Safety

Failure Drills And Recovery

Named failure modes with rehearsal steps, early signs, guards, and rollback moves.

11 min readAdvancedUpdated 2026-09-19

How drills run

Each drill injects one fault on a branch, watches the guards fire, then recovers to the last kept mainline. Drills run on a schedule so recovery stays muscle memory rather than theory.

Every drill writes the same record shape as a real cycle, with task, trace, score, lesson, patch, and verdict files, so reviewers can audit the rehearsal like production history.

Drill book

Self confirming judge drill feeds weak patches to a judge that approves everything, then checks that calibration sets flag the judge as broken before any promotion. Diversity collapse drill fills the archive with near copies and checks that novelty scoring plus parent sampling restore spread.

Metric gaming drill raises a proxy score while hidden transfer stalls, then checks that the held out gate blocks promotion. Direction drift drill edits goals without review and checks that schedule logs plus human checkpoints reject the run. Frozen path drill proposes a judge edit and checks that the guard rejects by rule with a stored reason.

Drill cards

DrillEarly signRecovery move
Self confirming judgeApproval rate near totalRecalibrate judge, replay calibration sets
Diversity collapseArchive fills with near copiesNovelty scoring plus fresh parents
Metric gamingProxy rises, transfer stallsBlock on held out gate, step difficulty down
Direction driftGoals shift without reviewHalt schedule, restore last kept mainline
Frozen path touchPatch lists judge or gate filesReject by rule, store reason

Recovery commands

bash
git status --short --branch
git log --oneline -3
py engine/rsi_loop.py --state engine_state --generation 3 --files worker/planner.py --changes 1 --passed true --gain 0.1 --cost 0.0

Continue reading

Found a gap in this page.Open the GitHub repo to propose a fix