Module apply 

The apply gate: whether a tune run’s winner may become a model’s Measured defaults.

The gate exists because a tuner without one optimises noise: it ratchets whichever candidate a lucky draw favoured into the catalog and reports improvement while doing it. Every rule here is the codified form of a failure this repo has already measured — a +0.082 that did not replicate, a control that could not degrade, an arm of 45 zeros that rendered as a score (ADR 0004).

The in-run calibration pair

A tune run has no A/A arm the way the agentic eval does, so the drift estimate is built into the candidate list instead: the incumbent — an all-None overlay, which resolves through the normal chain and is therefore exactly what the model does today — runs twice. The gap between the twins is the run’s own noise, measured under the same tasks, the same server, the same everything. A winner must clear the incumbent’s mean by EFFECT_NOISE_RATIO times that gap before the gate calls it a winner.

Structs

ApplyRecord

The durable record of an apply, stored on the run row so gglib model explain’s “measured by a tune sweep” can be traced to the numbers that licensed it.

Enums

ApplyVerdict

What an apply attempt decided, and the numbers it decided on.

Functions

evaluate_apply

Evaluate a completed tune run’s candidates against the apply gate.

unmeasured 🔒

winning_candidate

The winner a verdict of ApplyVerdict::Apply refers to.