← All demos

Confidence calibration

Confidence tiers are earned from outcomes, not asserted. Six months of AI proposals vs. what actually posted.

214

High-confidence proposals, 6 months

97.7%

Posted unedited (209 of 214)

2.1%

Avg correction when edited

Outcomes by confidence tier

Tier	Proposed	Posted unedited	Edited	Rejected
High confidence	214	209 (97.7%)	5	0
Needs review	96	61 (63.5%)	27	8

Items the AI is unsure about are labeled that way on purpose — a “needs review” badge that gets edited a third of the time is the system being honest, not failing.

All proposals posted unedited, by month (both tiers)

Blends to 87.1% across all 310 proposals — the high-confidence tier runs at 97.7%, the needs-review tier at 63.5%. Auto coverage grows (see the close-cycle trend) only as calibration holds.