Commit graph

2 commits

Author SHA1 Message Date
Shay
6582df3bae feat(parser): ADR-0122 rate/per-unit grammar (substrate-only; lift deferred)
First parser-expansion ADR after ADR-0121's deferral. Adds the rate
algebra substrate (Rate dataclass + apply_rate operation kind + parser
pattern + solver/verifier/realizer + en_arithmetic_v1 pack lemma)
mirroring the deferral pattern that ADR-0121 demonstrated for
capability promotion: substrate complete, gate refuses honestly.

Substrate
- Rate(value, numerator_unit, denominator_unit) frozen dataclass with
  strict positive-value + non-empty-distinct-unit refusal at construction
- apply_rate operation kind admitted in VALID_OPERATION_KINDS;
  Operation.operand widened to Quantity | Rate with kind-discriminated
  type enforcement
- Parser: _RATE_COST_EACH_RE + _RATE_COST_EACH_TRAILING_RE +
  _Q_RATE_AGGREGATE_RE patterns; actor_units state tracking;
  first-declaration-wins on redeclaration (ParseError); orphan-rate
  refusal at end of parse; three refusal paths in rate-aggregate question
- Solver: _apply_rate() reads denominator-unit state, multiplies by
  rate.value, writes numerator-unit state (denom preserved)
- Verifier: _verify_apply_rate_step() byte-equal replay
- Realizer: 'At {N} {numer} per {denom_singular}, {actor} spends ...'
  template containing required tokens
- Pack: en-arith-006 apply_rate lemma + gloss; SHA-256 checksums
  refreshed; manifest version 1.0.0 -> 1.1.0; provenance tagged
  adr-0122:rate_extension:2026-05-22

Measurement (honest)
- Sealed GSM8K correct_rate: 0/1319 (substrate matches zero real cases
  alone). Multi-construction barrier documented in the ADR: all 14 sealed
  cases matching 'each \w+ costs?' combine rate with at least one other
  class (aggregation 6, comparison 3, unit conversion 2, multi-actor 2,
  conditional 1)
- Sealed GSM8K wrong: 0 (load-bearing positive claim; grammar adds zero
  misparses across 1,319 real test problems)
- Anti-overfit lanes unchanged: OOD ratio, perturbation invariance
  preserving/breaking 1.0, adversarial wrong 0
- ADR-0121 invariants byte-equal preserved (6/6)
- 41 new ADR-0122 invariants pinned in tests/test_adr_0122_rate_per_unit.py
- 670 existing math + pack regression tests pass

Roadmap update
- Per-ADR lift expectation corrected: no single parser-expansion ADR
  will move sealed correct_rate alone. First lift signal will come
  from cumulative composition after 3rd or 4th class lands (rate +
  comparison + aggregation foundational set)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-22 21:24:28 -07:00
Shay
7ad3f72cb4 feat: ADR-0118 — stepped realizer (SolutionTrace → show-your-work prose)
Phase 4 of the ADR-0114 GSM8K-math roadmap. Consumes a SolutionTrace
and emits one sentence per step plus setup + answer sentences. Pure
function; same trace → byte-equal RealizedTrace.

What landed

generate/math_realizer.py
  - realize(initial_state, trace) -> RealizedTrace
  - Frozen RealizedTrace dataclass with canonical_bytes() + as_prose()
  - Per-kind sentence rules (add / subtract / transfer / multiply×2 /
    multiply×3 / multiply-general / divide)
  - Singular/plural surface rule matches parser canonicalization
  - Typed RealizerError on unrecognized step kinds

tests/test_math_realizer.py — 60 cases pinning five invariants:
  1. All 50 dev-set cases realize without error
  2. Determinism: byte-equal RealizedTrace across two calls
  3. Setup sentence count == initial_state count
  4. Step sentence count == operation count
  5. Answer sentence contains the resolved value + unit

ADR-0114a obligation discharge update

ADR-0118 hardens determinism (#9) across a third layer (realizer)
and makes #3 / #10 human-inspectable via the prose surface. No
obligation is directly newly discharged by ADR-0118; it's substrate
for ADR-0119 GSM8K eval lane.

Round-trippability of the prose through the parser is explicitly
out of scope for this phase. The trace is the verifiable artifact
(ADR-0117); the prose is human-readable documentation.

Tests: 60 new realizer cases; 546 total green across realizer +
parser + solver + verifier + OOD; 67/67 smoke green.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-22 17:11:10 -07:00