI flag
- delete tests/test_reader_coexistence.py (flag-ON/OFF premise dissolved);
fix 3 ADR-0174 build_report calls and 2 subprocess invocations
lifecycle.py and audit.py are KEPT — they are load-bearing for the ADR-0172
math-contemplation teaching corridor (audit_problem -> teaching/math_*),
which a pre-deletion trace surfaced. The parent ADR's plan to delete
lifecycle.py was wrong; only its GSM8K scoring dispatch was inert.
Net -1,038 LOC (code + tests). Behavior-preserving:
- train_sample 3/47/0, byte-identical verdicts to pre-5a baseline
- determinism holds; smoke/packs/runtime/cognition/teaching lanes green
- contemplation corridor + lifecycle/audit tests pass
Pre-existing (NOT introduced here; reproduce on base with changes stashed):
5 out-of-curated-lane stale committed-artifact / stale-assertion failures
(test_math_evidence_e2e, test_adr_0126_runner_wiring, G3/coverage_probe
report-match, test_refusal_taxonomy_lane rebuild).
ADR-0174 Phase 3b — emit N anchors for compound-clause discrete-count
sentences sharing one subject + one verb. Architectural substrate;
score on train_sample preserved at 3/47/0 (compound cases like 0027
admit past the recognizer-injection refusal but the rest of the
problem still has downstream complexity — fractions, percent — that
needs Phase 4 + solver work).
generate/comprehension/state.py:
HYPOTHESIS_CAP raised 4 → 8. Case 0040 emits 5 anchors; cap=8
gives headroom (7-item lists) without becoming permissive.
generate/recognizer_match.py:
_try_extract_compound_discrete_count_anchors() — new extractor
emitting tuple of anchors for compound sentences. Refusal-
preferring on:
- no conjunctive separator (single-anchor path)
- multiplicative/percent/fraction markers
- head verb not in whitelist
- any tail clause without grounded (count, observed_noun) pair
- exceeding HYPOTHESIS_CAP
- unaccounted digit in tail (wrong=0 hazard defense surfaced by
2026-05-28 implementation review: bogusnoun would silently fail
to produce anchor while leaving the digit unaccounted, admitting
partial state)
Wired into _match_discrete_count_statement dispatch as fallback when
single-anchor extraction fails.
tests/test_adr_0174_phase3b_compound_clause.py:
11 acceptance tests passing — pure conjunctive lists (proper-noun
+ pronoun-subject + single-actor antecedent), refusal-preferring
discipline (mixed-verb, multiplicative-tail, non-whitelisted-head,
partial-grounding all-or-nothing), HYPOTHESIS_CAP enforcement,
multi-actor pronoun defense preserved on compound, wrong=0 +
case-0050 canary.
tests/test_adr_0174_phase1_held_hypothesis_state.py:
Updated test_hypothesis_cap_is_four → test_hypothesis_cap_is_eight
with rationale for the raise.
Phase 3b implementation lookback review (per CLAUDE.md doctrine):
- Surfaced silent-partial-admission hazard in tail extraction;
fixed with digit-accounting check before commit
- Surfaced LATENT regex-path multi-actor pronoun hazard (not
introduced by Phase 3b; documented in test docstring with
cross-reference to project-adr-0174-multi-actor-pronoun-hazard
memory for follow-up)
- case 0040 ('He now has...') remains refused — 'now' adverb between
subject and verb defeats the existing canonical regex. Adverb-
stripping is separate scope (not Phase 3b).
Acceptance:
- 258/258 ADR-0174 + math_problem_graph tests pass
- Smoke 67/67, packs 141/141
- train_sample 3/47/0 preserved (wrong=0 held)
- Case 0027 'Malcolm has 240 followers on Instagram and 500 followers
on Facebook' now admits via the compound extractor — verified by
refusal moving to the next sentence (which has 'half' fraction)
All findings from the 2026-05-28 Phase 1-3a lookback review addressed
in one commit on the Phase 3a branch:
Wrong=0 hazard defense (the load-bearing fix):
- generate/math_candidate_graph.py: Phase 3a wiring now collects the
set of distinct proper-noun subjects seen in prior context. When
more than one exists, refuses with no_antecedent_ambiguous trace
event rather than guessing the most-recent (which was gender-blind
single-binding — wrong attribution in multi-actor problems).
- Refusals from the statement loop now preserve _statement_trace via
reader_trace in CandidateGraphResult (pre-existing latent issue:
Phase 2/3 trace events were dropped on early statement refusal).
- New tests assert: ambiguous case refuses with correct trace; single-
actor case still resolves normally.
Test coverage backfills (closes the 13 untested predicate-name gaps):
- TestCheckConstraintsInitialPredicateNames — 3 tests asserting the
exact predicate name on initial.value_grounds / initial.unit_grounds
/ initial.entity_grounds failure paths.
- TestCheckConstraintsOperationPredicateNames — 3 tests asserting
operation.verb_grounds / operation.value_grounds / operation.unit_grounds
failure-predicate-name parity.
- TestCheckConstraintsComposedInitialPath — 4 tests for the RAT-1
composed_initial path which was entirely untested in Phase 2
(parity manually verified during lookback review; now automated).
ADR amendment (honest doc vs impl drift):
- docs/decisions/ADR-0174-held-hypothesis-comprehension.md: appended
'Implementation Notes' section documenting:
- reevaluate signature differs from spec text (shipped is more
composable; treat as amended)
- Phase 2 wires per-candidate, not per-token (per-token is Phase 5)
- Lookback recompute is candidate-level, not token-level
- Hypothesis.constraint_state is never populated by Phase 2
- Multi-actor pronoun hazard defense rationale
- Honest LOC accounting: Phases 1-3a net +1,500 lines (Phase 5
delivers the projected net removal)
- Test coverage backfill summary
Cosmetic:
- lookback.py:297 unreachable raise — added # type: ignore[unreachable]
with comment explaining defensive future-proofing for Phase 3b.
Acceptance verified:
- 124/124 Phase 1+2+3a + reader tests pass (was 95/95 before backfills)
- Smoke 67/67, packs 141/141
- train_sample 3/47/0 preserved (wrong=0 invariant held)
- Multi-actor hazard live-tested: parse_and_solve refuses the
Alice/Bob/She case with no_antecedent_ambiguous trace event
See CLAUDE.md §Lookback Review Discipline and memory
feedback-lookback-review-discipline for the doctrine that surfaced
all of these issues at the right time.
ADR-0174 Phase 3a — substrate for held-hypothesis lookback.
Score unchanged at 3/47/0 (this PR is correctly-engineered
infrastructure; eval impact gated on ADR-0163.x recognizer expansion
documented in the follow-up brief).
Adds generate/comprehension/lookback.py:
- VALID_REFINEMENT_KINDS, VALID_UNRESOLVED_SLOTS — closed sets
contracted with reader_trace consumer
- PronounResolution refinement dataclass (pronoun + resolved_to +
evidence_source, all validated)
- Refinement Union (Phase 3b will widen with CompoundClauseExpansion)
- ReevaluateResult dataclass with admit/eliminate consistency
- reevaluate(hypothesis, refinement) operator — applies refinement,
re-runs check_constraints, returns refined Hypothesis or None.
- _rebuild_candidate_with_resolved_actor — rebuilds
CandidateOperation / CandidateInitial replacing the semantic actor
field (op.actor / initial.entity) while preserving matched_actor_token
/ matched_entity_token as the pronoun (so grounding still passes
against the held statement's source span).
Modifies generate/recognizer_match.py:
- _try_extract_discrete_count_anchor: pronoun-subject statements now
emit anchors with subject_role=<pronoun> + requires_pronoun_resolution
marker, rather than refusing at the _REFUSED_SUBJECT_TOKENS check.
The other narrowness layers (clause split, verb whitelist) still
refuse; only the pronoun layer changes.
Modifies generate/math_candidate_graph.py:
- After inject_from_match, when any parsed_anchor carries
requires_pronoun_resolution, the candidates are held as Hypothesis
objects with unresolved=('actor_pronoun',). The lookback path then
resolves via the existing _discourse_prior_subjects map and runs
PronounResolution refinements through reevaluate. Resolved
hypotheses flow into per_sentence_choices as if the regex parser
had produced them; unresolved hypotheses drop cleanly (refusal-
preferring). Emits 'lookback' JSON trace events with
outcome ∈ {admitted, eliminated, no_antecedent}.
Tests:
- tests/test_adr_0174_phase3_lookback.py — 17 acceptance tests
covering operator semantics on Operation/Initial, dataclass
invariants, closed-set constants, end-to-end wiring on synthetic
problems, and wrong=0 preservation on train_sample.
Phase 3.1 follow-up brief:
- docs/handoff/PHASE-3.1-FOLLOWUP-RECOGNIZER-EXPANSION.md documents
the empirical finding that the train_sample bottleneck is
verb-coverage (recognizer scope, ADR-0163.x) not lookback
(ADR-0174 scope). 11 verbs identified for HITL contemplation pass.
Recommends sequencing: Phase 3a now (substrate), ADR-0163.x verb
expansion next, Phase 3b after coverage matures.
Acceptance verified:
- 17/17 Phase 3a tests pass
- 95/95 existing tests pass (Phase 1 + Phase 2 + brief_11 + reader_phase2)
- Smoke 67/67, packs 141/141, lanes 8/8
- wrong=0 preserved, score unchanged 3/47/0 (intentional per brief)
Stacks on Phase 2 (PR #420). Rebases onto main after #416 + #420 land.
ADR-0174 Phase 2 — hoist _initial_admissible / roundtrip_admissible into
hypothesis-based constraint checks with structured elimination tracing.
Admission semantics are byte-equivalent to today; the change is structural.
Adds generate/comprehension/constraint_propagation.py:
- VALID_PREDICATE_NAMES: closed set of 17 sub-check names spanning
initial / composed_initial / operation admissibility predicates.
Adding new names requires an ADR amendment (structural contract with
reader_trace consumer).
- ConstraintResult dataclass: admitted bool + predicates_run trace +
elimination_reason. Validates admitted-vs-reason consistency.
- Elimination dataclass: confidence_rank + predicate + reason for one
hypothesis being eliminated. Serialisable as a reader_trace event.
- hypothesis_from_initial / hypothesis_from_operation: adapters wrapping
CandidateInitial / CandidateOperation as Phase-1 Hypothesis objects
with caller-supplied confidence_rank.
- _check_initial / _check_composed_initial / _check_operation:
decomposed sub-check implementations of the existing admissibility
predicates with first-failure short-circuit (matches current
semantics). Each sub-check populates predicates_run with (name, ok|
fail|skip) so the consumer sees exactly which predicate decided.
- check_constraints: dispatches on candidate type.
- eliminate_violating: bulk filter; returns (survivors, eliminations);
survivors are re-densified to satisfy ProblemReadingState's
open_hypotheses post_init invariant (dense-from-0 ranks);
eliminations carry the original confidence_rank for trace fidelity.
Wires into generate/math_candidate_graph.py at the recognizer
injection site (line 825+): replaces inline _initial_admissible /
roundtrip_admissible dispatch with eliminate_violating. Elimination
events become JSON entries in reader_trace with layer=
'constraint_propagation', phase=2, predicate, reason, sentence_index.
Phase 2 acceptance verified:
- 24/24 ADR-0174 Phase 2 tests pass (emission, parity with existing
predicates on 9 admit/reject cases, redensification, dataclass
invariants, integration).
- 71/71 existing reader + Phase 1 tests still pass.
- Smoke 67/67, packs 141/141, lanes 8/8.
- train_sample/v1 byte-identical across two runs with use_reader=True.
- Score preserved: correct=3 refused=47 wrong=0 — semantics identical
because the decomposed sub-checks short-circuit on the same predicates
the inline checks would have caught.
Trace-event behavior: today's injectors are conservative enough that
zero eliminations occur on train_sample/v1 (no false positives, no
mid-pipeline failures). The wiring is exercised by
test_phase2_event_shape_when_synthesized which proves the trace shape
on a synthetic CandidateInitial that fails initial.unit_grounds. When
Phase 3 begins emitting partial hypotheses from apply_word, the
elimination path will fire on real candidates and the trace will
populate.
Stacks on Phase 1 (feat/adr-0174-phase1-held-hypothesis-state, PR
#416). Merges cleanly into main after PR #416 lands.
MathProblemGraph.__post_init__ now raises MathGraphError when two
InitialPossession entries share the same (entity, unit) key but
declare different quantity values.
Pre-fix behavior surfaced by 2026-05-28 ADR-0174 Phase 3 post-merge
diagnostic: math_solver.solve() line 207 used last-write-wins dict
assignment when consolidating initial state. Two contradictory
inputs would silently overwrite without trace:
'Sam has 5 marbles. Sam has 3 marbles. How many marbles does Sam have?'
→ returned 3.0 (wrong=0 violation: definite answer from
contradictory input)
Post-fix: same input refuses with 'no branch produced a solvable
graph' — refusal-preferring discipline as wrong=0 doctrine requires.
Identical duplicates (same value) are admitted as redundant (no
contradiction). Different units for same actor admitted. Different
actors for same unit admitted. Single-value cases (the dominant
real-world pattern) unchanged.
This is an extraction-layer hazard discovered while investigating
Phase 3b scope: Phase 3b compound-clause held hypotheses would
emit multiple CandidateInitial entries per sentence, exercising
exactly this consolidation path. Fixing the silent overwrite NOW
ensures Phase 3b admission doesn't silently produce wrong answers.
Acceptance:
- 4 new tests in TestContradictoryInitialPossessionsRefuse
- 165/165 test_math_problem_graph tests pass (was 161/161)
- Smoke 67/67, packs 141/141 unchanged
- train_sample 3/47/0 unchanged (no real case exercised the
overwrite — but the hazard was latent)
References: CLAUDE.md §Lookback Review Discipline (the doctrine
that surfaced this), CLAUDE.md §Non-Negotiable Field Invariant
(make illegal states difficult to represent).
ADR-0174 Phase 1 — substrate only, no admission behavior change.
Adds to generate/comprehension/state.py:
- HYPOTHESIS_CAP (=4, structural assertion per ADR-0174 §Constraints)
- VALID_HYPOTHESIS_CONFIDENCE_RANKS (closed set, no probabilistic ranking)
- Hypothesis dataclass (frozen, slots) — candidate, category_assignments,
constraint_state, confidence_rank, unresolved. The 'candidate' field is
typed as object to avoid circular import on math_roundtrip /
math_candidate_graph candidate types; Phase 2 will pin canonical_bytes
contract over real candidates.
- UnknownHeld dataclass — token, position, narrowed_categories (frozenset).
Substrate for Phase 3 'hold instead of refuse' on unknown words; Phase 1
introduces only the type.
- ProblemReadingState.open_hypotheses + unknown_held fields, both default
to () (empty tuple). Defaults preserve today's single-committed behavior
exactly. Confidence-rank uniqueness + density-from-0 enforced at
__post_init__ as structural invariants.
- Canonical-bytes serializer extended to handle frozenset (sorted list).
Phase 1 acceptance verified:
- 29/29 ADR-0174 Phase 1 tests pass (construction, validation, cap
enforcement, canonical-bytes determinism, frozenset stability).
- 42/42 existing reader tests pass (test_brief_11_audit +
test_reader_phase2) — default-empty fields preserve byte-identity.
- Smoke 67/67, packs 141/141.
- train_sample/v1 byte-identical across two runs with use_reader=True.
- wrong=0 invariant held: 3/47/0 unchanged.
No apply_word body changes. The 'thread the hypothesis set' requirement
at Phase 1 is satisfied by field defaults that propagate through every
ProblemReadingState construction site in lifecycle.py without code edits.
Phase 2 (continuous constraint propagation) and Phase 3 (lookback
re-evaluation) will populate these fields with real hypothesis data and
wire the EMIT / ELIMINATE / HOLD operators.
_unit_grounds() previously refused multi-word units like 'Pokemon cards'
even when both component words appeared as tokens in the source span.
The function checked unit_token against the haystack as a single key,
but the tokenizer splits source into per-word tokens — 'Pokemon cards'
was never going to match.
Fix is conjunctive by design: every component word must appear in the
haystack. A missing component refuses, preserving wrong=0.
Truth-test: case 0023 (Nicole/Pokemon cards) previously refused with
'recognizer matched but produced no injection' on its first sentence.
After this fix, sentence 1 passes injection cleanly; the case now
refuses on sentence 2 (Cindy/Rex compositional clause) — a more
honest refusal reason that reflects the actual remaining gap.
Score unchanged at 3/47/0 (no overall lift; correctness win).
smoke 67/67, packs 141/141, lanes 8/8 all green.
Both INV-02/INV-21/INV-24 scan functions walked into .claude/worktrees/
and found vault recall/write callsites in the stale
step-3-submission-invariants checkout, producing 3 false failures.
Fix: add '.claude' to the os.walk exclusion set (INV-02) and to
EXCLUDED_DIRS (INV-21/INV-24). Defensive against any future worktree
that agents create under .claude/worktrees/.
Also pruned 58 stale worktree git-dir entries via git worktree prune
and removed the step-3-submission-invariants worktree directory.
Smoke suite: 67/67 passed.
C1: delete generate/math_versor_arithmetic.py and its 3 tests (ADR-0139
add-only arithmetic spike; no runtime consumers, no pipeline wiring,
follow-on lift paused per module docstring).
C3: gitignore engine_state runtime artifacts (manifest.json,
recognizers.jsonl, discovery_candidates.jsonl). Module code
(engine_state/__init__.py) remains tracked; generated checkpoint
files should not be.
C5: document reader zero-delta root cause in train_sample/v1/README.md.
Both Phase 2 (whole-problem) and Phase 1 (question-only) reader paths are
called but inert because all 47 refusals are statement-level NO_INJECTOR
gaps, not question-sentence gaps. Reader unblocks when injector coverage
expands (C2 work). report.json use_reader flag corrected to reflect last run.
C6: add deprecation header to generate/math_parser.py pointing at
generate.math_candidate_graph.parse_and_solve as the live path.
C2/C4 briefs: docs/handoff/CLEANUP-C2-run-lane-migration.md and
docs/handoff/CLEANUP-C4-compositions-compile.md added as operator
dispatch docs for the medium-scope wiring tasks.
Three review fixes:
1. Security: validate lane/split/version against ^[a-z0-9_]+$ before
building the runner module name. The runner_args list is passed to
subprocess.run without shell=True (no shell injection possible),
but defense-in-depth blocks arbitrary token characters from
reaching Python's -m module loader. Bad input now errors at the
CLI boundary with a clear message.
2. Bug-risk: _classify_refusal docstring referenced a
no_admissible_candidate bucket that the implementation never
emitted. Aligned docstring with actual buckets
(no_admissible_question / no_admissible_statement). Also made all
matching consistently case-insensitive (was mixed — some checks
used raw reason, one used .lower()).
3. Bug-risk: fetch_committed_baseline wrote to
.git/coverage_baseline_tmp.json. Replaced with tempfile.mkstemp in
the system temp dir — avoids (a) failures in non-git worktrees
where .git is a file pointer, (b) concurrent-access collisions
between simultaneous operators.
Tests (+3 new):
- test_classify_refusal_is_case_insensitive
- test_classify_docstring_matches_implementation_buckets
- test_fetch_committed_baseline_uses_system_temp
All 16 coverage tests green. Verified the validation:
core teaching coverage --lane 'evil; rm -rf /'
→ ERROR: lane='evil; rm -rf /' must match ^[a-z0-9_]+$
Brief D from PR #407. Closes the "flying blind on per-shape coverage"
gap identified in RAT-1's audit (finding 6).
After this PR, every operator can run a single command to see exactly
which refusal modes their work moved (or didn't), without re-eyeballing
report.json by hand.
Modules
-------
- teaching/coverage.py — pure aggregator:
- _classify_refusal — maps each per-case refusal reason to a
stable bucket (recognizer_empty_injection(<ShapeCategory>),
no_admissible_question, no_admissible_statement,
unexpected_question_count, other)
- build_coverage_report — reads a lane's report.json + emits a
CoverageReport with counts, refusal_taxonomy (sorted by count
desc), case_0050_verdict, optional delta vs baseline
- fetch_committed_baseline — uses `git show HEAD:<relpath>` to
pull the baseline report.json for delta computation
- core/cli.py:
- cmd_teaching_coverage — formats the report for terminal output
- core teaching coverage [--lane gsm8k_math] [--split train_sample]
[--version v1] [--use-reader] [--run] [--delta] [--json]
CLI output example
------------------
Lane: gsm8k_math/train_sample/v1 (use_reader=True)
Counts: correct=3 refused=47 wrong=0
Refusal taxonomy:
21 recognizer_empty_injection(discrete_count_statement)
6 no_admissible_statement
5 recognizer_empty_injection(multiplicative_aggregation)
4 no_admissible_question
4 recognizer_empty_injection(currency_amount)
3 recognizer_empty_injection(rate_with_currency)
2 recognizer_empty_injection(descriptive_setup_no_quantity)
2 recognizer_empty_injection(temporal_aggregation)
Wrong=0: ✓
Case 0050 hazard pin: refused ✓
Tests (13 new)
--------------
tests/test_teaching_coverage_cli.py — classification narrowness,
counts aggregation, case 0050 verdict capture, delta computation,
missing-baseline path, missing-report error, taxonomy sort order,
wrong=0 invariant visibility via as_dict.
Suite results
-------------
core test --suite teaching -q → 106 passed (93 → +13)
core test --suite runtime -q → 20 passed
core test --suite packs -q → 127 passed
core eval gsm8k_math --split public → 150/150, wrong=0
Note on Brief E (lexical auto-compile): the audit was WRONG. The
lexicon loader (generate/comprehension/lexicon.py::load_lexicon)
reads from the per-category source files directly; the compiled
lexicon.jsonl is only a manifest-checksum pin, not the source of
truth at runtime. apply_lexical_claim() writes a new entry → next
turn the loader sees it. Brief E is a non-issue; closing without a
code PR.
Verified by direct test: stage a clone of the math pack, write a
synthetic lemma to drain_token.jsonl, clear the lexicon cache, load
again → new entry present. So 3 of the 5 audit gaps closed (A, D,
E-as-correction); B and C remain as the next operator dispatch
targets.
Independent of PR #406 (RAT-1) and PR #408 (WAVE-A). Based on main.
Addresses 5 of 47 train_sample "recognizer matched but produced no
injection" refusals (the largest single failure-mode bucket
identified in RAT-1's audit).
Modules
-------
- generate/recognizer_match.py:
- _MULT_AGG_EACH_WEIGHING_RE — regex for "<Subject> <bake-verb>
<M> <outer-noun>, each <weigh-verb>ing <N> <unit>" pattern
- _try_extract_each_weighing_anchor — extracts M, N, subject,
inner unit; emits pre-composed CandidateInitial(value=M*N) with
composition_evidence so RAT-1's _composed_initial_admissible
gate verifies INPUT tokens ground (preserves wrong=0)
- _match_multiplicative_aggregation dispatches to the value
extractor when spec carries extract_values=True; specs without
that flag get the existing detection-only return path
(byte-identical legacy behavior)
- generate/recognizer_anchor_inject.py:
- inject_multiplicative_aggregation — new per-category injector;
narrow by anchor.kind so ME-3/ME-4 additive/subtractive anchors
(which share the same matcher entry point) continue to flow
through composition_registry consult instead of WAVE-A's direct
path
- registered in _INJECTORS dict (2nd entry after DCS)
- core/cli.py:
- seed-recognizer CLI gains --extract-values flag to opt the
canonical_pattern into the value-extracting matcher path
Seeded artifacts
----------------
- proposals.jsonl: rat1-seed-4dc30608fb783bc7 — multiplicative_
aggregation recognizer with anchor_kind=multiplicative_aggregate,
extract_values=True, observed_units covering ounces/strawberries/
questions/etc.
Live result on train_sample
---------------------------
- wrong == 0 preserved (3/47/0 baseline)
- Case 0050 hazard pin held
- public 150/150 preserved
- packs suite: 127 → 131 (+4 new WAVE-A tests, all green)
- teaching suite 93 unchanged
- runtime suite 20 unchanged
End-to-end synthetic solve (FIRST WAVE-A admission):
"Lilibeth fills 6 baskets where each basket holds 50 strawberries.
How many strawberries does Lilibeth have?" → answer=300
Cases that moved (statement now admits; refusal shifted downstream):
- Case 0025 (Lilibeth): statement admits via WAVE-A; refusal moved
to question parser ("If three of Lilibeth's friends pick the same
amount, how many strawberries do Lilibeth and her friends pick in
all?")
- Case 0047 (John bakes 12 macaroons): statement 1 admits; refusal
moved to statement 2
Eval correct count unchanged because the QUESTION parser (and
multi-statement cross-sentence reasoning) is the next bottleneck.
RAT-1's audit identified that gap; WAVE-A closes the injector half.
The remaining 3 multiplicative_aggregation refusals (0006, 0013,
0045) have different shape patterns the WAVE-A regex does not yet
cover; they're follow-up matcher extensions in the same architecture.
Tests
-----
- tests/test_wave_a_multiplicative_aggregation_injector.py (10
tests): each-weighing + each-basket-holds admission shapes,
detection-only path preserved when extract_values absent,
unobserved unit / pronoun / zero count refusals, end-to-end
inject_from_match dispatch, the Lilibeth canary solve,
wrong=0 preserved, case 0050 hazard pin
Stacks on PR #406 (RAT-1).
Adds surface_pattern, composition_category, and polarity to the
proposed_change_payload for composition_reclassification proposals so
operators can call apply_composition_claim() without field synthesis.
Dispatch by missing_operator:
- quantity_extraction → multiplicative_composition + bound(count) × bound(unit_cost)
- multi_quantity_composition → additive_composition + bound(qty_a) + bound(qty_b)
All other change kinds (matcher_extension, injector_sub_shape,
frame_reclassification) keep the existing evidence-aggregation payload.
Legacy fields (evidence_count, group_key, modal_sub_type) preserved.
Adds tests/test_contemplation_ratifiable_payload.py with 11 tests
including a round-trip from decompose_audit → apply_composition_claim.
The user's question — "shouldn't we be running it multiple times so
it can learn? or is that part broken?" — exposed that the math
teaching loop's `ratify → admit` closure had been structurally
broken at the connector between operator ratification and runtime
visibility. The handlers wrote source files (compositions/, frames/)
that the runtime loader never read because no compile step
regenerated the runtime artifacts.
This PR fixes the gap end-to-end AND fires the first live composition
admission on the canonical pack.
Modules
-------
- language_packs/compile_pack.py — unified compile step that
regenerates frames.jsonl + compositions.jsonl + updates
manifest.{frame,composition}_checksum atomically. Idempotent.
- teaching/math_composition_ratification.py — apply_composition_claim
now calls compile_pack at end of successful ratification. Closes
the source-file→runtime-artifact gap.
- teaching/math_frame_ratification.py — same auto-compile wire for
apply_frame_claim.
- generate/math_candidate_parser.py — CandidateInitial gains optional
composition_evidence Mapping field. When populated, signals the
candidate was produced by a registry-gated composition (ADR-0169);
the value/unit/entity are DERIVED arithmetic over grounded inputs.
- generate/math_candidate_graph.py — new _composed_initial_admissible
predicate that branches on composition_evidence. Wrong=0 preserved
by requiring each composition INPUT token (count, amount) to ground
in source_span literally; the derived value is admitted because the
arithmetic over grounded inputs is deterministic.
- generate/math_candidate_graph.py — discourse-level prior_subject
tracking: capture proper-noun subjects from ALL statement sentences
(including ADR-0136.S.0 context-filler sentences that get filtered
out before the candidate loop). Without this, "John adopts a dog"
(no numbers) is dropped and the cross-sentence subject resolver for
case 0019 sees prior_subject=None.
- generate/recognizer_match.py — all four composition matchers
(ME-1 currency-per-unit same-sentence, ME-2 cross-sentence, ME-3
additive, ME-4 subtractive) now populate composition_evidence in
CandidateInitial. Also added standalone " each " / " apiece " to
_PER_UNIT_TOKENS so currency_amount detection-only matcher refuses
per-item costs instead of swallowing them.
CLIs
----
- core teaching compile-pack — explicit operator surface for
regenerating runtime artifacts. JSON output for CI integration.
- core teaching seed-recognizer — operator surface for seeding a
RatifiedRecognizer entry in the proposal log for a given
(shape_category, anchor_kind). Writes created + transition(accepted)
events directly via ProposalLog._append.
Seeded artifacts (the actual loop closure)
------------------------------------------
- proposals.jsonl: new rat1-seed-48dd2673d6ad673d RatifiedRecognizer
entry for shape_category=rate_with_currency,
anchor_kind=currency_per_unit_composition.
- compositions/multiplicative_composition.jsonl: ratified
"bound(count) × bound(unit_cost)" affirms entry sourced from
case 0019 evidence.
- compositions.jsonl + manifest.composition_checksum: compiled
runtime artifact + manifest pin (RAT-1 auto-compile).
Live result on train_sample
---------------------------
- wrong == 0 preserved (3 correct / 47 refused / 0 wrong)
- Case 0050 hazard pin holds (refused)
- public split 150/150 preserved
- Case 0019 sentence 1 ("requires 3 vet appointments, which cost
$400 each") NOW ADMITS via composition. Previously refused with
"recognizer matched but produced no injection". The refusal moved
downstream to sentence 2 (a different currency_amount detection
bottleneck that is its own follow-up).
This is the first time a composition ratification on the canonical
pack actually reaches the runtime. The flywheel turned one
revolution.
Tests
-----
- tests/test_rat1_end_to_end_admission.py — 4 new live tests:
composition statement admits on isolated synthetic problem, case
0019 cross-sentence admission, wrong=0 preserved on train_sample,
case 0050 hazard pin.
- tests/test_consumption_empty_registry_no_op.py — refactored to use
isolated synthetic packs (the canonical pack may now carry ratified
entries).
- tests/test_math_{frame,composition}_ratification.py — updated
"manifest checksum unchanged" tests to "lexicon checksum
preserved" semantics: RAT-1 auto-compile may add the new optional
checksum fields; pre-existing lexicon checksum stays untouched.
Suite results: teaching 93, packs 131 (+4), runtime 20. All green.
Final PR of the matcher-extension wave. Ships:
1. tests/test_me5_all_categories_integration.py — 4 new tests:
- test_all_three_canaries_admit_through_full_pipeline: stages a
pack with all three SAFE_COMPOSITION_CATEGORIES entries +
ratifies, runs Maria/Sam/Tom canaries through matcher →
inject_from_match, asserts admission for all three
- test_partial_pack_only_admits_present_categories: refusal-
preferring when only one category is ratified
- test_all_safe_categories_have_extension_admission: pins that
SAFE_COMPOSITION_CATEGORIES is exactly the three covered
categories (breaks if future ADR widens without matcher)
- test_falsifies_uniformly_suppresses_across_categories:
polarity discipline holds across all three matchers
2. docs/handoff/ME1-ME5-MILESTONE.md — wave milestone doc:
- architecture diagram (audit → ratify → compile → load →
match → consult → admit)
- SAFE_COMPOSITION_CATEGORIES coverage matrix
- invariants preserved across the entire stack
- scope boundary (what does NOT fire yet — RAT-1 follow-up)
- recommended next dispatch
3. Test registration in core/cli.py packs suite.
Across the full ME-1..ME-5 stack:
- 5 stacked PRs (#400/#401/#402/#403/#404)
- 1 foundation PR (#398 — consumption wiring)
- 114 new tests, all green
- packs suite 127 passed
- core eval gsm8k_math --split public → 150/150, wrong=0
- All three SAFE_COMPOSITION_CATEGORIES have matcher extensions
Anti-regression invariants preserved across the entire stack:
- wrong == 0 on public split
- Case 0050 hazard pin (parametrized over all three categories)
- ADR-0166 — no new eval lanes
- ADR-0167 partition — no cognition imports
- ADR-0169 mutation boundary — registry is a gate, not arithmetic
- All matcher detection paths byte-identical
- engine_state/* never committed
- SAFE_COMPOSITION_CATEGORIES enforced at write AND load
- polarity falsifies honored uniformly
Live train_sample admission requires operator-seeded ratifications
(RAT-1 follow-up). Wiring is end-to-end correct, verified by ME-5
integration tests.
Memory: milestone-me1-me5-matcher-extensions-complete saved.
Stacks on PR #403 (base: feat/matcher-extension-subtractive).
