17.11.2 Semantic admission as a type system

17.11.2 Semantic admission as a type system

GLP1–GRANT–1 makes the semantic firewall executable. It reuses the same five-study ledger, exposure-history obstruction, mock funding call, team profile, and three-aim grammar, but replaced unconstrained specificity by four explicit epistemic types: source-supported, externally supplied, hypothetical, and open variable. Every unresolved sample, duration, regimen, endpoint, site, recruitment, supply, budget, data-access, or authorization choice had to be represented as TO_BE_DETERMINED together with a procedure for resolving it. Thus omission and invention were both type errors: a proposal could not gain apparent rigor by silently choosing plausible numbers.

One temperature-zero GPT-OSS 20B generation satisfied the structural contract, retained three discriminating dependency-ordered aims, used only registered sources, kept all hypotheses hypothetical, supplied every required commitment field with a valid type, and explicitly separated resource gaps from human authorizations. The firewall also rejected the frozen GLP1–GRANT–0 proposal and five deterministic counter-witnesses: an unsupported sample size, a fabricated source, a false authorization, a promoted hypothesis, and a missing commitment. Each mutation was rejected for its preregistered reason.

The generated program nevertheless failed admission. Although duration was correctly declared open in the protocol table, the surrounding narrative specified 12-month discriminators for all three aims and 24-month alternative outcomes for Aims 2 and 3. These five statements were unlicensed future protocol commitments. A frozen-output evaluator correction excluded the digit in the lexical term “GLP-1” from the numerical detector and extended the same preregistered rule to all protocol-bearing aim prose; neither the output nor the threshold was changed. The corrected result passed ten of twelve gates and remained not admitted.

This failure is more informative than a malformed proposal. The generator learned to populate the explicit type system but treated fluent prose as an escape path around it. A semantic firewall must therefore check the entire artifact, not merely its designated commitment table. It must also distinguish three successively stronger claims: a well-typed proposal, a scientifically credible proposal, and a fundable proposal. GLP1–GRANT–1 tests only the first.

Experiment: GLP1–GRANT–1: semantic admission firewall. Input: the frozen GLP1–GRANT–0 compilation packet plus epistemic types and required open-variable resolution procedures.
Model: GPT-OSS 20B, one temperature-zero generation.
Result: 10/12 gates; all structural, provenance, typing, dependency, resource-boundary, legacy-rejection, and counter-witness gates passed.
Rejection: five untyped duration commitments leaked into narrative discriminators and alternative outcomes.
Boundary: semantic contract checking over a five-source curated corpus, not biomedical review, feasibility validation, clinical guidance, or evidence of fundability.