AI method: recorded run, failed attempt and limits

The retained private store contains one completed extraction run for one C-014 source document. It is not evidence of an absolute historical provider-request figure, so none is published.

Completed run

Uncached input tokens

3

Cached input tokens

17,682

Output tokens

1,053

Recorded cost

USD 0.00161784

Recorded latency

10,017 ms

The cache read shows that a prefix was reused. The request which primed the cache is unestablished; no origin or historical request total is inferred.

Failed attempt on 13 August 2026

A later authorized request returned an incomplete response. The adapter rejected it as provider_response_incomplete. It produced no stored extraction run and no stored claim.

The adapter was subsequently corrected so a future authorized call can preserve content-free usage, cost, latency, incomplete reason and rejection details while still failing closed. That correction does not reconstruct the missing accounting or establish the cause of the failed attempt.

What has and has not been demonstrated

The current recorded repository suite passes 178 tests covering schemas, storage, evidence and publication gates, adapter behavior, static export and leak scans. Engineering tests do not measure model extraction quality.

No accuracy, precision or recall result exists. The scored golden set is empty, so there is no basis for an extraction-quality percentage or a comparative model claim.