Confidence

A confidence assertion belongs to a specific claim. It records:

• the score and its explicit numeric scale
• whether the score was reported, derived, or assessed
• what the score describes through scope and optional granularity
• whether the score is calibrated
• the entity, activity, or agent that supplied it
• the aggregation method and sample count when known

Explain how a score was produced

A confidence assertion can carry structured provenance for the score itself:

• kind says whether it was reported, derived, or assessed
• method identifies the algorithm or assessment method
• sources names the entities, activities, or agents behind it
• sampleCount records how many observations contributed
• attributes preserves method-specific parameters or intermediate statistics

For example, a transform that was proven to preserve the same calendar date can record:

{
  "score": 1,
  "scale": { "min": 0, "max": 1 },
  "kind": "derived",
  "scope": "transformation",
  "method": "openparser.transform.date-time-format.v1",
  "calibrated": false,
  "sources": [{ "type": "activity", "id": "activity:transform" }]
}

The OpenParser profile stores bounded method-specific proof on the transform derivation. See Transform verification.

Keep source semantics

Retain confidence at the granularity supplied by the source. OCR providers may report confidence for symbols, words, blocks, table cells, or pages. A source-region aggregate still describes OCR recognition in that region; it is not the probability that a downstream extracted value is correct.

Fallback does not change ownership. If page confidence is the nearest score available for a cited cell, keep the assertion on the page entity and reference it as an inherited page recognition score. Do not copy the number onto the cell, evidence claim, or extracted field, because placement would imply unsupported cell- or field-level precision.

Keep scores from different scopes, scales, and providers as separate assertions. Combining OCR recognition and extraction confidence into one number would erase what each score measures.

Reported and derived assertions

Use kind: "reported" when a provider or model returned the score directly. Preserve the original value and scale in namespaced attributes when the source did not use the normalized range.

Use kind: "derived" for an aggregate or calculation. Record the method, sample count, and source entities or activity so consumers can inspect how the score was produced.

Use kind: "assessed" for a human or automated evaluation applied after the value was produced.

Calibration

calibrated: true means the score has a documented probabilistic interpretation for its declared scope. Most provider OCR confidence is uncalibrated and should remain false.

Consumers should compare confidence only when scope, granularity, scale, provider semantics, and calibration are compatible.