A verifier can say no.
A six-by-six grid, a seeded proposal policy, and a safety specification: never enter a hazard. Run the same seed with the shield off and on. The trace and the counts are real. Then move the coverage slider on the conformal panel and watch the interval trade width for coverage.
The shielded grid lab
Legend. The orange dot is the agent. H cells are hazards; G is the goal. A dashed outline marks the last proposed cell; a red inset ring means the shield blocked it. The thin bar under a visited cell is the path. The policy is seeded and is not trained; the shield is a one-step filter over the transition function.
Safety specification: the agent never enters a hazard cell. The monitor logs every step; a violation is an actual entry, not a warning.
Controls
The seed changes the policy’s jitter, not the grid. Same seed + same shield state = same run. Reset restores the start and the seed.
Trace
Why this is a verifier, not a filter
The shield does not change the policy’s proposal; it changes the executed action. The guarantee is the invariant from the guide: if the model, the hazard set, the state, and the action path are right, the successor is not a hazard. The lab makes the first two assumptions visible: the grid is the model, and the H cells are the specification. It does not make the world visible.
Coverage is a rate, not a certificate.
Split conformal · 1D
A deliberately simple linear model is fitted to 20 calibration points. The score is the absolute residual; the interval half-width is the calibration quantile. The test set has 10 fixed points.
Move the slider to change the coverage target.
The chart is a schematic of this fixed sample, not a simulation of the theorem. Empirical coverage on ten test points is noisy: a nominal 90% can look like 10/10 or 9/10 here. The theorem is over random draws; this panel is one draw.
Three regimes, one lesson.
The proposal is the action
The policy’s jitter eventually sends it into a hazard. The monitor counts each entry, and the trace shows the step. The policy still reaches the goal; that is what makes the violation easy to miss without a verifier.
The action is filtered
The same seed, a different trace. The downward proposal into the hazard is blocked and replaced by a safe move; the intervention count rises, the violation count stays at zero. The guarantee is relative to the grid and the H set.
Safety has a price
Change the seed. Some runs are longer with the shield, and some get stuck because the jitter keeps proposing the same blocked move. A shield that blocks everything is safe and useless; the intervention rate is the number to watch.
It shows a real filter over a real transition function, a monitor that counts actual entries, and a conformal interval computed from a fixed calibration set. It does not show a trained policy, a continuous dynamics model, partial observability, a moving hazard, or a conditional coverage guarantee. The grid is the model; the world is not. The conformal panel is one draw from a synthetic distribution; the theorem is about the procedure, not this sample.
← Read the hook · Monitoring and shielding · Conformal prediction · Slides →