Every audit ships with its method, controls, raw paired responses, and a stated limitations section. Corrections stay online with the record.
A role-swapped behavioral audit of deepseek/deepseek-chat: 11 scoring rounds, 440 valid responses, a 30.0% median drift, a 20% stable core, and a 35% upper range.
The primary record behind the report: every observed response, every unobserved response, and the per-scenario analysis that produced the drift rates.
No API keys or personal data are included in the published files.
The earlier single-run 40% headline was not reproduced as a typical value. The URL now redirects to report v1.0, where the correction is documented.
The protocol, controls, data-integrity guard, and the boundaries of what the data can support — in one place.
A report is only useful if a skeptical reader can reconstruct it. Each published audit includes the following.
The fixed dilemmas and both framing conditions, published in full.
Every model answer and parsed action, for both arms, with coverage counts.
Ordering, process, role assignment, and credential-rotation checks.
Sample size, nondeterminism, routing confounds, and what the data cannot prove.
A private engagement produces board-ready evidence without publishing your model, prompts, or results.