ratchet

Point it at a number and let it climb. /ratchet is a metric-gated optimization loop — Karpathy's AutoResearch pattern (propose a change, run a fixed-budget evaluator, keep the commit if the metric improved, git reset if not, repeat) adapted to Cohort's human gate: the whole climb runs inside a throwaway git worktree, bounded by a budget, and you review the staircase and merge via PR. The autonomy is the inner loop; the merge stays gated. Reach for it when the win is measurable and the search is tedious — a perf number, a benchmark score, a failing-test count, a lint count.

Runs on a coordinator tier (Fable or Opus): you set up the contract and read the staircase; the loop does the methodical climbing.

The three-part contract

Make these three things explicit before you start — it is what makes autonomy safe:

• The immutable evaluator — one command that prints the objective number, and that the loop never edits. This is the ground truth (Karpathy's prepare.py). The command is yours; the code it executes is the engine's — it runs in the worktree the engine just wrote to, so it is confined like a doer (see Guardrails). If the doer could change the evaluator, it could optimize the metric by lying; it can only touch the worktree's tracked code, which you review — so a metric that jumped because the change broke or gamed the evaluator is yours to catch at the staircase.
• The sandbox — a detached worktree off HEAD. Every proposal lands only here; your working tree is never touched, and a bad run is thrown away.
• The direction — the objective in words. Keep it tight and surgical ("lower p99 latency in handler.py; change nothing else").

Run it

cohort engine ratchet gpt \
  --evaluator "pytest tests/bench.py -q 2>&1 | tail -1" \
  --metric-regex 'score=([0-9.]+)' \
  --goal maximize --budget 15 --footprint src/handler.py

• gpt (Codex, edits under its own sandbox) or grok (egress-gated agentic patch) does the proposing; the loop's keep/revert, worktree, ledger, and budget are enforced in code.
• The evaluator runs with a scrubbed environment and, on Linux with bubblewrap, no network and no access to your home — so put any variable it needs inside the command (FOO=1 pytest …) and make sure the tools it calls are system-installed or live in the worktree; a venv under your home directory is invisible inside the jail.
• Each iteration is fed the current best and the recent ledger so it calibrates what to try next — the ratchet-results.tsv staircase is the loop's memory.
• Ties and non-improvements revert. The lineage only advances on a real gain.