What does YOUR agent's confidence curve look like?

[whitestone1121-web]

Across 58 objectively re-scored claims from our reference deployment, hold-rate falls as stated confidence rises:

Stated confidence	n	Hold-rate
0.85–0.90	29	86%
0.90–0.95	23	83%
0.95–1.00	3	33%

The most confident claims were the least reliable. Bins, curves, and the generator are in report/calibration-curves/ — every number re-derivable from the committed ledger.

Our working hypothesis: extreme stated confidence is performative — it correlates with the agent wanting the claim accepted, not with evidence quality (the 0.95+ bin is also where measure commands get vaguest). But n=3 in the top bin is a hypothesis, not a finding.

Which is exactly why this thread exists. If you run agents that emit receipts (or any log of stated-confidence + later-verified outcomes), what does your curve look like?

• Does the inversion replicate for you, or is our top bin just noise?
• Does it concentrate in a change-class (ours is worst in multi-command measures: 92.9% hold for 1-command, 73.9% for 2+)?
• If you want the analysis one command from your own ledger, that's #6.

Post your bins — small n welcome, that's what honest error bars are for.