core/docs/decisions/ADR-0178-compositional-structure.md

ADR-0178 — Compositional Structure: Comprehension-Guided Multi-Step Derivation (Gap B)

Status: Proposed Date: 2026-05-28 Author: Shay Anchor: thesis-decoding-not-generating Builds on / synthesises:

• ADR-0174 — Held-Hypothesis Comprehension — its held-hypothesis / eliminate_violating / reevaluate / contemplate substrate is repointed here from reading-to-parse to reading-to-structure (where it finally becomes load-bearing).
• ADR-0176 — Multi-Step Composition — the chain model + the self-verification gate that scores each structure.
• ADR-0177 — Cue-Precision Learning — the learned (cue→op) reliabilities that guide the structural search and that this search feeds.
• The comparatives pack (and a future relational-cue / superordinate-units pack) — the irreducible relational world-facts.

---

Context — the actual flip lever, and what it really is

MS-3 proved the multi-step search is built and wrong=0-safe but low-coverage: blunt whole-problem shapes (product-of-all / sum-of-all) flip only the one case where the shape happens to match (0021), and a rich unguided search drowns in disagreement-refusals. The dominant failure (0019, 0041) is compositional structure — which quantities group, via which ops, in what order. That is Gap B, and ADR-0177 established it is the real coverage lever (cue-precision is its safety gate and prune, not the unlock).

The central tension: a richer search produces more self-verifying candidate structures that disagree, so uniqueness refuses more. Coverage cannot come from enumerating more structures — it must come from steering to the right one.

The reframe that resolves it — structure-from-reading, not enumeration. The text already encodes its own compositional structure: clauses group local quantities, and relational cues across clauses say how the pieces combine. Every gold structure fits a sequential, clause-by-clause read:

Case	Read sequentially
0021	one clause: 15 × 10 × 3
0003	48 → ×24 (24 in each box) → ×0.75 (sell for $0.75 each) = 864
0024	20+36+40+50 (the and-list) → ×3 (three times as many) = 438
0033	12 → ×7 (7 times) → ÷2 (half) → +5 (5 older) → …

So Gap B is, at bottom, comprehension: read the problem, build the derivation structure as you read, holding alternatives when a clause is ambiguous and revising on lookback when a later clause disambiguates. This is precisely the original project articulation — "articulate word to word, with lookback re-evaluation, with problem-solving throughout, without insane complexity." It is also the synthesis of every prior strand: the reader (ADR-0174) supplies structure, the solver/ gate (ADR-0176) scores it, the learning (ADR-0177) guides it, the packs supply the relational facts.

Decision

A comprehension-guided sequential composer: read the problem clause-by-clause; within a clause, derive a local sub-result by a small, bounded search over that clause's few quantities; across clauses, combine the running result with the next clause via the relational cue the text presents (per/each/of → multiply; same- unit and-list → sum; comparative → scale; more/older/then → add; fewer/less/lost → subtract; fraction-of → scalar). When a clause's relation is ambiguous, hold multiple candidate structures and eliminate them downstream (repointed ADR-0174 machinery); when a later clause disambiguates an earlier choice, reevaluate (repointed). Every structure is scored by the ADR-0176 self-verification gate (grounding ∧ cue ∧ unit ∧ completeness) + uniqueness; the relational cue→op choices are guided/ranked by ADR-0177 cue-precision.

Why this gets coverage without breaking wrong=0: locality bounds the search (each clause has few quantities → no flat explosion), and the text's relational cues constrain the cross-clause op (so the surviving structure set is small enough that uniqueness can resolve, not just refuse). Where the text is genuinely ambiguous and no constraint or learned cue-precision resolves it, the held- hypothesis set stays >1 and the composer refuses — refuse-preferring, unchanged. Coverage rises exactly as far as the reading constrains the structure; no further.

Repointing ADR-0174 (where the inert reader becomes load-bearing)

ADR-0174's held-hypothesis substrate was built for reading-to-parse and is inert (0/50). Gap B is its true home: ProblemReadingState.open_hypotheses becomes the set of candidate derivation structures; eliminate_violating prunes structures a later clause contradicts; reevaluate revises an earlier structural commitment; contemplate consults vault/packs when a relation is unknown. These are reading- coupled today (per ADR-0175's audit) and must be generalised off their parse- specific types — not a drop-in, but the conceptual machinery is exactly right.

wrong=0 obligations (must be proven, not asserted)

1. Per-step self-verification. Each clause-local and each cross-clause combination is grounded + unit-consistent before it joins the structure (the ADR-0176 gate, applied per step). A test fails if an ungrounded/unit-incoherent step is admitted.
2. Irreducible ambiguity refuses. When the held structure set stays >1 with no eliminating constraint or decisive cue-precision, the composer refuses (does not pick). A test fails if it resolves an unresolved hold.
3. Completeness + uniqueness (ADR-0176) over the whole structure: all body+ question quantities used; a single distinct verified answer or refuse.
4. No spurious structure. A structure whose grouping is not licensed by the reading (relational cue absent/ungrounded) cannot self-verify even if its value matches gold (extends invariant #2 to trees/chains).
5. Determinism/replay; sealed (no serving import; serving stays 3/47/0 until Phase 5 ratifies).

The honest hard parts

• Relational-cue precision — the cross-clause op hinges on relational lexemes (per/each/of, times/half, more/older, fewer/less, then). Their op-mapping is partly packs (comparatives) and partly learned (ADR-0177 cue-precision, which is itself bottlenecked + data-starved). This is the same Gap-A/Gap-B co-dependency: Gap B produces gold-matching chains that give cue-precision its signal, and cue-precision ranks Gap B's structural choices. They co-evolve.
• Clause/clause segmentation — deciding clause boundaries is itself a parse decision; start with sentence + simple conjunction splits, refine cautiously (ADR-0165: orthographic/lexeme splits, not grammar templates).
• Referent binding — whose value a clause refers to ("Brooke does 3× as many as Sidney" → scale Sidney's result). Cross-sentence entity tracking (ME-2 / the discourse-subject machinery already exists) must say which running value a relation applies to.
• Branch/DAG structures (the hard end) — 0033 reuses 12 in two branches (12×7… and 25−12); a pure sequential chain cannot express quantity reuse. DAGs are GB-5, after chains work.
• Data/scale — coverage and the cue-precision it feeds compound only with volume (ADR-0163 §Phase F). On 50 cases this is mechanism + incremental flips, not a finished solver.

This is the comprehension core — the largest remaining capability and the place serving lift actually materialises. It is multi-phase by nature; each phase is gated wrong=0-first and measured honestly.

Sub-phases

• GB-1 — clause segmentation + clause-local sub-derivation. Split the problem into clauses (lexeme/sentence-level); within a clause, a small bounded search over its quantities → a local sub-result (self-verified). Tests: locality bounds the search; a clause-local product/sum self-verifies; ambiguous clause holds/refuses.
• GB-2 — sequential combination (chains). Combine running result with the next clause via the relational cue → op (guided by ADR-0177 + comparatives). Flips the sequential-chain gold cases (0003/0024-class) under self-verification + uniqueness.
  • Shipped narrower than scoped (drift G1, audit AUDIT-ADR-0178-GB1-GB2.md). What landed is a same-unit list-sum-then-scale slice (compose_sequential) that re-extracted from the whole problem — it did not consume GB-1's ClauseResults and did not chain across clauses, so it never flipped real 0003/0024 (those stay blocked on extraction + cross-clause chaining). Treat the landed code as GB-2a (list structure), not the full sequential chainer.
• GB-3 — lookback / reevaluate (repoint ADR-0174 reevaluate): a later clause revises an earlier structural choice. wrong=0-first.
  • GB-3a — referent guard (landed). The lookback review proved GB-2a's whole-problem aggregation admitted wrong structures across referents/scopes (hazards H1/H2/H3 returned 12/20/13). compose_sequential is now clause-scoped: the list-sum must be licensed within one clause, and a comparative outside that clause refuses. This is the wrong=0-first floor (compose_sequential is sealed substrate, not yet wired to a scorer, so serving stayed 3/47/0); tests in tests/test_adr_0178_gb3_referent_guard.py would fail against the pre-guard code.
  • GB-3b — referent-aware accumulation chaining (scoped). Consume GB-1 ClauseResults and combine across clauses referent-safely: the first reading is single-referent accumulation (X has N. He buys M more. → N ± M), licensed by a closed change-verb cue set + a minimal "no new actor" referent guard (generalises GB-3a: chain-when-licensed rather than refuse-all-multi-clause). Measured target: the 46 sum-of-all == gold additive practice cases (all currently refusing) + the subtract analog — the first chunk where comprehension visibly flips at scale. Full scope, cross-references (ADR-0164.2/.3, ADR-0174 multi-actor hazard), wrong=0 obligations, and increments in ADR-0178-GB3b-referent-accumulation-scope.md.
• GB-4 — held structural hypotheses + eliminate (repoint eliminate_violating / contemplate): hold >1 structure on ambiguity, eliminate downstream, refuse if irreducible.
• GB-5 — branch/DAG structures (quantity reuse, 0033-class). The hard end.
• GB-6 — measurement + perturbation + the cue-precision/scale dependency.

Acceptance criteria (Proposed → Accepted)

1. GB-1/GB-2 land; the per-step gate + completeness + uniqueness hold; sequential- chain gold cases flip under self-verification with wrong=0 in serving.
2. Invariant: spurious/unlicensed structures and irreducible holds refuse (proven).
3. Flips hold under ADR-0114a perturbation; determinism/seal invariants hold; capability lanes G1–G5/S1 stay 100% wrong=0.
4. The measurement reports honestly against the relational-cue-precision + scale dependencies (no coverage claim the substrate can't support).

Cross-references

• Repointed substrate: ADR-0174 (held hypotheses, eliminate_violating, reevaluate, contemplate, ProblemReadingState) — generalised from parse to structure.
• Scoring/gate: ADR-0176.
• Guidance/learning (co-evolves): ADR-0177.
• Relational facts: the comparatives pack + a future relational-cue / superordinate-units pack (ADR-0175 §10 self-proving-vs-pack split).
• Scale: ADR-0163 §Phase F.
• Thesis & origin: thesis-decoding-not-generating — this is the word-to-word, lookback-revised, problem-solving-throughout reader the project set out to build; Gap B is where it lands.