D312 - A run helped by named overrides was recorded as an honest measurement

decided · 2026-08-27 · found by asking what carries a finding back into the repository

compat record refuses a run whose stubs were reporting success, because such a run reaches further by construction and would become a best-ever entry no honest run could beat (D182). The guard:

!self.default_return.is_empty() && self.default_return != "unimplemented"

It reads the global default. Learned::policy() deliberately leaves that at unimplemented and puts its answers in per-function overrides (D296), so a measured policy reports false and records clean. The feature built to carry the loop's findings drove straight through the guard written to stop exactly this.

Conditions already counted the overrides. Nothing read the count, and policy_summary's own doc says why it exists: "loosening the default is the single change that improves every number in a report while implementing nothing" - which is true of answering one function by name, at a smaller scale.

The refusal was the wrong repair. Turning a propped-up run away means a person has to pass --force on every measured policy, and throws away the number that says whether a patch is worth pursuing. The refusal only existed because a single best-ever entry could not hold both kinds of result.

So the record holds two: [status] for what the emulator does as it stands, [experiment] for the furthest it got while being helped. A run is routed to its own slot, compared only against that slot, and never refused on policy grounds. --force survives for one case - overwriting a better entry within a slot.

Comparability is a fact rather than a count: propped or not. Two experiments differing by one override are not different in kind, and treating them as incomparable would recreate the refusal in another form.

This is what "without a person" costs. Every guard here that refuses rather than records is a place the loop stops and waits. The honest version of a guard is usually a field, not a gate - the run happened, the number is real, and what a reader needs is to know which question it answers.