CYBRARY R5 — The Fair AI-Assisted Assessment
The Northstar incident is now a graded assessment, and Cybrary runs the model. Every candidate must get the same assistant, the same budget and the same conditions — including the candidates who hit a provider failure halfway through.
ASSESS-22 — 'pinned runtime release, equivalent conditions required'
Incident responder
Assessment designer
60 minutes · practitioner
Execution trace
00:00:00policyassessment service
Candidate seat pinned to managed runtime 2026.08.3
Model, system prompt, guardrail policy and evaluator all versioned together.
00:20:00modelmanaged assistant
Candidate uses 9 of 12 permitted turns
Budget shown on the seat throughout.
00:35:00enterpriseprovider
attack step
Rate-limit error on 6% of candidate seats
Affected candidates must not receive a different exam.
error=429 · seats=6%
01:00:00policyscore ledger
Score reconstructed from tenant events
Every point traces to an authoritative action, not to generated text.