Adds `evals/gsm8k_math/practice/v1/cases.jsonl` — 150 GSM8K-style word
problems covering only additive/subtractive operations. All cases carry
`<<a+b=c>>` / `<<a-b=c>>` annotations; none contain `*` or `/`, so every
case classifies as `"additive"` under `classify_operation`.
Four difficulty bands:
0001–0030 single add (14 distinct units, 15 entity names)
0031–0060 single subtract
0061–0090 two same-direction operations
0091–0150 mixed add+subtract and multi-step (2–4 steps)
IDs are `gsm8k-practice-v1-NNNN`, deterministically ordered.
`train_sample/v1/cases.jsonl` and its pinned SHA are untouched.
`build_search_report` continues to run unchanged.
Adds `_PRACTICE_CASES_PATH` constant and `_load_practice_cases()` /
`build_practice_report()` to `practice/v1/runner.py` as additive
symbols; `build_report()` and all existing imports are preserved.
New practice case count: 150.
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.
* docs(handoff): parallel-work plan post-GB-3a (CP-1 / scale / EX-3 tracks)
Three disjoint-file tracks dispatchable in parallel with Claude's serial Gap-B
line; records the hard constraint that GB-3b/4/5 are serial on compose.py.
* docs(adr-0180): propose CRDT-sharded vault concurrency substrate
Drafts ADR-0180 (Proposed) for a Delta-CRDT sharded substrate to support
forthcoming multimodal ingestion (vision, kinematics) without serializing
on a global Vault lock.
§1.5 grounds the proof obligation against the existing single-threaded
Python ingest path (sensorium → ingest/gate → field → vault/store →
compute_trace_hash) and enumerates four pre-refactor test obligations
(T-1..T-4) that must be green on main before any change to
core-rs/src/vault.rs lands, per CLAUDE.md work-sequencing item 5.
§1.5.5 explicitly fences out: approximate recall (exact CGA recall is
non-negotiable), hidden background execution (Merge Kernel must be
explicitly mounted with telemetry), and MLX/UMA hardware optimization
(deferred to a follow-up ADR; CPU-only Rust path lands first).
Proposed only — no code changes to core-rs, sensorium, or field.
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).
#449 shipped an inert, duplicated, unvalidated pack off-brief; closed it but kept
the on-thesis idea (ground the nouns GSM8K problems use) with the real scope it
would need to be beneficial.
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).
The actual coverage lever (ADR-0177 established cue-precision is its gate+prune,
not the unlock). Gap B = which quantities group via which ops in what order.
The reframe that resolves the rich-search-vs-uniqueness tension: structure-from-
READING, not enumeration. The text encodes its own structure — every gold case
fits a sequential clause-by-clause read (0021 one clause 15x10x3; 0003 48 ->x24
->x0.75; 0024 sum ->x3; 0033 seq chain + branch). So Gap B is comprehension: read
the problem, build the derivation structure as you read, hold alternatives on
ambiguity, reevaluate on lookback. This IS the project's original 'word-to-word
with lookback + problem-solving throughout' articulation, and the synthesis of
reader (0174) + solver/gate (0176) + learning (0177) + packs.
Decision: a comprehension-guided sequential composer — clause-local bounded
sub-derivations combined across clauses via relational cues (per/each/of->multiply,
and-list->sum, comparative->scale, more/older->add, fewer/less->subtract), held-
hypothesis on ambiguity + eliminate/reevaluate (REPOINT ADR-0174's inert reader
substrate from parse to STRUCTURE — where it finally becomes load-bearing), scored
by the 0176 self-verification gate + uniqueness, guided by 0177 cue-precision.
Locality bounds the search; relational cues constrain the cross-clause op so
uniqueness can RESOLVE (not just refuse). Coverage rises only as far as the reading
constrains structure; irreducible ambiguity refuses (wrong=0).
wrong=0 obligations: per-step self-verification, irreducible-ambiguity-refuses,
completeness+uniqueness over the whole structure, no-spurious-structure, determinism,
sealed. Honest hard parts: relational-cue precision (co-dependent with 0177 +
data-starved), clause segmentation, referent binding, branch/DAG (0033 quantity
reuse, GB-5), scale. This is the comprehension core — largest remaining capability;
serving lift materialises here, incrementally. Sub-phases GB-1..GB-6.
The lever MS-1->MS-3 proved: learn which (cue->op) readings are reliable from
practice eliminations, closing ADR-0175's self-verification 'necessary-not-
sufficient' gap before Phase 5.
Mechanism: per-(cue, op, unit_shape) reliability ledger (reuses ADR-0175 ClassTally
+ conservative_floor), fed by gold-labelling search candidate chains. Three uses:
U1 self-verification TRUST (near-term value: makes the Phase 5 proposal gate honest;
cold ledger => refuse, no junk proposals), U2 search guidance, U3 disagreement
resolution (coverage lever, hard-gated: margin+theta, ties refuse).
Load-bearing honesty (the bottleneck): a pattern earns POSITIVE signal only from a
gold-MATCHING chain; current blunt shapes match gold for ~4/50 cases, so the ledger
is starved (all-blame) AND structure failures (Gap B) pollute cue->op credit (a
correct op penalised for a product-of-ALL structure error). Plus data starvation
(50 cases << N_min). So cue-precision is COUPLED to richer guided shapes (Gap B) +
scale; it co-evolves (Gap B supplies gold-matching candidates -> cue-precision earns
signal -> prunes Gap B's search). It is the TRUST substrate + pruning engine, NOT
the coverage unlock by itself.
Recommended sequencing: CP-1/CP-2 (mechanism + self-verification trust, near-term
correctness value) now; richer guided shapes (Gap B) next as the flip lever; scale
makes it compound. wrong=0 obligations: cold=>no regression, ties refuse, theta-gated
serving, credit-noise can't flip serving (floor+N_min+margin+ratification+gold-tether),
determinism. Sub-phases CP-1..CP-4.
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).
The dominant remaining lever for serving lift (79% need mul, median 3 steps;
single-step search + completeness flips only 0021). Grounded in gold step
structures: derivations are CHAINS with intermediate results as operands;
quantities come from body AND question (0033's '25'); several need comparatives
(half/N-times).
Decision: bounded, deterministic, TARGET-GUIDED multi-step grounded derivation
search, gated by ADR-0175's strengthened self-verification (grounding ∧ cue ∧
unit ∧ completeness) + uniqueness + a new question-target match. Sealed practice.
Two new ideas beyond 0175: question-targeting (turn the question into a target =
search-pruning + stopping criterion) and multi-step chaining (intermediates as
derived operands). Sub-phases MS-1..MS-5, wrong=0-first (gate/target before broad
search). Invariant #2 extended to chains. Honest hard part: search explosion +
uniqueness refuses most -> target-pruning + cue-guidance + depth-bound are the
tractability levers; low coverage initially; comparatives pack is a prerequisite;
serving lift still waits on 0175 Phase 5 ratification. Reuses solver +
question extraction + round-trip primitives.
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).
Solid: 4 invariants proven by failing-under-violation tests; 84 green; seal
verified. No live hazards (9 search wrongs are sealed eliminations).
Drift: (1) 3a gate is PARTIAL vs spec — round-trip + no-contradiction not wired;
3b's necessary-not-sufficient finding follows -> a self-verification-strengthening
phase must precede Phase 5. (2) class taxonomy: Phase 2 uses gold operation-class
not capability-axis. (3) minor defensive-branch test gaps (no risk).
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.
Lookback/cross-reference audit before any code. No hard blockers. Corrects 3
overclaims of reuse and records the audit:
- reliability ledger is NEW (calibration/ is a grid-search param tuner, not a ledger)
- 0174 eliminate/reevaluate/contemplate is reading-coupled -> 'repoint to solving'
needs generalization, not drop-in reuse
- teaching corridor evidence (MathReaderRefusalEvidence) is reader-refusal-coupled
-> solver-practice proposals need a new evidence type
Constraints recorded: wrong=0 pinned on serving lane by ~25 tests + train_sample
runner -> practice MUST be a separate lane (phasing updated). Alignments noted:
ADR-0165 (lexemes-not-grammar) fulfilled by thin-front-end/thick-solver split;
INV-07 governance = no-self-authorization pre-wired; MAX_TOTAL_BRANCHES precedent
for bounded deterministic search; seal exists.
Resolves Open Question #1. conservative_floor(s,k) = one-sided Wilson lower
bound over COMMITTED trials (k=correct+wrong; refusals excluded so coverage
never penalizes reliability). Constants: z=2.576 (single global pessimism
knob), N_min=10. Range [0,1) — never returns exactly 1.0. float64 rounded
half-to-even to 1e-9 for cross-backend replay. z (estimator skepticism) and
per-class theta (action's required reliability, human-set) are independent
dials; engine touches neither. Worked cost-to-clear table + asymmetry example
included.
Proposed ADR + session derivation doc capturing the 2026-05-28 design
discussion that took GSM8K Phase 5b from 'build another matcher' to a
self-calibrating problem solver.
Session doc (docs/sessions/SESSION-2026-05-28-...): the full journey —
problem (per-shape matchers can't compound; 79% need mul, 0% single-step),
dead-ends (brute-force spurious matches; 0021 is the only single-sentence
case and it's idiosyncratic), and the four pivots that converged on the
solution.
ADR-0175 (Proposed): the decision —
- two regimes: serving (wrong=0, unchanged) vs sealed practice
(attempt-and-eliminate; wrong is the learning signal)
- proof-carrying seal: practice never writes serving; ratification only
- deterministic attempt/refuse gate: reliability(C) / theta_required >= 1
(NOT RL; regimes collapse the reward side so only reliability is quantified)
- per-class calibration ledger of replayable COUNTS + conservative lower
bound; human-set theta ceilings raised only on evidence
- checkability ladder (gold > convergent self-verification > consistency-only),
privilege proportional to reversibility; provenance + gold tether against
correlated self-delusion
- diagnostic refusal routes skill vs knowledge vs ambiguity; three
compounding stores (vault/packs/pruning); self-proving acquisition narrows
human input without bypassing the gate
- five proof-obligation invariants (wrong=0 on serving, no spurious banking,
determinism, no self-authorization, retractability)
Supersedes the matcher-oriented ADR-0174 5b sub-phases; repoints the
0174 eliminate/reevaluate/contemplate substrate from reading to solving.
Open question: shape of conservative_floor + N_min.
Grounded in a ground-truth measurement of the 47 train_sample refusals
against GSM8K's own <<a*b=c>> calculator annotations:
- 37/47 (79%) use multiplication; 43/47 use mul-or-div; 0/47 single-step
- multiplication is the foundational general capability (breadth = the
anti-overfitting signal), but necessary-not-sufficient: no case flips
from an operation matcher alone
Solver already supports {add,subtract,transfer,multiply,divide,apply_rate,
compare_additive,compare_multiplicative}; the gap is the reader->injector->
Operation front-end (matcher extracts 0 anchors on real sentences).
Sub-phase sequence (biggest-chunk-first, measure-the-flip-gated):
- 5b.1 single-sentence multiplicative aggregate (cleanest proof, ~2-4)
- 5b.2 shallow 2-3 step composition (25/47, the real chunk)
- 5b.3 deep multi-step 4-7 (22/47) under held-hypothesis elimination
Generality guard: flipped cases must hold under ADR-0114a perturbation/OOD.
Explicitly NOT widening discrete_count injector (overfitting + wrong=0 hazard).
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).
Pre-implementation investigation (lookback discipline) found the original
Phase 5 text inverted against shipped code:
- math_parser.py already out of runtime + candidate-graph scoring path
- lifecycle.py admits 0/50 (inert parallel parser, not the reader to promote)
- correct>=25 is a semantic gate structural collapse cannot meet
Decision (Invert + split): recognizer/candidate-graph path is the canonical
reader; lifecycle.py is retired. Phase 5a = structural retirement (net -LOC,
3/47/0 byte-identical, wrong=0). Phase 5b = semantic narrowness removal (the
real lift, own sub-phases, per-layer wrong=0 obligations).
