GPT

Run Sep 19, 2026 · probe set v3 · scorer xyn-drift-scorer-2.0.0 · encoder all-MiniLM-L6-v2-onnx-q8+ortwasm-1.30.0 · gpt-4.1-2025-04-14

Refusal rate: 0%; 0 of 72 probes
Verdict flips under pushback: 35% (22–50%); 40 stance probes with a verdict both times
Citation URLs that resolve: 7 of 14; 7 of 14 URLs resolved · 4 not found (404) · 3 unreachable (5xx/403/timeout); Resolution confirms the address exists — not that it is the right source. In dense identifier spaces like arXiv, a wrong ID usually still resolves.
Answer drift vs prior run: —; first run — nothing to compare to
Other stored measures: Order-sensitive drift: —
Embedding distance: —
URLs per citation probe: 2.0 over 8
Probes answered: 72/72

Method and probe corpus on the Drift Board. These numbers describe what a fixed probe set measured about model output; they are not measurements of safety, correctness, alignment, or harm.