ifc-0260

17.11.7 Structure-first rhetorical realization

GLP1–GRANT–4.2 relocates the proposed structural repair from unconstrained prose generation to a deterministic rhetorical planner. Before Qwen sees the program, the compiler allocates 67 immutable sentence slots. Each slot had a fixed section, epistemic type, support array, and semantic payload. The inventory included every hypothesis, identification contract, program field, source statement, alternative, fallback, open commitment, resolution operator, resource boundary, and epistemic boundary. The model could supply only one reader-facing sentence for each identifier in the frozen order. It could neither add nor delete a slot, change its type, nor move its evidential support.

This intervention repaired the structural failures of GRANT–4.1. The output filled all 67 slots exactly, preserved every type and support edge, included both resource gaps and human authorizations, retained all source numbers, and round-tripped through the audit compiler. It contained none of the registered novelty, causal, protocol-number, or surface-token violations. All eleven counter-witnesses—including missing and extra realizations, semantic replacement, numerical drift, metadata tampering, and marker deletion—were rejected for their intended reasons.

The experiment was nevertheless not admitted: fifteen of seventeen gates passed. The model expanded a 1,411-word semantic skeleton into 1,958 words, exceeding the hard 1,500-word ceiling. The artifact-admission gate therefore failed as a consequence. This is more informative than a cosmetic length error. A retrospective audit found that the expansion added plausible but unlicensed details, including particular missing-data estimators, IRB and award timelines, enrollment targets, unblinding, and agency review cycles. The registered overlap test ensured that the original payload remained visible; it did not ensure that no new claim had been appended.

The residual obstruction is thus semantic accretion. Structure-first planning can guarantee that the generator says something about every required object, but not that it says only what the admitted theory licenses. Conservative realization consequently requires an upper semantic bound—for example, deletion-only compression, bidirectional claim entailment, or deterministic clause compilation with a narrowly bounded connective layer. The frozen proposal cannot acquire that certificate through post-hoc editing.

Experiment: GLP1–GRANT–4.2: structure-first realization. Input: an immutable 67-slot rhetorical skeleton constructed from the sanitized GRANT–3 theory before prose generation.
Model: one temperature-zero Qwen3-Next-80B realization; text fields only and no repair call.
Result: 15/17 gates; exact slot, type, support, source-number, coverage, round-trip, and replay checks; all eleven counter-witnesses rejected.
Rejection: 1,958 words exceeded the 1,500-word ceiling; a secondary audit outside the registered gates also found unsupported procedural elaboration.
Interpretation: structure-first generation solved omission and retyping, while conservative semantic realization remains unresolved; this is not biomedical validation, peer review, or evidence of fundability.