Warrant

Did agent-merged work deserve to exist? The unwarranted-merge rate — agent PRs later reverted because the change shouldn't have been built (a hallucinated premise), not because of a bug — over the trailing 90 days. As of 2026-08-27 07:50 MST.

Unwarranted rate

0.2%

Reverted

5 (0.5%)

Merged agent PRs

961

Bug-reverts

3 (0.3%)

By producer model

Model	Merged	Reverted	Unwarranted	Bug-reverts
deepseek/deepseek-v4-flash	75	1 (1.3%)	1 (1.3%)	0
claude-sonnet-4-6	662	3 (0.5%)	1 (0.2%)	2
claude-sonnet-5	114	1 (0.9%)	0 (0.0%)	1
claude-opus-4-7	4	0 (0.0%)	0 (0.0%)	0
claude-opus-5	2	0 (0.0%)	0 (0.0%)	0
deepseek/deepseek-chat	98	0 (0.0%)	0 (0.0%)	0
moonshot/kimi-k2.7-code	4	0 (0.0%)	0 (0.0%)	0
openai/mimo-v2.5-pro	2	0 (0.0%)	0 (0.0%)	0

By repo

Repo	Merged	Reverted	Unwarranted	Bug-reverts
Platform & internal	127	1 (0.8%)	1 (0.8%)	0
Lyra	508	2 (0.4%)	1 (0.2%)	1
Corvus	214	2 (0.9%)	0 (0.0%)	2
Vela	105	0 (0.0%)	0 (0.0%)	0
Pavo	1	0 (0.0%)	0 (0.0%)	0
Orion	6	0 (0.0%)	0 (0.0%)	0

A hallucinated-but-correct PR scores positively on every other axis — merged, survives 7 days, +velocity, zero security findings — so it's invisible unless measured here. Code review can't catch it: the defect is the premise, not the code.
Unwarranted = reverted because the change shouldn't have existed (the spurious-work rate, the headline). Bug-reverts = reverted because it broke something (a quality-escape, complementing the survival metric). The two need opposite fixes: unwarranted → better upstream task-grounding; bug → tighter review/CI.
The signal is the revert, classified by a Revert-reason: tag on the revert PR (unwarranted | bug | superseded | other). A detected revert with no tag reads unclassified until a human labels it — never silently dropped.
The rate colouring is a soft read (zero clean, above flagged), not a policy threshold — that gets set from the accumulated baseline once there's enough to measure.