THE GATE
Every AI-written change gets one verdict. Every SKIP says why.
Dependencies, build, lint, tests, replay fixtures and a headless Gazebo scenario, run in a pooled workspace that never touches your tree. The receipt on the bench on the home page is a real run.
EVIDENCE BUNDLE a3f9c2…
- MCAP slice of the run
- Foxglove layout, ready to open
- verdict, logs and source JSON
- offline report: no server, no network
What the gate gives you
- One-command setup. lcc init reads your repo and writes the config.
- Your tree is never touched. Pooled, warm workspaces rebuild in 1.6 s instead of 7.3 s.
- Evidence bundles. Content-addressed, with an MCAP slice and a Foxglove layout.
- On the pull request. In CI the bundle goes to an evidence branch and the verdict to a PR comment.
- Proof comparison. Original and candidate verdicts side by side with lcc compare.
- Human GO. Hardware-in-the-loop waits for an explicit GO or NO-GO from a person.
the receipt on the bench · robot@4730087c · humble/amd64 · 25.4 s · 9 of 10 stages ran
- S1 workspacePASS
- S2 dependenciesPASS
- S3 buildPASS
- S4 lintPASS
- S5 testPASS
- S6 target_buildSKIP
- S7 replayPASS
- S8 scenarioPASS
- S9 evidencePASS
- S10 forwardPASS
S6 SKIP · no deployment targets declared · not run, so not proven
Seeded, paired experiments
Two revisions, the same seeds, behaviour budgets you set.
launchcommit.experimentsFault injection and search
Inject a bounded fault, find the revision that broke it, shrink the fault.
experiment_searchAgent gym
Measure which agent or model works on your stack, in isolated KVM machines.
flightops_gymNEXT DRAWER