Offline research engine · Evidence engine for human-automation readiness

Average accuracy is a poor answer to a redundancy question.

Reiyah is an offline research engine for comparing perception systems under uncertain reference evidence. It computes bounds on a shared comparison and checks the accompanying certificates. HARBOR remains the broader research program; its architecture instrument and the Engine have separate release histories.

Open ReiyahEvidence repo

What the live surface shows

  • Six kinds, kept apart

    Observation, latent belief, decision, intervention, outcome, and evidence are separate object kinds. A record that omits a kind, impersonates another, or crosses kinds through an untyped reference must fail validation. The separation is a repository assertion about static artifacts, not a claim that future data respects it.

    GA-04 · reiyah.claim.repository.kind-separation · proposed

  • Unknowns that cannot collapse

    Observed, missing, unmeasured, out-of-distribution, sensor-invalid, and abstained are distinct epistemic states. None may validate as a zero, a false, a default, or an alias of another, and none may be rendered as a result. The console shows them as first-class states for the same reason.

    GA-05 · reiyah.claim.repository.epistemic-separation · proposed

  • The joint silent miss

    Two individually accurate actors can share blind spots. Reiyah proposes that modeling dependence between human and automation miss events changes estimated joint silent-miss risk relative to a declared independence model. It is a falsifiable proposition with a frozen comparator and no retained evidence.

    reiyah.claim.scientific.joint-silent-miss · proposed

  • Gate A and the operator

    In the retained September 7 architecture packet, completeness was checked by an offline validator and operator acceptance remained a separate decision on the evidence-index digest. That historical gate does not describe the current Engine's implementation status.

    GA-17 external · historical packet, 7 September 2026