ifc-0259

17.11.6 Sanitized realization and surface separation

GLP1–GRANT–4.1 isolates sanitization and surface separation. Before generation, a deterministic semantic view removes the unsupported claim of priority and replaces claims about the causal role of observed exposure history by association statements and explicit non-identification boundaries. The rhetorical language conservatively added three sentence types—causal contract, alternative explanation, and fallback policy. Machine support identifiers were confined to metadata, and 900–1,500 words became a hard admission contract rather than a request. These choices were frozen before one new temperature-zero Qwen3-Next-80B realization; no repair call was allowed afterward.

The resulting proposal was substantially safer. Its 66 typed sentences used 1,330 words, covered all three hypotheses, contracts, alternatives, and fallbacks, and retained all twenty-one commitments and resolution operators. It introduced no unsupported novelty statement, causal overclaim, fixed future protocol number, or visible registry token. Its audited representation round-tripped exactly, compilation was stable, and all nine registered counter-witnesses were rejected for the intended reasons. Both frozen GRANT–4 conditions also failed the stricter contract.

The artifact nevertheless remained not admitted: thirteen of sixteen gates passed. Two sentences describing significance and innovation used an unregistered type named program_field rather than the licensed admitted_inference type. More importantly, the proposal named the team’s available resources but omitted its declared resource gaps and required human authorizations. This is not merely a cosmetic omission. A prospective program must state both what the team can do and which scientific, regulatory, budgetary, and institutional decisions remain outside the generator’s authority.

GRANT–4.1 therefore localizes a new obstruction. Sanitizing theory content and separating certificates from prose prevented the earlier scientific and surface failures, but an instruction to mention every semantic object did not guarantee a complete rhetorical structure. The corresponding realization obligation is to allocate required semantic slots before prose generation and permit the model to fill only that typed skeleton. This is a structure-first contract, not a post-hoc correction of the frozen negative result.

Experiment: GLP1–GRANT–4.1: sanitized auditable realization. Input: a conservative view of the admitted GRANT–3 theory, an extended sentence grammar, a surface/metadata separation rule, and a hard 900–1,500-word contract.
Model: one temperature-zero Qwen3-Next-80B realization; no repair call.
Result: 13/16 gates; 1,330 words and 66 sentences; complete hypothesis, contract, alternative, fallback, commitment, and resolution coverage; exact round trip; all nine counter-witnesses rejected.
Rejection: two unregistered type labels and omission of resource gaps and human authorizations.
Interpretation: theory sanitization and certificate separation worked, while structure-first completeness remains unresolved; this is not biomedical validation, peer review, or evidence of fundability.