Skip to Content
CLI referencesections (ablate)

ablate sections

Full section attribution: which lines earn their keep. ~$110 default at K=12, ~90 min. Also registered as ablate ablate (same command — sections is the form used everywhere so it never reads as a typo).

ablate sections CLAUDE.md
ablate sections CLAUDE.md --confirm 5 --granularity bullet

Arguments & flags

Argument/flagDefaultWhat
FILEauto-detectedConfig file.
--tasks DIR.ablate/tasksDirectory of task YAML files.
--tasks-only IDSnoneRun only these task ids or globs.
--tasks-tag TAGSnoneRun only tasks carrying any of these tags.
--trials N5Replicates per cell, confirmatory stage.
--agent NAMEclaude-codeclaude-code or codex.
--model IDSonnet 5Target model for the confirmatory stage.
--screen-model IDHaiku 4.5Model for the screening stage.
--no-screenoffScreen on the target model instead (~4× the price).
--confirm N3How many sections to run a confirmatory A/B on.
--granularitysectionsection or bullet. Bullet is literally per-item and roughly triples the cost.
--concurrency N4Parallel trials.
--max-turns N40Turn budget per trial.
--max-spend USD25Hard ceiling.
--margin PCT0.05Equivalence margin for the confirmatory stage.
--alpha A0.05Significance level.
--seed NnoneDesign and bootstrap seed.
--yesoffSkip the confirm prompt.
--freshoffIgnore cached trials.
--keep-worktreesoffDon’t clean up trial worktrees afterward.

What it does

Two stages:

  1. Screening — every section is independently kept or dropped across a randomized subset design (R = 2K + 8 variants) on a cheap model, one trial per (variant, task). Fits a ridge regression with task fixed effects to rank sections by estimated effect. This stage ranks; it does not produce verdicts. Read the order of the table, not the coefficient magnitudes — they’re shrunk toward zero by design.
  2. Confirmation — the top-ranked sections and the largest low-signal (“deletion candidate”) sections get a dedicated, pre-specified paired A/B on the target model, at full --trials. Only this stage produces a verdict, and only an EQUIVALENT verdict licenses a deletion.

ablate apply performs the actual edit on sections this command recommends dropping.

Last updated on
compareapply