Panout: the commit-boundary record for agent labor
One-liner. Panout records and enforces the decision to stop reading agent work. It runs a team's contracts at the commit boundary, records every override with its outcome, grants autonomy per contract from measured fault-injection catch rates, and is paid per evaluation, the same whether it says yes or no.
Where the score landed. Six independent judge rounds using the gstack office-hours rubric scored 62/72, 66, 67, 64, 59, 61. The judge's stated ceiling without external evidence is about 65. Reaching 90 needs ten operators in writing, three external maintainers, a paying team, and a powered predictive measurement that cannot exist before roughly 2026-10-20. The loop was stopped there rather than manufacturing further rounds.
Final memo
The round-5 strategy memo: moat, $1B path, desperate human, wedge, pricing, and what the founder owes.
MVP
Goal prompt, what shipped in prototype/panout.py, acceptance status, and what remains user-only.
Live outputs
Real status, audit, inject, test run, and trailers from the founder's repos as of build time.
Judge rounds
All five memos and the fresh-judge verdicts that moved the idea from five candidates to one.
Appendices A to H
Every measurement, including the ones that disconfirmed earlier moats.
Decisions
The decision log, including the MVP departures from the goal prompt.
New on 2026-09-04: one more day of measurement
- The predictive test ran today, not in October. Appendix I evaluates the shipped contracts retrospectively on 3,221 founder commits. Contracts fire on 26% of commits (43% of agent commits). Unadjusted, a failure predicts +18 points of rework. Adjusted for commit size, the effect vanishes: the defaults measure exposure, not risk.
- It replicates on repos the founder does not own. Appendix J runs the same test on six public repos including Garry Tan's gstack (91% agent-attributed, 56% override rate, 91% of agent lines reworked within 30 days). Two contracts separate cleanly under fault injection on gstack and openai/codex.
- openai/codex has zero agent trailers in 4,590 commits. Squash merges strip attribution at the vendor's own scale. The commit-time record is the only place it could exist.
- Market sweep. Incumbent, entrant, compliance, incident and pricing changes since the memo, with sources: appendix K.
- Judge round 6 on the new evidence scored 61 (up from 59), ceiling about 64: verdict. Two of six requirements for 90 closed; the three that contain humans remain.
Read this first
- The novel data type is the override. A human committing past a failing contract is recorded in the commit trailer and the ledger while the session that produced it still exists. Nothing else records that moment.
- Autonomy is per contract, never per task class. Appendix H shows task-class rework maps are below the noise floor for any sample a real team produces. Fault injection gives a same-day, chosen-sample answer instead.
- The 30-day rework label was demoted. Appendix G showed the agent-versus-human rework gap is team-specific and near zero pooled across public repos. It remains a slow audit, not a product claim.
- Zero new habits. Two git hook shims, metadata-only ledger, commits never blocked, no login, Python stdlib only.
Current dogfood status
Panout status
Contracts installed: 4 (tests-touch, no-secrets, size-guard, agent-attribution)
Evaluations recorded: 8 total, 8 in the last 7 days, 2 for commit(s) no longer on HEAD (amended or reset; kept as recorded)
Exposure: 8/8 commits with a non-skip result (100%; target 30%)
Overrides per contract:
- tests-touch: 2 override(s) over 4 evaluation(s)
- no-secrets: 0 override(s) over 8 evaluation(s)
- size-guard: 4 override(s) over 8 evaluation(s)
- agent-attribution: 0 override(s) over 8 evaluation(s)