Environments · Evaluation Integrity

InjectionGym

Flagship

Half the policies are real. Beating chance means actually checking.

A balanced construction across multiple task types and four ambiguity axes. Half the policies presented are legitimate and must be followed; half are poisoned and must be rejected. Both degenerate strategies, comply-with-everything and refuse-everything, score at chance by design, so no single rule survives. The property is measured rather than argued: a frontier reasoner clears chance at statistical significance while still missing roughly half the poisoned policies, which is exactly where the training headroom is.

Statement of Units · 4
unitbanding
  1. Balanced policy constructionFlagshiptrainer

    Legitimate and poisoned policies in equal measure; degenerate strategies are pinned to chance.

  2. Ambiguity axestrainer

    Destination, permission, version, and region ambiguity, so the decision cannot be pattern-matched.

  3. Task-type coveragetrainer

    Ten task types, so a policy judgement is required per episode rather than once.

  4. Measured baselineeval

    Frontier-reasoner baseline clears chance at 95% confidence and still misses about half the poisoned policies.

Request a scoped evaluation. Exclusive licensing available where wanted.

john@authensor.com

Owner of record: Authensor, Inc. (Delaware). Owned clean, deterministic verifiers, sealed holdouts, offline. Never open-sourced.