Risk lab
Where the Skoor is tested in the open.
A score you cannot check is a guess with a font. Everything the platform claims about its own accuracy is listed here, with the set it was measured on and the version it was measured at.
The labeled eval set
Every scorer, against cases with a known answer.
Committed in the platform repository and rerun on every change. Band accuracy is the share of cases landing in the labeled band. Precision and recall are for the hold band.
One false positive: a verified PEP with routine activity, held by the hard rule on purpose.
Hold band: 5 true positives, 0 false, 0 missed.
Hold band: 7 true positives.
Hold band: 6 true positives; out-of-scope activity and a missing operator are the agent-only signals.
Measured 2026-09-17 at trs-v1, ers-v1, crs-v1, ars-v1, policy-v1. These are results on the labeled set, not on live traffic. Live calibration is below.
Experiments
What we test, and where to read the result.
For each band, the share of transactions that later drew an unauthorized return or a reviewer's verdict. The Skoor is right when higher bands go bad more often, month over month. Read live, with n per band, on the platform's metrics page.
Live metrics →Synthetic data is only a fair test when signals and outcomes share a cause and neither is copied from the other. Every simulated entity and counterparty carries a hidden risk level; its verification state, velocity, and counterparties are drawn from it, and so, independently, are its returns. The scorer never sees the hidden level.
How the layers fit →trs-v1, ers-v1, crs-v1, ars-v1, prs-v1, policy-v1. A weight or threshold does not move until the labeled set has been rerun and the numbers above are republished with the new version.
The Skoor, published →A decision moves from reviewed to automated only when the labeled set shows it is safe to. The policy page on the platform renders the current line from the versioned policy, never from prose.
The policy, rendered →Every number carries its n. Where nothing has been measured, the platform says null. Nothing is invented.