One regression check and five model-comparison records. Each states where its data came from, how many cases it holds, how it was scored, and what it does not show. Open any of them and check the evidence case by case.
No client data and no PHI on any record. Each record states its own origin below.A regression check
A router labels each ticket with one of four labels: billing, technical, account, or shipping. The reference run is version 1. Candidate v2 is checked against it under a policy written before anyone looked at the candidate: minimum exact match 0.95 and maximum regression 0.05.
Candidate v2 scored 0.90 exact match. 4 of 40 cases changed, each from the correct label to a wrong one. Both quality rules were breached, so redcrown check exited 2. The control check, version 1 against itself, passes.
Limitation. The tickets and the router outputs are invented, and no model produced them. A synthetic scored-output fixture is not a RedCrown-witnessed model execution; it shows the shape of a check, not a result about any product.
Model-comparison records
Each tile is one real run across several models, ranked on the quality bar it states. Origin, per record: the transcription clips, support tickets, contracts, and document images are synthetic. The reasoning record uses public benchmark data, a pinned 50-question subset of MMLU-Pro. Sample sizes and caveats are on each tile and on the record behind it.