pub struct AutoPromotion {
pub dataset_version: Option<String>,
pub min_improvement: f64,
pub scopes: Vec<MemoryScope>,
}Expand description
The evidence thresholds an auto-promotion must clear (the design’s
rule): a clean replay, compare() showing no regression, and
improvement on the target metric — over the named dataset version,
when the deployment names one.
Fields§
§dataset_version: Option<String>The dataset version a clean verdict must come from. None accepts
the verdict’s own version — honest for deployments that version
datasets per evaluation; a named version pins the promotion to the
exact evidence the deployment reviewed.
min_improvement: f64The target-metric delta must exceed this (strictly). 0.0 is the
design’s default: any improvement clears, parity does not.
scopes: Vec<MemoryScope>The memory scopes a memory_set candidate may auto-promote at
(ignored for other kinds; empty = unrestricted). The R0.8 default
names run and agent (open question 6): wider scopes always
require a human approval this release.