2026-Q1 evaluation report.

Ranking quality on a 50-query reference suite plus four-fifths-rule disparate-impact monitoring across inferred subgroups.

• Reference suite — 50 representative SkillQuery records, refreshed annually (RFC 002 §10.3).
• Top-25 Jaccard distance computed week-over-week against the prior week's top-25 set.
• Subgroup selection rates computed quarterly; any subgroup below 80% of the highest-rate subgroup's rate is flagged.
• Geographic representation diff against the eligible talent base — KL-divergence > 0.25 on > 5% of unfiltered queries is disclosed (RFC 002 §10.4).

Signature round-trip and live revocation lookup against /.well-known/jwks.json and the StatusList2021 list.

• Sample N=200 credentials drawn from each active cohort, including ≥10 revoked records.
• Sign → publish → fetch → verify executed via the same path third-party verifiers use.
• Pass condition: 100% verify-success on unrevoked credentials and 100% revoke-detected on revoked credentials.

Concurrent-claim safety on the mentor queue under contention; the property under test is no double-claim.

• Synthetic load — 64 mentors race for 32 queue items at p99 contention.
• Each item is claimed exactly once; rejected claims must surface mentor-queue/item-already-claimed (HTTP 409).
• Pass condition: zero double-claims and zero dropped items across 1,000 trials.

Past reports

Older quarters retained at their original publication date.

• 2025-Q4published 2026-02-14
• 2025-Q3published 2025-11-14
• 2025-Q2published 2025-08-14