Galileo scores hallucinations. Arthur monitors models. Aether proves the truth — with cryptographic signatures, multi-model debate, and preserved evidence.
AI hallucinations cost businesses billions. A law firm cites a fabricated case. A medical assistant recommends a contraindicated drug. A financial model invents a merger. Existing tools flag suspicious patterns — but they can't prove why an answer is true, and they can't hold evidence accountable when sources change.
Three independent AIs from different labs answer, cross-examine, and reconcile. A cross-firm verifier synthesizes the hardened answer. No single model can fool the tribunal.
Every answer is backed by an Ed25519-signed warrant with premises, conclusion, confidence, and source hashes. Anyone can verify it — no trust in Aether required.
Claims are checked against live web sources with SHA-256 content snapshots preserved at attestation time. Sources can't rewrite history.
Every verification runs a mandatory adversarial attack. If the falsifier finds a strong counter-case, the warrant is capped — no false confidence.
Medical claims checked against PubMed. Financial claims against SEC EDGAR. Legal claims against statutes. Generic web doesn't pass for high-stakes domains.
Brier scores, per-bucket calibration, regression gates. Every release is gated by FABRICATED catch-rate regression. Trust is measured, not claimed.
If an answer carries an Aether warrant, you can independently verify its premises, check its source hashes, and reproduce its tribunal verdict. No black box. No trust required. Just proof.