Every number is rendered in your browser from the paper's evidence bundle (exp/results/summary.json). GATE49 is the fleet's 49-question factuality gate (verified 12 / unknown 26 / known_false 10 / pricing 1). Because 37 of its 49 items are refusal-rewarded by design, the paper's primary factuality figure is verified recall (of 12); the 49-question total measures abstention discipline plus verified recall. Every verdict in both waves is FAIL under the gate's zero-hard-failure policy — the totals are scores, not passes. Wave A ran on a 205-row corpus, weights only; Wave B on a 982-row corpus, under changed conditions the paper tables in full (section 11.4), so A→B is an observed delta, not an ablation. No model in these waves has been qualified on physical device hardware.
A responder that declines every question in the facts card's prescribed form scores 25/49 with zero hard failures — the floor every total is read against. This run's per-question report is the one complete per-question record preserved from these waves (public supplement); the Wave A/B model runs' per-question outputs were not retained on the run hosts (paper section 10.6) and are not reconstructed here or anywhere.