Expand description
Repeat sets and rustev.repeat-report/1 (spec 029, 3.1, 3.5, 3.6).
A repeat set is N evaluations of one configuration, each with its
reconciled case measures. The repeat report states, per cohort (all,
then each declared slice) and per metric, the per-run values, mean,
sd, min, max, pooled, the case-clustered interval and the noise
band with its basis, and names every document and seed it came from. It
is evaluation evidence for the host, never a grant, a selection or a
quality claim.
Structs§
- Cohort
Report - Held
OutInput - The host’s held-out inputs (3.7.3): its lineage handle, its ledger or the explicit statement that the ledger is new, and the time to record.
- Metric
Report - Repeat
Inputs - Repeat
Report - Run
- One run: its evaluation and its case measures.
- RunRef
- The record digests of one run’s documents and the seed handed to it.
Enums§
- BandDoc
- Basis
Doc - Cohort
Status - Figure
Doc - HeldOut
- The held-out record (3.7.4). Without a ledger the count is unknown, never zero.
- Interval
Doc - Observation
- Whether a latency, queue or cost figure was observed by the evaluated plan (3.11.2).
- Repeat
Error - Why no repeat report was produced (
invalid_input, 3.14). - Source
- Where a repeat set’s runs came from (3.1.2).
- Split
Role - A split’s fixed role (3.7.1).
Constants§
Functions§
- repeat
- Assemble a repeat report from N reconciled runs.
- repeat_
held_ out - A held-out repeat report counted against the host’s ledger (3.7.4): the report records the counts and heads, and the new ledger (the input with one entry appended) is returned for the host to keep. The input ledger is never modified.
- split_
role