Evidence

We publish the method before the result.

Every claim on this site that carries a number is backed by a study you can inspect. Each one is pre-registered and hash-locked before any evaluation data is opened, and reported with its denominator, its confidence interval, and its failures.

Published studies

01July 2026Internal · not independently reviewed

Government of Canada AIA policy-conformance challenge

Deterministic enforcement of team-compiled rules derived from a frozen snapshot of Canada's Directive on Automated Decision-Making, evaluated on department-disjoint held-out cases.

Read the study

Method

Pre-registration

Thresholds, metrics, denominators, and rule logic are frozen and hashed before any evaluation data is retrieved. The lock file refuses execution if anything changed.

Held-out separation

Calibration, validation, and held-out sets are split by organisation, not by sampling. The held-out set is opened exactly once.

Reported failures

Aborted cases, unmet targets, and under-powered axes are published with their reason codes. A study that misses its planned sample size says so.