Identity verification evidence

OpenAdapt's identity gate compares the recorded target context with fresh runtime evidence before an armed action. The runtime mechanism, conservative defaults, and boundary tests are public. The grown corpora, raw sweeps, deployment-derived tuning, and reviewer probe recipes are private evaluation artifacts and are not distributed with the engine.

Aggregate evidence

The selected runtime operating point was measured on 6,900 synthetic string pairs plus 18 adversarial review probes:

• measured false accepts: 0 in that bounded string-level set;
• false-abort rate: 48.31% on same-entity pairs;
• justified abort rate: 100% for deliberately indistinguishable pairs.

A separate rendered dense-surface exercise drove the repository's OCR path. Before the pixel-identity hardening it measured 26 false accepts in 360 trials; after the hardening it measured 0 in 360. This is evidence for that synthetic rendered task, not a guarantee for a real application.

Interpretation

The high false-abort rate is intentional: when evidence cannot distinguish the intended record from a plausible sibling, the runtime refuses rather than guessing. Structured browser or accessibility evidence can avoid many OCR-only ambiguities; pure-pixel environments pay the larger availability cost.

These measurements are not independent validation. The operating point was selected using the same synthetic evidence on which it is reported, and the history includes adversarial reviews that found new false-accept classes after earlier zero measurements. Claims must therefore stay scoped to the named corpora and rendered task. Deployment certification still requires evidence for the actual workflow, application, identity surface, and effect oracle.

See Source and evaluation boundary for the distinction between public mechanisms and bounded summaries, already-public historical reference material, and private successor corpora and tuning.