Consolidating ratification of the GSM8K design of record. Ratify the built comprehension/derivation substrate, freeze the serving regex recognizer/ injector path to lexemes + refusal-only, pin Phase 5b execution to WIRING -> COMPOSITION -> LEXICON. - ADR-0207: new consolidating decision (Accepted, ratified 2026-06-03). Supersedes ADR-0163 §Phase B-E + ADR-0136 regex sentence-template prescriptions. Freeze + wrong=0 gates (22-case corpus + sealed 1,319). - ADR-0164/0165/0174/0178/0179: -> Accepted (ratified by ADR-0207, 2026-06-03). 0164 keeps its implementation clause (Phase 1+2 shipped; remainder per §5) so Accepted != fully built. - composition_validation/v1: 20 -> 22 cases (2nd R4/R5 positives, dataset-sourced golds), +contract invariants 6-7, +dataset-gold test. Baseline 4/18/0; 47 passed. - docs/analysis: extraction-richness audit (read-only) reconciling ADR-0179 to the tree (EX-1/2/4/5/6 landed; EX-3 deferred). Non-serving (evals/docs/tests only). train_sample 6/44/0 unchanged; no-ref <N> times hazard stays refused. GB3b/0136 untouched.
127 lines
7.4 KiB
Markdown
127 lines
7.4 KiB
Markdown
# ADR-0179 — Extraction Richness: feeding the comprehension composer real quantities
|
||
|
||
**Status:** Accepted (ratified by ADR-0207, 2026-06-03)
|
||
**Date:** 2026-05-28
|
||
**Author:** Shay
|
||
**Anchor:** [[thesis-decoding-not-generating]]
|
||
**Builds on / unblocks:**
|
||
- [ADR-0176](./ADR-0176-multistep-composition-question-targeting.md) (MS-3 product search), [ADR-0178](./ADR-0178-compositional-structure.md) (GB-1/GB-2 composer) — built but **starved** by thin extraction.
|
||
- [ADR-0177](./ADR-0177-cue-precision-learning.md) — richer extraction produces the gold-matching chains that give cue-precision its signal.
|
||
- Reuses `en_numerics_v1` (cardinal table) + `WORD_NUMBERS` + the existing currency/fraction/compound grounding in `generate/math_roundtrip._value_grounds` (ADR-0128 / 0131.G.3).
|
||
|
||
---
|
||
|
||
## Context — the wall every recent measurement hit
|
||
|
||
The structure machinery (MS-3 search, GB-1 clauses, GB-2 list-sum) is built,
|
||
deterministic, and wrong=0-safe — and flips almost nothing on *practice* for one
|
||
recurring reason: **extraction is too thin to feed it.** Concretely, with
|
||
`extract.py`'s current `(\d+(?:\.\d+)?)\s+([a-z]+)` (digit + single following word):
|
||
|
||
| Case | Gap | Effect |
|
||
|---|---|---|
|
||
| 0024 | `"36 on Tuesday"` → unit `"on"` | non-uniform units → GB-2 list-sum never fires |
|
||
| 0003 | `"$0.75"` → bare `0.75`; `_value_grounds("0.75")` fails (currency branch needs the `$`; digit-runs `0`/`75` not checked for a bare decimal) | correct product (864) refuses |
|
||
| 0024/0033 | `"three times"` → `three` not a digit | word-number quantity lost |
|
||
| 0016/0024 | `"jumping jacks"`, `"stop signs"` | multi-word unit truncated |
|
||
| 0033 | `"Rachel is 12."` → no following unit word | sentence-final quantity lost |
|
||
|
||
The capability is there; the inputs aren't. **Extraction richness is the
|
||
prerequisite the prior phases kept naming** — and the place coverage should finally
|
||
move (modulo cue precision for op-disambiguation, and scale).
|
||
|
||
## Two layers — and their very different risk
|
||
|
||
- **(A) The sealed derivation extractor** (`generate/derivation/extract.py`).
|
||
Enriching it is **safe**: it feeds only the (sealed, practice-only) derivation
|
||
search; over-extraction is caught downstream (the gate's completeness + grounding
|
||
+ uniqueness; refuse-preferring). The bulk of this ADR lives here.
|
||
- **(B) The shared grounding primitive** (`_value_grounds` / `_tokens` in
|
||
`math_roundtrip`). **wrong=0-sensitive**: the *serving* round-trip filter uses it.
|
||
Exactly **one** change belongs here — bare-decimal grounding — and it must be
|
||
proven to leave serving `3/47/0` byte-identical.
|
||
|
||
## Decision
|
||
|
||
Enrich extraction, lexeme-level (ADR-0165 — orthographic shapes, never grammar),
|
||
refuse-preferring, reusing existing numerics machinery:
|
||
|
||
1. **Word-number quantities** (extract.py) — `three`, `a hundred`, hyphenated
|
||
compounds → values, via `en_numerics_v1` / `WORD_NUMBERS` (already the grounding
|
||
table; extraction just *uses* it). Unblocks comparative/word operands.
|
||
2. **Currency + decimal extraction** (extract.py) — `$0.75` → `(0.75, "dollars")`
|
||
preserving the money form; **plus** the one shared-primitive fix: `_value_grounds`
|
||
grounds a **bare decimal** `"N.M"` when both digit-runs `N` and `M` appear as
|
||
tokens — symmetric with the existing `$N.NN` and `N/M` logic. wrong=0-gated.
|
||
3. **Multi-word units** (extract.py) — `jumping jacks`, `stop signs` (a bounded
|
||
unit-phrase: number + 1–2 following content words), so same-unit comparison works.
|
||
4. **List-unit inheritance** (extract.py) — in a list (`20 jumping jacks on Monday,
|
||
36 on Tuesday, …`), a quantity with no unit word of its own inherits the list's
|
||
head unit. This is what makes 0024 a same-unit list → GB-2 list-sum fires.
|
||
5. **Sentence-final / unit-less quantities** (extract.py) — `"Rachel is 12."` →
|
||
`(12, <inherited/empty unit>)`, so the quantity isn't lost.
|
||
|
||
Each is independently bounded and refuse-preferring; (4) and (1) are the highest-
|
||
leverage (they unblock 0024 and the comparatives).
|
||
|
||
## wrong=0 obligations (must be *proven*, not asserted)
|
||
|
||
1. **Shared-primitive change keeps serving byte-identical.** After the bare-decimal
|
||
grounding fix, serving stays `3/47/0` *byte-identical* (the `verify pinned lane
|
||
SHAs` gate). If a decimal case now grounds and shifts a serving verdict, evaluate
|
||
it explicitly; any `wrong > 0` shift reverts the change. (Grounding a true
|
||
token-run is a false-negative fix; the round-trip + disagreement + solver still
|
||
gate any new admission.)
|
||
2. **Over-extraction cannot break wrong=0.** Spurious/mis-attributed quantities flow
|
||
into the derivation gate, where completeness + grounding + uniqueness refuse
|
||
them. A test: an over-extracted quantity yields refusal, never a wrong commit.
|
||
3. **Determinism/replay**; extraction is deterministic and lexeme-level (no grammar
|
||
templates — ADR-0165).
|
||
4. **List-unit inheritance must not invent units** — a quantity only inherits a unit
|
||
present in its list head; otherwise it stays unit-less (refuse-preferring on
|
||
same-unit comparison).
|
||
|
||
## Honest payoff
|
||
|
||
This is the prerequisite, not a coverage promise by itself — but it is the one that
|
||
**unblocks built capability**: uniform units → GB-2 list-sum fires on 0024-class;
|
||
decimal grounding → MS-3 product flips 0003-class; word-numbers → comparative
|
||
operands resolve; and every newly-extracted gold-matching chain **feeds cue-precision
|
||
its signal** (ADR-0177's bottleneck). Expect the practice flip-curve to finally move
|
||
on the extraction-blocked cases — measured, not assumed; op-ambiguity (cue precision)
|
||
and volume (scale) still gate the rest.
|
||
|
||
## Sub-phases
|
||
|
||
- **EX-1 — word-number quantities** (extract.py; reuse `en_numerics_v1`/`WORD_NUMBERS`).
|
||
- **EX-2 — currency/decimal extraction** (extract.py) **+ bare-decimal grounding**
|
||
(shared primitive, wrong=0-gated: serving `3/47/0` byte-identical). The one
|
||
shared-primitive touch; ship it alone with the lane-SHA proof.
|
||
- **EX-3 — multi-word units** (extract.py).
|
||
- **EX-4 — list-unit inheritance** (extract.py) — unblocks 0024 same-unit list.
|
||
- **EX-5 — sentence-final / unit-less quantities** (extract.py).
|
||
- **EX-6 — measurement.** Re-run MS-3 + GB-2 on the enriched extraction; report the
|
||
flip-curve delta and the residual (cue-precision + scale) honestly.
|
||
|
||
## Acceptance criteria (Proposed → Accepted)
|
||
|
||
1. EX-2's shared-primitive change proven to keep serving `3/47/0` byte-identical
|
||
(lane-SHA gate); capability lanes G1–G5/S1 stay 100% `wrong=0`.
|
||
2. Over-extraction is refuse-preferring (proven: a mis-attributed quantity refuses,
|
||
never wrong-commits).
|
||
3. The measurement shows the previously extraction-blocked cases (0003, 0024-class)
|
||
reach a gold-matching chain (flip or honest hold under cue ambiguity), under
|
||
ADR-0114a perturbation; determinism/seal invariants hold.
|
||
|
||
## Cross-references
|
||
|
||
- **Unblocks:** [ADR-0176](./ADR-0176-multistep-composition-question-targeting.md),
|
||
[ADR-0178](./ADR-0178-compositional-structure.md); **feeds**
|
||
[ADR-0177](./ADR-0177-cue-precision-learning.md).
|
||
- **Reuses:** `en_numerics_v1`, `WORD_NUMBERS`, and the existing currency/fraction/
|
||
compound grounding in `_value_grounds` (ADR-0128, ADR-0131.G.3) — extends, never
|
||
reinvents.
|
||
- **Constraint:** [ADR-0165](./ADR-0165-regex-scope-rule.md) — lexeme/orthographic
|
||
extraction only, never grammar templates.
|
||
- **Thesis:** [[thesis-decoding-not-generating]] — read the quantities the text
|
||
actually states; don't pattern-match a shape.
|