Concepts

Transparency & Evals

A verdict you can verify is only half the promise. The other half is whether the verdict is any good. Fidacy publishes that too: accuracy, the direction of its errors, and how much of the decision runs on the deterministic hot path are exposed on a public endpoint anyone can read. These are calibration signals, not a substitute for a customer's own validation of the controls it deploys.

The public transparency endpoint

GET /v1/transparencyreturns the engine's current evaluation metrics, computed from a labeled, human-reviewed ground-truth set. No authentication, no account.

GET/v1/transparency
{
  "labeled_cases": 0,
  "accuracy": null,
  "cost_weighted_error": null,
  "error_favors_fail_safe": null,
  "hot_path_share": 0,
  "model_version": "…",
  "generated_at": "…"
}
FieldWhat it tells you
labeled_casesSize of the human-reviewed ground-truth set the metrics are computed against.
accuracyShare of cases where the engine's decision matched the human label.
cost_weighted_errorError weighted by how costly each mistake is, so a wrongful approve counts far more than a wrongful review.
error_favors_fail_safeWhether the remaining errors lean toward the safe side (review/deny) rather than wrongful approve. The invariant the whole engine is built around.
hot_path_shareFraction of the decision made by the deterministic, replayable hot path (versus the optional reasoning layer).
model_versionThe engine model version represented by the aggregate report. Check the live endpoint rather than treating this example as current performance.
Errors that favor fail-safe. Accuracy alone can hide a dangerous failure mode. A model that is wrong by wrongly approving is far worse than one that is wrong by asking for review. cost_weighted_error and error_favors_fail_safe make that direction public, not just the headline number.

How the engine improves: the eval loop

The deterministic core does not drift on its own. It improves through a human-in-the-loop cycle, and the result is published, not asserted:

  • ·Capture. Real assessments are captured as cases, structured signals only, no payment credentials or PII.
  • ·Label. A human reviews a case and assigns the correct outcome, building a ground-truth set.
  • ·Evaluate. A pure, deterministic eval engine replays the cases and computes the metrics above, including the cost-weighted error and the fail-safe direction.
  • ·Propose. A change to the rules is proposed against the ground-truth, with its expected effect on the metrics. A human approves or rejects it; nothing ships itself.
  • ·Version. An accepted change bumps the policy_version, which is stamped into every signed verdict from that point on.

Tie a verdict to a published version

Every signed Risk Payload carries the policy_version it was produced under. Combined with the public metrics, an auditor or counterparty can take a verdict from months ago, read which model version produced it, and see the accuracy and error profile of that version, without contacting Fidacy.

A signature proves who signed a payload and whether it was altered after signing. The transparency endpoint publishes an aggregate snapshot. A small or zero labeled set is not a statistical performance claim.