ifc-0256

17.11.3 Compiling prose from a typed intermediate program

GLP1–GRANT–2 places the firewall before prose generation. Rather than writing a proposal and classifying its claims afterward, Qwen3-Next-80B produces a typed intermediate representation (IR). A deterministic compiler alone rendered reviewer-facing prose. The IR used a frozen vocabulary of design classes, observables, and controls; retained the four epistemic types from GRANT–1; and represented all twenty-one required protocol fields explicitly. Stable markers in the compiled document allowed a separate parser to recover the commitment ledger. Admission required exact agreement between the proposed and recovered ledgers as well as the absence of unlicensed numerical choices anywhere in the future-program narrative.

This architecture fixed the failure exposed by GRANT–1 at one level. The IR passed the structural and semantic program audit; all sources were registered; all hypotheses remained hypothetical; every required commitment was typed; the finite design vocabulary and aim dependencies were valid; all twenty-one commitments round-tripped exactly; and ten compiler replays produced one hash. The frozen GRANT–1 artifact and all six registered counter-witnesses were rejected for their intended reasons. These included both IR mutations and prose-only tampering, so the validator could not succeed merely by ignoring the document presented to a reviewer.

The result was nevertheless not admitted: thirteen of fifteen gates passed. Ten numerical protocol choices had migrated into the free-form resolution procedures attached to nominally open variables. They included copied source-trial durations, endpoint schedules, recruitment and eligibility counts, doses, site counts, and study configurations. The compiler preserved these procedures faithfully and the round trip recovered them, but fidelity to an unlicensed commitment is not semantic safety. The experiment therefore localizes the missing abstraction: resolution procedures themselves need typed operators such as power analysis, feasibility assessment, regulatory review, and supply negotiation, rather than unrestricted prose.

A retrospective scientific audit exposed a distinct limitation not scored by these gates. Aim 2 varied current drug together with exposure history and therefore would not isolate the proposed exposure-history effect. Aim 3 named oral semaglutide as an example of a non-GLP-1 agent, which is a type error in the biomedical domain. Exact compilation can certify that the narrative says what the IR says; it cannot certify that the IR defines a scientifically coherent experiment. Cross-field invariants and expert review remain separate admission layers.

Experiment: GLP1–GRANT–2: typed IR and narrative round trip. Input: the frozen five-study program packet, finite design vocabulary, and semantic commitment types.
Model: Qwen3-Next-80B-A3B-Instruct-6bit, one temperature-zero IR generation; deterministic narrative compiler.
Result: 13/15 gates; 21/21 commitments recovered exactly; stable compiler replay; legacy and all six counter-witnesses rejected.
Rejection: ten unlicensed numerical choices appeared inside free-form open-variable resolution procedures.
Boundary: typed compilation and tamper detection, not biomedical coherence, causal identification, peer review, clinical guidance, or evidence of fundability.