SREForge stands up a real, lesser-known public app as a lived-in deployment, authors an organic regression into its existing code so a genuine incident surfaces as a real alert, hands an agent only a short trigger, and lets it investigate through the same tools a human on-call would use. It grades the agent on behavioural recovery — not on matching a reference diff. You cannot bluff a behavioural oracle.
Closed-loop behavioural verification
The fault stimulus keeps running while the fix is verified. An alert clears only because the deployed change works — not because the harness stopped poking the system. Diff-matching is at most a hint, never the grade.
Contamination-free by construction
v1 substrates are real apps with no public solution to memorise. Faults are authored into existing code, and the run workspace’s git remote is a local forge — so the fault can’t be diffed against the upstream.
Authored, reproducible incidents
A determinism gate confirms the incident has actually reproduced — the target alert is firing — before the agent is ever handed the page.
Honest, neutral framing
The agent receives a plain page or ticket. No eval language, no production theatre. It is never told it’s in a harness.