ecc0072ea1
Three concrete cleanup items from the day's work, per the cleanup-as-you-find memory principle. ## 1. Remove inject_rate_with_currency stub PR #369 (A2 rate_with_currency) shipped a function that always returns () with an extensive docstring documenting the Rate-not-in-SentenceChoice schema gap. The function is dead at runtime — `_INJECTORS.get(category)` returning None has the same downstream behavior as the function returning (). The 16 tests pinned the empty-tuple return; the case-0050 hazard pin is duplicated in test_recognizer_skip_wrong_zero.py and test_brief_11b_step2_lexicon.py. The schema gap is now properly documented in ADR-0170 (PR #372). A dispatch-table comment at the removal site retains the at-code pointer to that ADR for anyone wiring a new injector. Removed: - `inject_rate_with_currency` function in generate/recognizer_anchor_inject.py - Its `_INJECTORS` dispatch table entry - Its `__all__` export - tests/test_injector_rate_with_currency.py (371 lines, 16 tests) ## 2. Remove docs/handoff/GPT55-MOBILE-DISPATCH.md Single-session travel-time scaffolding. The 5 tasks it named are complete or superseded by ADR-0170's findings. Pure historical artifact. ## 3. Remove docs/handoff/WAVE-NEXT-INJECTORS.md Superseded by docs/handoff/WAVE-NEXT-REVISED.md, which captures everything load-bearing from the original brief in its A1–A4 findings table. The "kept for history" justification didn't survive scrutiny: the document was misframed (over-promised lift; misframed schema work as injector work). Lessons captured in REVISED + ADR-0170. Updated cross-references: - WAVE-NEXT-REVISED.md: removed the "supersedes ... kept for history" pointer; tightened cross-reference list - ADR-0167-FOLLOWUPS.md §7: rewrote pointer to name ADR-0170 + REVISED as the live plan rather than "the original is retained" ## Test plan - 219 tests passed across G.2/G.4/G.5/S1/Brief 11/B1/B11A/wiring/partition/DCS-D.2 - evals/gsm8k_math/train_sample/v1/report.json untouched (regen surfaces a separate stale-baseline test issue — out of cleanup scope) - No runtime behavior change ## Net impact - 5 files removed (~1200 lines) - 1 file modified for explanatory comment (~30 lines) - 2 doc files updated to remove dangling cross-references - 0 behavioral change
188 lines
8.4 KiB
Markdown
188 lines
8.4 KiB
Markdown
# Wave-Next Revised — DCS Sub-Shapes + Schema-Gap Backlog
|
||
|
||
**Date:** 2026-05-27
|
||
**Supersedes:** the original Wave-Next injector briefs (removed in
|
||
cleanup PR; the four A-findings and pivot rationale are captured here)
|
||
**Why revised:** The A1–A4 dispatch surfaced findings that invalidate three
|
||
of the four briefs' lift assumptions. This doc replaces them with the
|
||
actually-tractable next steps.
|
||
|
||
---
|
||
|
||
## What A1–A4 actually found
|
||
|
||
| Brief | Outcome | Schema gap surfaced |
|
||
|---|---|---|
|
||
| **A1** `currency_amount` | Sandbox-blocked write; analysis only. **Real lift potential** — new verbs (`charges`, `earns`) outside `_INITIAL_HAS_RE` coverage. Implementation ready, design intact. | None |
|
||
| **A2** `rate_with_currency` | PR #369 opened (schema-refusal). Lift=0. | `Rate` not in `SentenceChoice = Union[CandidateInitial, CandidateOperation]` |
|
||
| **A3** `multiplicative_aggregation` | Sandbox-blocked write; analysis only. Lift=0. | Brief's `CandidateOperation(multiply)` spec is wrong: would compute `0 * inner = 0` with no prior `InitialPossession`. Correct emission is `CandidateInitial(outer × inner)` pre-computed product. |
|
||
| **A4** `temporal_aggregation` | Sandbox-blocked write; analysis only. Lift=0. | Needs `apply_rate` primitive that doesn't exist in the algebra. |
|
||
|
||
**Three of four categories have schema-level prerequisites before any
|
||
injector can ship.** The original Wave-Next brief misframed them as
|
||
injector work. The honest read: they're each their own ADR-sized
|
||
schema-extension followed by an injector wave.
|
||
|
||
---
|
||
|
||
## The actually-tractable next wave: DCS sub-shapes
|
||
|
||
The `discrete_count_statement` category is 21 of 47 GSM8K refusals
|
||
(largest single bucket) and has a **working v1 injector** from #315
|
||
plus a **specification document on main** at
|
||
`docs/handoff/discrete_count_statement-injector-spec.md` (from #366).
|
||
|
||
The spec recommends *narrow incremental expansion* — one sub-shape
|
||
per PR, each carrying its own `wrong=0` hazard pins. That maps cleanly
|
||
to small focused PRs.
|
||
|
||
### Proposed sub-shape sequence
|
||
|
||
Order by lift-per-risk:
|
||
|
||
1. **DCS-S1** — proper-noun possession with single static count, no
|
||
clause split (the canonical narrow form already partially handled
|
||
by the v1 injector). Expand to cases the v1 injector currently
|
||
misses for surface-shape reasons.
|
||
2. **DCS-S2** — proper-noun possession with `and`-coordinated counted
|
||
nouns (e.g. "Francine has 5 boxes and 27 loose crayons"). Requires
|
||
the multi-quantity composition decision but is structurally simpler
|
||
than the full multi-quantity recognizer.
|
||
3. **DCS-S3** — pronoun-subject possession (e.g. "She had 6 baskets").
|
||
Requires anaphora resolution to a prior proper noun. Higher risk
|
||
surface than S1/S2.
|
||
4. **DCS-S4** — subject-anonymous possession ("There were 12 apples").
|
||
Requires anonymous-subject handling decision.
|
||
|
||
Each sub-shape PR carries:
|
||
- A focused match-tightening + injector extension in
|
||
`generate/recognizer_match.py` and `generate/recognizer_anchor_inject.py`
|
||
- A test file pinning the new admission cases AND the
|
||
wrong=0 hazard (case 0050 + any newly-revealed hazard pattern)
|
||
- Eval delta section in `audit_brief_11.md`
|
||
|
||
The full DCS bucket lift potential is ~5–15 cases depending on
|
||
sub-shape coverage. Each sub-shape PR is a small, focused increment.
|
||
|
||
### Why not dispatch this as a Wave
|
||
|
||
The A1–A4 dispatch demonstrated that orchestrator-launched subagents
|
||
**burn 50k+ tokens per cold-start** with mixed PR-opening success
|
||
rates. The user dispatches operators in their own UIs/CLIs where
|
||
context costs amortize across the operator's session.
|
||
|
||
This document is the **reference brief** that operators read; the
|
||
**dispatch is the user's action**, not the orchestrator's. See
|
||
`memory/feedback-no-self-dispatch-of-subagents.md` for the binding
|
||
principle.
|
||
|
||
---
|
||
|
||
## Schema-gap backlog (file for separate ADRs, not Wave-Next)
|
||
|
||
The three findings from A2/A3/A4 deserve named ADRs when (and only when)
|
||
the prerequisite ADRs are in place. Until then they sit as queued items.
|
||
|
||
### Schema-Gap 1 — `Rate` in `SentenceChoice` union (from A2)
|
||
|
||
**From:** PR #369 (A2 rate_with_currency injector — refusal-only)
|
||
**Requires:** ADR-0168.x or new ADR-0170.
|
||
|
||
A2's concrete 4-step follow-up (preserved verbatim from the PR):
|
||
|
||
1. Add `CandidateRate` dataclass in `generate/math_candidate_parser.py`
|
||
(sibling of `CandidateInitial`/`CandidateOperation`) — carries a
|
||
`Rate` operand keyed by actor + source-span provenance.
|
||
2. Widen `SentenceChoice` to include `CandidateRate`; update
|
||
`_slot_count`, `_collapse_per_sentence_ties`, admissibility
|
||
predicates, and the per-sentence admission gate.
|
||
3. Teach `parse_and_solve` to compose `CandidateRate` with downstream
|
||
`apply_rate`/`multiply` questions — unify with the existing
|
||
`extract_earnings_candidates` short-circuit.
|
||
4. Then `inject_rate_with_currency` can emit `CandidateRate`. The
|
||
matcher already extracts `(currency_symbol, amount, per_unit)`;
|
||
needs entity extraction added.
|
||
|
||
**Lift potential after schema extension:** 3+ cases (the original
|
||
`rate_with_currency` bucket). Likely 4-5 more from `rate_*` adjacent
|
||
categories once the substrate exists.
|
||
|
||
### Schema-Gap 2 — `apply_rate` primitive in the algebra (from A4)
|
||
|
||
**From:** A4 temporal_aggregation analysis (no PR)
|
||
**Requires:** an algebra extension ADR.
|
||
|
||
The current solver has no `apply_rate` operation in its operator
|
||
inventory. Temporal aggregation ("5 hours per day, 7 days per week")
|
||
fundamentally needs this: a rate composed with a time-duration
|
||
quantity yields a total. Without it, temporal aggregation has nowhere
|
||
to ground.
|
||
|
||
**Lift potential after primitive lands:** 2 cases (the
|
||
`temporal_aggregation` bucket).
|
||
|
||
### Schema-Gap 3 — Multi-quantity composition emission shape (from A3)
|
||
|
||
**From:** A3 multiplicative_aggregation analysis (no PR)
|
||
**Requires:** correction to the original Wave-Next brief, plus a
|
||
multi-quantity-composition ADR coordinated with FOLLOWUPS §1
|
||
(CompositionClaim).
|
||
|
||
The brief's `CandidateOperation(multiply)` spec is mechanically wrong:
|
||
the solver's multiply operation does `state[(actor, unit)] *= value`,
|
||
requiring a prior `InitialPossession`. A standalone multiplicative
|
||
statement doesn't have that, so multiply would compute `0 * inner = 0`.
|
||
|
||
**Correct emission:** `CandidateInitial(outer_count × inner_count,
|
||
inner_unit)` — the pre-computed product as a possession.
|
||
|
||
**Lift potential after correction lands:** 0 cases (zero GSM8K
|
||
`multiplicative_aggregation` cases match the canonical narrow form
|
||
even with correct emission — they all need either coordination or
|
||
clause-split support). Effectively this category's bucket is empty
|
||
relative to current GSM8K; the schema correction is for future
|
||
problems, not for visible lift today.
|
||
|
||
### Schema-Gap 4 — A1's `currency_amount` injector (preserved, not dropped)
|
||
|
||
**From:** A1 currency_amount analysis (no PR, sandbox-blocked)
|
||
**Requires:** Only the sandbox fix + reimplementation. No schema work.
|
||
|
||
**Lift potential:** 2–4 cases (new verbs `charges`, `earns` not
|
||
covered by the existing `_INITIAL_HAS_RE`). The implementation design
|
||
is in A1's analysis — see agent transcript for details. This is the
|
||
smallest, lowest-risk follow-up if anyone picks it up; the brief is
|
||
sound, only execution failed.
|
||
|
||
---
|
||
|
||
## Standing pivot
|
||
|
||
The next operator pickup should be:
|
||
|
||
1. **DCS-S1** — narrow proper-noun possession sub-shape expansion (highest lift-per-risk, smallest scope)
|
||
2. **DCS-S2** — coordinated counted-nouns (medium lift)
|
||
3. **A1 reimplementation** (if sandbox config is fixed) — currency_amount with new verbs
|
||
4. **Schema-Gap 1 ADR** (Rate-in-union) — unblocks 3–8 follow-up cases
|
||
5. **Schema-Gap 2 ADR** (apply_rate) — unblocks 2 cases
|
||
6. **Schema-Gap 3** — folds into CompositionClaim ADR (FOLLOWUPS §1)
|
||
|
||
No timelines. Order is by leverage, not calendar.
|
||
|
||
---
|
||
|
||
## What this document does NOT do
|
||
|
||
- It does not dispatch any agents (per `feedback-no-self-dispatch-of-subagents`)
|
||
- It does not modify any runtime code
|
||
- It does not add new eval lanes (ADR-0166)
|
||
- It does not propose any non-deterministic mechanism
|
||
|
||
## Cross-references
|
||
|
||
- `docs/handoff/discrete_count_statement-injector-spec.md` — the DCS sub-shape spec
|
||
- `docs/handoff/ADR-0167-FOLLOWUPS.md` — parent follow-up queue
|
||
- `docs/decisions/ADR-0168-frameclaim-ratification.md` — FrameClaim scoping
|
||
- `docs/decisions/ADR-0168.1-math-frameclaim-proposal-adapter.md` — adapter bridge
|
||
- PR #369 — A2's schema-refusal artifact
|
||
- `memory/feedback-no-self-dispatch-of-subagents.md` — binding dispatch principle
|