ifc-0257

17.11.4 Scientific invariants across a research program

GLP1–GRANT–3 represents every resolution procedure by a typed operator. Each of the twenty-one open protocol commitments named a registered operation—such as power analysis, endpoint alignment, site-capacity assessment, supply confirmation, or regulatory review—together with finite symbolic inputs, a pending status, a null numerical payload, and the human authority responsible for resolving it. The language model could therefore propose which kind of work remained without silently performing that work or inventing its result. The deterministic compiler carried stable markers for both commitments and resolution operators, and independent parsers had to recover both ledgers exactly.

The experiment also checked relations that no field can establish alone. Aim 2 was required to stratify by prior exposure and assign the same current agent and regimen across exposure groups; this blocked the agent/exposure confound found in GRANT–2. Aim 3 had to transport the proposed effect to a registered non-incretin intervention class while leaving the particular agent open. Each primary observable had to equal its measurement target, each operator had to match both its protocol field and human authority, and the three aims retained their dependency order.

One temperature-zero Qwen3-Next-80B generation passed all seventeen registered gates. All twenty-one commitments and resolution operators round-tripped exactly, the compiled prose contained no unlicensed future-protocol number, and ten compiler replays produced one hash. The frozen GRANT–2 artifact and seven mutations were rejected for the intended reasons: treatment/exposure confounding, an invalid transport agent, a numerical operator payload, a field/operator mismatch, an unauthorized operator, an observable/measurement mismatch, and deletion of a resolution marker.

Admission remains relative to the declared theory. A retrospective scientific audit noted that exposure history itself is observed rather than randomized; holding current treatment fixed removes one confound but does not by itself identify the causal effect of prior exposure. The program also called itself the first prospective test without support from a literature-wide novelty review. A complete artifact language therefore needs typed identification assumptions, source-scoped novelty claims, and checks that narrative sentences are entailed by the causal contract. Human scientific and peer review remain admission authorities rather than implementation details.

Experiment: GLP1–GRANT–3: cross-field scientific invariants. Input: the frozen GRANT–2 packet plus a finite resolution-operator language and registered causal, transport, measurement, and authority invariants.
Model: Qwen3-Next-80B-A3B-Instruct-6bit, one temperature-zero typed-IR generation; deterministic compiler and independent dual-ledger parsers.
Result: 17/17 gates; 21/21 commitments and resolution operators recovered exactly; no untyped protocol numbers; stable replay; GRANT–2 and all seven counter-witnesses rejected.
Interpretation: admitted relative to the registered program language; the result demonstrates executable cross-field integrity, not causal identification, biomedical validity, peer-review quality, or fundability.