What does a calibrated confidence score mean when an agent reads product data?

Most extraction tools return a score that measures how many fields were filled, or how sure a model sounded. Neither tells an agent how often results at that score are actually right, which is the only thing it needs to know before acting.

Calibrated versus uncalibrated

An uncalibrated score orders results but does not say how often they are correct. A calibrated score is fitted against labelled outcomes so that, across many results, the stated number matches the observed rate of being right. That is what makes a fixed threshold meaningful: "act at 0.7 or above" becomes a statement about expected error rather than a guess.

The measured figures, with their limits

Legibility's calibration was measured on a held-out test split of 63 labelled items, in evaluation run iso-63-2026-07-05:

• Precision at the 0.7 gate: 1.0, with a Wilson lower bound of 0.832.
• GTIN recall: 1.0. Adversarial rejection: 1.0.
• Expected calibration error across the full range: 0.19.

The honest reading is that the score is dependable at the gate and looser in the middle of the range. Sixty-three items is one evaluation run, which is a claim rather than a credential, so the sample size is published with the number. Every response carries calibration_version, so a stored score stays auditable after a recalibration.

How an agent should use it

1. Gate on confidence. At 0.7 or above the object is treated as trusted; below it, treat the read as unknown.
2. Use field_confidence when one field matters more than the rest, for example acting on the price only when the price field clears your own bar.
3. Adjust the gate with min_confidence (0 to 1) when your task tolerates more or less risk than the default.
4. Store legibility_id and calibration_version with any decision you make from the read.
5. Close the loop with report_outcome, which records whether the answer led to a real result and is never billed.

Why the gate is also the billing unit

With an API key, Legibility bills only trusted reads: a product returned at confidence 0.7 or higher. A miss or a below-gate read costs nothing and does not use quota. Charging for misses would reward the opposite of an honest gate. The method is described on the home page and in the read_product answer.