loading…
Governed inference
Every AI decision arrives with a signed receipt you can verify yourself. The trust math is a machine-checked theorem. When the model is not sure, it refuses rather than guesses. It runs on your own hardware, air-gapped if you need it.
The plain-English story. Every claim below links to the check that proves it.Each answer carries a cryptographic receipt (DSSE / ECDSA-P256). You re-check the signature in your own browser — no trust in us required.
Eight formulas (the "locked-8") are proven in Lean 4 and re-checked on every build. Honest caveat: the Λ trust score itself is Conjecture 1 — advisory, never shown as "green", never a gate.
We meter the real energy each answer costs and write joules-per-token onto the receipt. When no hardware meter is attached, it says so — we never fake the number.
Nothing runs unless policy explicitly allows it. Untrusted input is quarantined; a state-changing step can require a human to approve it first.
The whole substrate deploys on-prem with zero runtime dependency on outside services — suitable for classified and air-gapped environments. Open-weight models only.
Trust is capped (never claimed at 100%). Below the confidence bar, the system abstains and says why, rather than inventing an answer.