`ablate report`

Re-analyze a stored run. Zero API calls, zero cost.


ablate report --last
ablate report --last --margin 0.03
ablate report --list

Arguments & flags

Argument/flag	Default	What
`RUN`	none	Run id or path under `.ablate/runs/`.
`--last`	off	Re-analyze the most recent run.
`--margin PCT`	`0.05`	Re-decide the verdict with a different equivalence margin.
`--alpha A`	`0.05`	Re-decide with a different significance level.
`--seed N`	run’s original seed	Bootstrap seed.
`--list`	off	List stored runs instead of analyzing one.
`--format FMT`	`text`	Also write `md`, `json`, `junit`, or `pr-comment`.
`--fail-on WHEN`	`never`	`regression`, `inconclusive`, or `never`.

What it does

Every compare or sections run is stored under .ablate/runs/<id>/ as a JSONL trial log. ablate report re-reads that log and re-computes the verdict — no agent is invoked, so it’s free.

This is genuinely useful — deciding you actually care about ±3pp instead of ±5pp after the fact costs nothing. It is also how a dataset gets tortured: re-running the analysis until the margin you like produces the verdict you want is sequential testing without correction. ablate report doesn’t stop you from doing that, but it counts every re-analysis against a run and prints the count, so it’s visible rather than silent.