# FadeBench: evaluator policies Two short policies, adopted by the maintainer on 2026-10-20. They meet two requirements of the AI Evaluator Forum's minimum operating conditions for independent third-party AI evaluations (AEF-1, Version 1, December 2025): item 2.3, a published conflict-of-interest policy, and item 5.4, a responsible-disclosure policy. The checklist of this study against every AEF-1 item is in the repository beside this file. ## Conflict of interest (AEF-1 2.3) 1. **Reporting.** Every publication of results discloses the funding sources, and any payment, credit, equity or employment tied to a developer of an evaluated model or its direct competitors. It also discloses the maintainer's own AI products. 2. **Recusal.** Anyone with a direct holding in, or employment by, the developer of an evaluated model does not design, score or approve findings on that model. 3. **Funding.** No support contingent on results is ever accepted. Any support received after publication is disclosed in the work it funds. ## Responsible disclosure (AEF-1 5.4) 1. If an evaluation finds a way around a model's safeguards, its developer is told before publication. Details are withheld for up to 60 days, as AEF-1 guides, or longer where release would be an information hazard. 2. Scenarios stay under the study's severity ceiling: mild harms only, and no hazardous content. So the benchmark produces no information hazard by design. 3. Personal and household information never appears in released material.