ifc-0261

17.11.8 Local metric and semantic bounds

GLP1–GRANT–4.3 treats verbosity as a local rather than merely global obstruction. Each of the 67 frozen slots receives a minimum and maximum word count before generation, with its maximum no larger than the admitted semantic payload. The sum of the local maxima was 1,453 words, so satisfying every local fiber automatically respected the 1,500-word document ceiling; a 900-word global floor prohibited empty or trivial compression. A second bound made semantic fidelity bidirectional: the earlier coverage test prevented replacement of the payload, while a normalized vocabulary test limited new content not licensed by it.

The single temperature-zero Qwen3-Next-80B realization passed all nineteen registered gates. It filled every slot in order, preserved all types, supports, source numbers, and program objects, and contained 1,453 words. It introduced no registered novelty, causal, numeric-protocol, or surface-token violation; its sentence ledger round-tripped exactly; and all fourteen counter-witnesses were rejected for their intended reasons. Relative to this bounded rhetorical language, the artifact was admitted.

The secondary diagnostic is essential to interpreting that success. All 67 output strings were byte-for-byte identical to their semantic payloads. Faced with strict upper bounds, the model chose exact transcription rather than paraphrase. This is a legitimate conservative solution and a useful contrast with the accretion in GRANT–4.2, but it adds no demonstrated rhetorical value. The copied compiler text retains awkward constructions such as “Alternative explanations include Weight,” “remains open because Requires,” and “The sites remains open.” Neither readability nor reviewer preference was a registered outcome.

GRANT–4.3 is therefore an admission of bounded faithful realization, not of successful grant writing. A decisive comparison retains exact copying as a conservative baseline while testing a controlled paraphrase language under bidirectional claim entailment and blinded human readability. That would ask whether a realization layer can improve prose without weakening the semantic firewall.

Experiment: GLP1–GRANT–4.3: bounded realization. Input: the 67-slot structure-first skeleton with local word intervals, a global 900–1,500-word contract, and lower and upper semantic bounds.
Model: one temperature-zero Qwen3-Next-80B realization; no repair call.
Result: 19/19 gates; 1,453 words; exact structural, semantic, numerical, round-trip, and replay checks; all fourteen counter-witnesses rejected.
Diagnostic: 67/67 sentences exactly copied their payloads.
Interpretation: admitted as bounded faithful transcription, not as a readability, peer-review, biomedical, or fundability result.