core/docs/analysis/comprehension-coverage-wall-map-2026-06-14.md
Shay 04e0b0fbe9 docs(analysis): wall map — the recognizer/coverage wall is COMPOSITION
Evidence-based map of the comprehension coverage wall (the C decision). GSM8K
diagnostic: 4/0/46, no single barrier >10%, ~24 fragmented primary barriers with
heavy multi-barrier co-occurrence (compound+comparative+fraction+coreference stack
on the same cases). Capability lanes are narrow curated-gold conformance
(combined-rate: 6 solvable shapes, 19/19 oracle). Single-shape recognizers are
proven metric-inert (4th confirmation). Diagnosis: shape-recognizers don't COMPOSE
within compound statements / across coreference. Leverage: Tier 1 recognizer
additions ~+5 cases then cap; Tier 2 compositional reading is the only real lever
(architectural, must stay wrong=0). Recommendation: don't add isolated recognizers;
design a compositional reader.
2026-06-14 15:10:30 -07:00

90 lines
5 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# Wall map: the recognizer/coverage wall is COMPOSITION
**Date:** 2026-06-14
**Status:** diagnostic map (the real capability lever, per the C decision — point at
the recognizer/coverage wall). Evidence: committed refusal taxonomy + live lane runs
+ 3× prior empirical confirmation (ADR-0191/0192/0193 milestones).
**Bottom line:** more single-shape recognizers are **proven metric-inert.** The wall
is that shape-recognizers do not **compose** within a compound statement or across
statements with coreference. The capability gain lives in compositional reading,
not in widening any one shape.
## Evidence
### 1. GSM8K diagnostic — the composition signature
`train_sample` serving: **4 admitted / 0 wrong / 46 refused** (8% admission). The
curated refusal taxonomy (`refusal_taxonomy_v3.json`, 50 cases) shows **no single
barrier exceeds 5/50 (10%)** and the barriers are fragmented across ~24 primary
categories:
```
primary_barrier (top): compound_statement 5 · novel_initial_form 5 ·
novel_initial_verb 4 · fraction_operand 4 · conditional_question 3 ·
context_filler 3 · compound_comparative 3 · rate_price 2 · … (long 1-2 tail)
secondary_barriers (co-occurring): compound_comparative 5 · percentage_of 5 ·
fraction_operand 4 · rate_price 4 · multi_step_complex 4 · rate_comparative 4 ·
coreference_pronoun 3 · …
```
The secondaries are the tell: refused cases **stack 3-4 structures at once**
(a compound statement *with* a comparative *and* a fraction operand *and* a pronoun
coreference). The two raw refusal modes confirm it:
- `recognizer matched but produced no injection` (majority) — the category is
recognized (`discrete_count_statement`, `rate_with_currency`,
`multiplicative_aggregation`, …) but the injection layer can't structure the
*composed* sentence.
- `no admissible candidate` — the composed shape isn't recognized at all.
### 2. Capability lanes are narrow curated-gold conformance, not open coverage
- `evals/combined_rate_oracle`: **19/19 valid**, `by_expect = {solved: 6,
solver_refuses: 5, reader_refuses: 8}` — the combined-rate reader handles exactly
**6 solvable shapes**; everything else it correctly refuses. It is an *oracle*
(does the reader match the curated gold), not a coverage measure.
- `evals/comprehension/*`: per-domain conformance runners (propositional, syllogism,
set-membership, total-ordering, relational metric/predicate) — the flagship
deductive lanes, each 100%-conformant to a curated gold.
So each capability is a **shape-specific reader at 100% of a small gold.** Capability
= gold scope. The readers are individually correct and individually narrow.
### 3. Single-shape widening is metric-inert (already proven 3×)
ADR-0191/0192/0193 confirmed empirically that adding one operator/recognizer is
metric-inert on the real corpus — because the refused cases need *composition*, not
one more shape. This map's barrier data is the fourth confirmation.
## Diagnosis
The wall is **compositional reading**: the organ recognizes individual structures
(rate, fraction, comparative, count, percentage, coreference) but cannot **compose**
them — combine multiple recognized sub-structures within one compound clause, and
carry referents across statements. Real GSM8K problems are compositions; the readers
are a bank of isolated shape-recognizers. That is why admission caps at ~8% and why
every single-shape fix has been inert.
## Leverage (two tiers, honestly)
- **Tier 1 — near-term recognizer coverage (small, capped):** the pure
*single-barrier* misses (`novel_initial_form`/`novel_initial_verb` and a few
unrecognized rates/temporals like "Every week, he gets 6 cards", "Mark does a gig
every other day for 2 weeks") can be admitted by adding their recognizers. Honest
estimate: **~+5 cases** (4→~9/46), then it hits the composition wall hard, because
the remaining ~37 refused cases carry *multiple* co-occurring barriers that one
recognizer doesn't clear.
- **Tier 2 — compositional reading (the real lever):** a layer that **composes**
recognized sub-structures within a compound statement and **resolves coreference**
across statements, so a sentence that is rate+comparative+fraction+pronoun reads as
one composed structure. This is the only path past ~10/46, and it is an
architectural build (not a recognizer addition). It must preserve wrong=0 (a
composed reading still passes the self-verification/disagreement gate).
## Recommendation
**Do not pour effort into more isolated shape-recognizers — it is proven inert.**
The capability gain is Tier 2 (compositional reading). Tier 1 is a small honest
warm-up at best; I would not lead with it.
Tier 2 is make-or-break and the solution space is wide (how to compose, where the
composition layer sits relative to extract/clauses/compose, how coreference is
resolved deterministically, how composed readings stay wrong=0). That warrants a
**design effort** — multiple independent architecture attempts, judged and
synthesized — before any build. Proposed as the next step, on sign-off.