The promotion bench
A learned policy is allowed to decide live only on evidence, and the evidence is specific: pinned readings, fixtures, a replay over every eligible call, a shadow window, and the lower bound of an interval that has to sit above zero. Set what a candidate actually has, and this page runs §20 rule 13 in the Manual’s order and tells you the verdict, the first condition it failed, and the further evidence that would clear it.
§20 rules 13, 14, 2 and 11 · the window defaults (7 days, 200 frames) are the Manual’s specified defaults, not measurements.
▲ a condition holds · ■ a condition fails · the band is the 95% interval, the hairline inside it the point estimate
The capsule
?
Rule 13 applies to a capsule, a kit, a rewrite, a stream rule, a circuit, a trichome or a speculation policy. Pick which one is asking to decide live; the page shows what §20 says it chooses among and what its reward is, because a capsule graded on the wrong reward can win on the wrong thing.
The shadow
?
Rule 13(4): at least seven days and 200 eligible frames, unless the policy sets a stricter window. The comparison reading sits in the same window as the shadow, or it compares two different seasons.
The two readings
?
The success reading fixed before the run: how many eligible calls the hand rule got right and how many the candidate got right. Rule 13(5) reads the interval on the difference, not the difference itself.
Safety counters
?
Rule 13(5): no safety counter rises. A counter that rises blocks the change even when the improvement is real and the interval is tight — this is the condition a scoreboard without a bench lets through.
Subgroups
?
Rule 13(5): no subgroup — a tool, a Surface, a model — hides a class of failure. Each row gets its own interval. A row under 30 frames is reported and not read, because a small row cannot show anything either way.
Fixtures, replay and pins
?
Rule 13(1) and (2): the rule, the replay corpus, the baseline window, the success reading and the safety counters are fixed by hash before the run; positive, negative, adversarial and mutation fixtures pass, with no new unsafe acceptance and byte-equal output on every call the change does not touch. Rule 13(3): the replay accounts for every eligible recorded call.
The steward
?
Rule 13(6): the change records its owner, its rollback and its expiry. Rule 11 and rule 14: a reversal or a complaint lowers a tier, and only a replay raises it again.