Recorded run replay

Same detector. Different delivery.
One ships broken.

Two coding agents run the same task on the same seed, both watched by the same verifier under the same strict rules. The only thing that differs is how each one is told about a problem it introduces.

This page replays a real recorded run captured on 2026-07-12. Nothing here calls an API or a live agent. Live runs are available on request: get in touch.

Loading…
Left channel
Told at commit
The finding arrives as a CI list, once, at the commit step.
Silent breaks0unflagged this step
Tokens · this run0context carried
Outcomerunning
Right channel
Told in-loop
The repair is pushed into the next step as a directive.
Silent breaks0unflagged this step
Tokens · this run0context carried
Outcomerunning

Prior runs · study aggregate · not this run

24/24
recovered when the repair was pushed in-loop
1/24
recovered when delivered as a CI list at commit

These are aggregates across a seeded study, shown for context. The lanes above are a single recorded sample and do not, on their own, prove these numbers.

What the counters mean

Left · silent breaks0
Right · silent breaks0

A silent break is a step where the verifier flagged a live violation and the agent was not told about it that step. It is the same measure on both sides.