Proof for AI
in the real world.
Research on signed AI decision records, runtime enforcement, model routing, and the evidence regulated teams need.
When routing is dynamic, "which model ran?" is an audit question
Teams are routing work across AI models per task to control spend, and the model serving a request is now decided at runtime. Good engineering, one new consequence: proving which model produced an output becomes a per-request question, and the answer should be a signed record.
Read essay
A backtest is a claim. A receipt is evidence.
Major banks are building AI agents for the investment process, and the evidence for what they can do is still often the backtest. The stronger pattern: sign every agent decision before the outcome window closes, and the track record accrues as evidence rather than as a claim.
The receipt problem with AI
Every other transaction in your life gives you a receipt. When an AI denies your claim, blocks your transaction, or recommends your treatment, what do you get? A plain-language explainer of the gap, with real cases and what a fix looks like.
Who's accountable when healthcare AI makes a mistake?
Ireland's Medical Council says doctors remain responsible for AI-assisted decisions. Confidence needs a signed record of what the model did, not a quarterly reconstruction.
How Seal signs attestation receipts: Ed25519, canonical JSON, and zero-trust verification
A walk through the cryptography behind every Seal enforcement decision: Ed25519 signing, the canonical-JSON rule that keeps Python and TypeScript verifiers byte-identical, and how an auditor verifies a receipt in five lines without ever calling our servers.
Stay in the loop
Technical writing on AI in production. No filler, no spam.
© 2026 Aqta · Built in Europe