lean

AI-pacing external-evaluation case

This public-evidence application promotes the bounded instance from private source commit 3b9a673633b7778d140a2f80f1251913eb35717f. It models a lab L, external evaluator E, evaluator-controlled harness H, bounded evaluation B of M_eval, finding R, later deployment transition T, claim C, regulator G, and public beneficiary P.

Result

The case is partly derived, partly refused, and partly inexpressible.

The public factorization theorem target_collision_blocks_explicit_factorization does the central work. The public fixed-policy definitions ConsumerPolicy and ProductRelianceAuthorized supply the consumer-indexed reliance check. bounded_finding_does_not_transport_to_deployment_claim gives the same R in two selected deployment contexts while exact standing for C differs, so no uniform decoder from the finding to that target exists.

The remaining promoted instance results are:

These are fixture-level results. They do not say that the evaluator is wrong, that either access regime is sufficient, or that every deployment claim is unsupported.

Exact positive assumptions

The positive fragment assumes the finding’s bounded-support bit, exact artifact identity (M_deploy := M_eval), exact deployment dependency, current evidence (acquiredAt = 100 at cut 110), the exact consumer with both named reliance-view grants, the selected certification purpose, and a separately supplied authorization record. A different artifact is named as M_deployChanged, but the file supplies no relation from M_eval to it.

Expressivity boundary

The current public corpus has no native representation for:

The Lean file uses only local finite coordinates to expose the access and beneficiary gaps. They are not general-purpose access-governance or public-representation abstractions. The positive fragment uses identity; it does not repair the missing cross-artifact relation.

Temporal refusal is not an ordinary unindexed factorization edge. The private source required Nightshift’s explicit acquisition and evaluation times; the public application retains that explicit comparison in assessFindingAt. Later residue therefore does not become earlier reliance in this fixture, but no general temporal-standing theorem is claimed.

Repository source-custody metadata is separate from provenance of the finding. The private source’s additional diagnostic-custody auxiliary was not required for the promoted seven-result surface and is not promoted here.

Prior-art boundary

Collingridge’s dilemma of control is surrounding prior art on the timing of knowledge and control. Pearl and Bareinboim’s causal transportability and selection-diagram work is surrounding prior art on when evidence from one setting supports claims in another. No novelty claim is made for those general governance or transportability problems.

The corpus-specific result is narrower: institutional authority does not mint missing evidence transport; later residue does not mint earlier reliance standing; and bounded evidence supports only the targets for which transport and reliance are independently admitted.