# PALO agentic trustworthiness research profile 0.1.0

This additive prototype screens supplied evidence records. It does not authorise actions, authenticate evidence, make legal or fairness findings, or establish conformity. It is not wired into the deployed PALO runtime.

Run the reproducible checks with a supported Node.js runtime:

```sh
node extension/evaluate.test.mjs
```

Import `evaluateProfile(profile, expectedContext, assessmentTime)` from `evaluate.mjs`. Supply the expected case, system version, scenario, Action Claim digest and purpose from a separate trusted context. Pass an explicit ISO assessment time. The JSON fixture is entirely synthetic and uses illustrative thresholds.

Results are `eligible_for_review`, `hold` and `insufficient_evidence`. A passing screen only says that the selected structural and numerical checks over the supplied records passed. The configured dimensions, cohorts, thresholds, evidence acceptance and workload envelope require external review. A caller must not let a subject weaken these choices to evade assessment.

The timing decomposition is an engineering proposal. It does not transform component percentiles into an end-to-end percentile or confirm that a declared safe state was reached. The intervention profile requires stop and at least one of disregard, override or reverse; a real equivalent procedure would need a reviewed profile variant.

The illustrative bias screen assumes independent binomial cases. It uses an envelope formed from separate 95% Wilson intervals, not a calibrated simultaneous confidence interval for the gap. It does not handle dependent trajectories, multiple testing, continuous outcomes, causal discrimination, metric selection or study power. Those require an appropriate statistical protocol. The fixture's sample and five-percentage-point threshold are not recommendations or legal thresholds.

Production admission would require a trusted evidence registry, authentication and revocation, domain-valid measurements, immutable reviewed policy selection, tenant/role controls, representative-user studies, qualified connectors and a versioned compatibility migration. Start any integration in shadow mode. Do not add unknown fields to strict Action Claim 1.4 contracts.

The full supplied oversight draft was reviewed. The full supplied N1240 bias revision has now been reviewed; agentic application remains a PALO research proposal. See the parent analytical dossier and crosswalk for sources and limits.


## Bias reconciliation update, research 0.2

The full 106-page N1240 prEN 18283 sneak-peek revision has now been reviewed. The original evaluator above is retained with its original limitations. Its separate Wilson envelopes do not implement the draft's predefined family-wise statistical protocol.

`bias-evaluation.mjs` and `bias-fixture.json` implement a separate fixed-sample, unweighted independent binary-case screen with simultaneous Hoeffding bounds. Run `node extension/bias-evaluation.test.mjs` to reproduce 30 tests. See the dossier Sections 7A–7F and `bias-crosswalk.json` for source interpretation, equations, assumptions, metric selection and unresolved draft issues.

The independent expected digest must be supplied through a trusted external policy path; a digest bundled with attacker-controlled evidence provides no authentication. Registration dates, sampling reviews and group disjointness are assertions requiring external qualification. The screen cannot deduplicate raw cases, establish a causal effect, validate all bias metrics, or make an execution, risk-acceptance or conformity decision. Do not use the old Wilson result as a substitute for a complete registered statistical family.