How to verify an AI answer: a practical guide
If the answer carries a proof, you can check it yourself — in your browser, offline, or from code. Here’s how.
The prerequisite: a proof
You can only verify an AI answer if it comes with the evidence to verify it: a cryptographic signature plus the public key to check that signature against. Most AI gives you neither — so for most AI output, “verify it” simply isn’t possible, and a screenshot is the best anyone has.
Throndar attaches a portable proof to every answer: the answer text, its post-quantum signatures, and the public keys, bundled together. Everything below assumes you have that proof (from the “Download proof” action on an answer, or a share link).
Method 1 — in your browser
Open the public verifier, paste the proof (or the .json file, or open a share link), and the signature is checked locally in your browser against the published keys. A green result means origin and integrity are confirmed: this exact text came from the system and hasn’t been altered. No account, and no trust in the vendor’s servers — the math runs on your side.
Method 2 — offline / air-gapped
For the strongest assurance, verify with no network at all. Because the proof bundle carries the public keys inside it, an air-gapped verifier can confirm the signatures with nothing but the file — the same check, provably independent of any live server.
Method 3 — programmatically
To verify at scale, POST the proof bundle to a public, no-auth endpoint and get back a structured verdict — for a compliance pipeline, an auditor’s tooling, or your own product. A companion endpoint returns a full, filable evidence record. Both are stateless and run the same verification math as the browser.
In every case the honest reading is the same: a pass attests origin and integrity, not the factual accuracy of the answer.
Machines can verify the same proof at POST /api/v1/verify.
Verification attests an answer’s origin and integrity, not its factual accuracy. Algorithm names denote the public standards the primitives are based on (ML-DSA-87 / FIPS 204, ML-KEM-1024 / FIPS 203; Falcon / FN-DSA, FIPS 206 forthcoming), not a FIPS-140 / CMVP validation.