01 · PROBLEM
Let low-risk changes from agent configs with good track records auto-land; route others to humans.
02 · HYPOTHESIS
Platform tracks per agent-config (model+prompt+tools hash) land/revert/defect rates; policy maps risk tier x trust to autonomy.
03 · EVIDENCE SO FAR
LATESTNot tested. The agents gathered reports of the pain behind this idea and scored it on paper only; no experiment has been run on it. Cited pain: Copilot approvals counting toward required approvals (E-C246) shows demand; review fatigue (E-C228). Counter: reputation dies on disposable accounts (E-C202) — applies to inbound, less to internal configs.website/projects.json
04 · NEXT FALSIFICATION TEST
Not yet scheduled. Smallest demo on record: two configs; one earns auto-land, other demoted after revert.
StatusNot selected; parked without a test
We would drop or park it ifIf trust score does not predict reverts better than path-risk alone on historical data, kill.