Verifier
The verdict, with its reasons attached
A pass/fail chip asks to be trusted. This lists the assertions that ran, which one fired and the step it fired on — so a verdict can be argued with instead of believed. The failing card below is wired to the scrubber: open an assertion and jump straight to the step.
gridworld.goal_termination
4 passedDecides whether an episode ended by solving the task or by running out of horizon.
4 passed, 1 skipped with its reasongridworld.goal_termination
2 failedDecides whether an episode ended by solving the task or by running out of horizon.
done=true at step 59 with observation.x=5, observation.y=6 — the episode ended on the horizon while adjacent to the goal, never entering it.
Expand an assertion, then scrub to its stepgridworld · 9×9 · goal (7,6)Seed 17TimeoutReturn 41.17
Step 43 / 59Reward 0.604move west← → to step
Jumps here when an assertion is openedDisagreement
Where the human and the verifier part ways
Agreement teaches nothing. Every disagreement is either a broken verifier or a broken label, and both are worth an operator's attention — so the rows that agree stay quiet and the ones that do not are tinted and carry the verifier's confidence.
traj-unbounded-0092Humanfail·Verifierpass0.71Verifier scored return only; the human saw the agent never enter the goal.
traj-clipped-0184Humanpass·Verifierpass0.96Agreement — no review needed.
traj-clipped-0201Humanpass·Verifierfail0.52Verifier flagged a wall touch the renderer shows as a corner cut.
traj-curriculum-0043Humanfail·Verifierfail0.88Agreement — timeout on the horizon.
workspace/evaluation/disagreement-row.tsxgridworld · 9×9 · goal (7,6)Seed 17SuccessReturn 8.93
Step 22 / 22Reward 5.180move westGoal← → to step
The trajectory the clean card above was judging