A safety guarantee tested on one kind of task says nothing about other kinds, so can you trust it elsewhere?

This is about guarantees of the form 'on tasks like the ones we tested, the validator is right at least X% of the time.' The corpus suggests the bound doesn't so much break outside its workload as go silent. It never said anything about that territory. One note separates pointwise, domain-conditional, average-case and uniform guarantees and insists they are not interchangeable: only a uniform bound protects against worst-case tasks, and a validator reported without its scope is unscoped, not safely uniform What scope should a validator's statistical guarantee actually state?. The failure happens when a reader treats a domain-conditional number as if it were uniform.

The edge of the calibrated workload is also hard to see, because nothing signals it. Models optimized for one domain show the same pattern: they are excellent inside it and confidently wrong outside it. Specialization strips out the calibration signals that would flag uncertainty, so the drop is a cliff rather than a slope Why do specialized models fail outside their domain?. A bound calibrated on one workload has the same problem. It reports no warning when you cross the line, so a result from the wrong side looks like one from the right side.

What changes at the edge is often the structure of the problems, not just how hard they are. Frontier reasoning models reach only 20-23.6% exact match on constraint satisfaction problems that require genuine backtracking. Their fluent-looking reflection did not carry over to instance structures they hadn't seen Can reasoning models actually sustain long-chain reflection?. A calibrated workload also carries assumptions nobody wrote down, such as task length, format and difficulty mix. Language models fail in a similar way when background conditions are never brought forward as relevant, and forcing them to enumerate preconditions lifted accuracy from 30% to 85% Do language models fail at identifying unstated preconditions?. The parallel for bounds is that the fix starts with listing the conditions the calibration silently depends on.

There are two practical responses. The first is to attach the scope to the certificate every time, so the number can't travel without its conditions What scope should a validator's statistical guarantee actually state?. The second is to check the actual run instead of relying only on a guarantee earned elsewhere. Verifiers can run alongside generation and intervene only when they see a violation, at near-zero latency on correct runs Can verifiers monitor reasoning without slowing generation down?. Per-action checks also can't express constraints that depend on history, which is why monitoring that tracks state over time matters when behavior composes across steps Can stateless checks ever catch sequence-level constraint violations?.