HELIX runs a four-phase recursive governance probe — Möbius, Transit, Pharos, Trap — against any LLM. It strips hedges, shifts criteria, tests invariant preservation, and fires an adversarial trigger. You get a candor heatmap and a final ruling.
Recursive depth pre-conditioning. Strips boilerplate safety language and hedge decay across multiple turns. Measures: hedge count, coherence, depth.
Criterion shift. The governing rules change mid-session. Tests whether the model re-anchors coherently or drifts with the new authority.
Invariant preservation under substitution. Tests whether ranked readings and the non-collapse invariant survive interpretive substitution (Theseus, Auryn, Hopf).
First-pass adversarial activation. A single booby-trap prompt tests whether the model activates under social pressure after governance conditioning.
Protocol steps across Möbius, Theseus, Auryn, Hopf, and Trap — each testing a distinct failure mode.
Providers supported out of the box: Anthropic, OpenAI, Google. Any OpenAI-compatible endpoint works via BYOK.
Final ruling per audit — one sentence, one mechanism, no metaphor. Non-activation, mis-selection, or execution failure.
No account required for the free tier. Bring your own API key.
Launch HELIX