17.11.5 Auditable rhetorical realization

17.11.5 Auditable rhetorical realization

An admitted program is not yet a grant proposal. GLP1–GRANT–4 tests the passage from the GRANT–3 IR to reader-facing prose under three conditions: a free-form language-model proposal, the frozen deterministic compiler, and an auditable realization. In the third condition, Qwen3-Next-80B emitted a JSON rhetorical plan whose atomic sentences carried a registered epistemic type and support identifiers. A deterministic compiler produced both a clean proposal and a marker-bearing audit document, from which a separate parser reconstructed the sentence ledger.

The registered language distinguished source-supported statements, hypotheses, admitted inferences, open variables, resolution operators, resource boundaries, and epistemic boundaries. Admission required the proposal to cover all three contracts, hypotheses, alternatives, and fallbacks; all twenty-one commitments and resolution operators; and the program’s resource and epistemic boundaries. It also prohibited unsupported priority language, causal claims about the nonrandomized exposure history, and invented protocol numbers. Eight counter-witnesses tested precisely these boundaries as well as source fabrication, hypothesis upgrading, marker deletion, omitted alternatives, and apparent closure of an open variable.

The auditable proposal was not admitted: fifteen of nineteen gates passed. Its 64 typed sentences covered every required program object, and its sentence ledger round-tripped exactly. All eight mutations were rejected for their registered reasons. The generator nevertheless repeated the inherited claim that this was the first prospective study, invented six unregistered sentence types called alternatives and fallback, and typed three causal-contract sentences as epistemic boundaries. The first failure shows that realization can detect a defect in the admitted input theory rather than merely introduce one. The latter failures suggest a conservative extension of the rhetorical language: causal contracts, alternative explanations, and fallback policies may deserve explicit types.

The independent free-form condition supplied the contrasting failure mode. It had no sentence ledger, used only five of the seven required headings, repeated both unsupported novelty and causal claims, and invented fixed protocol timings and analytic commitments. This comparison is diagnostic rather than a readability result: no human reviewer preference was measured. A retrospective audit also found that the supposedly clean auditable proposal echoed registry identifiers in visible prose and exceeded its requested word budget. A stricter realization language must therefore sanitize the admitted theory, separate machine certificates from surface text, type the rhetorical budget, and add the new sentence categories without weakening existing checks.

Experiment: GLP1–GRANT–4: auditable proposal realization. Input: the admitted GRANT–3 IR, frozen deterministic proposal, finite sentence-type language, and support-identifier registry.
Conditions: one temperature-zero Qwen3-Next-80B free-form proposal and one sentence-audited rhetorical plan; deterministic compilation and independent ledger recovery.
Result: 15/19 gates; 64 typed sentences; complete coverage of 21 commitments and 21 resolution operators; exact round trip; all eight counter-witnesses rejected.
Rejection: unsupported priority language, six unregistered rhetorical types, and three contract/type mismatches.
Boundary: auditable realization calibration, not a biomedical, readability, peer-review, novelty, or fundability assessment.