Når AI får handlekraft, trenger vi bevis
Vi skal ikke måtte stole på agenten for å kunne stole på resultatet. Tilliten må ligge i grensene som er satt på forhånd, kontrollene som kjører, og beviset som fanges underveis.
Blog
Field notes on cryptographic timestamping, advanced electronic seals, AI evidence, and the regulations driving them.
Sigill og NSMs kryptografiske anbefalinger: Hvor står vi?
NSMs kryptografiske anbefalinger handler om mer enn valg av algoritmer: nøkkelbeskyttelse, evnen til å bytte kryptografi, og forberedelsene på det som kommer etter dagens løsninger. Vi har sett på hvordan Sigill står seg.
Your PDF never leaves your systems: delegated sealing in SDK v0.3.0
SDK v0.3.0 seals PDFs with an embedded PAdES seal — up to the archival B-LTA level, with qualified timestamps — while transmitting only a 32-byte digest. The PDF stays in your environment, the guarantee is enforced in code, and the code is open source. Validated against the EU's reference validator.
JAdES is live: ETSI-standard seals for JSON
JSON is what AI systems and APIs actually produce — and it now has a first-class seal format on sigill.ai. Detached JAdES seals per ETSI TS 119 182-1, hash-only by default, with optional qualified timestamps and a post-quantum ML-DSA-87 hybrid signature. Available in the API and both SDKs today.
Timestamp every AI conversation from your browser
A Manifest V3 browser extension that automatically creates RFC 3161 timestamps for every Claude, ChatGPT, and Gemini conversation turn — without sending a single word of your conversation to Sigill.
Sigill is now public
Proof only works if it exists before it is questioned. We are making Sigill's independent cryptographic evidence layer public today.
Bring your own TSA: use your own timestamping authority in Sigill
Paid tenants can now connect an existing RFC 3161 timestamping authority to Sigill, keep their provider policy intact, and use Sigill for routing, evidence, sealing, and verification.
Sigill now speaks MCP: timestamp and seal from inside Claude
Sigill's API is now available as a remote MCP server. Claude can timestamp a document, verify a TSR, look up a hash, seal a PDF, verify a sealed document, and check your plan quota without leaving the conversation — using the same API key you already have.
Sealing Claude Code sessions with the Sigill SDK
The Sigill SDKs are not just for production AI calls. This post shows how a small Claude Code Stop hook can use sigill-python to turn local AI coding sessions into sealed, verifiable evidence envelopes.
SDKs are out: sigill-python and sigill-dotnet
Two open-source SDKs for building, sealing, and verifying AI evidence envelopes — Apache-2.0, byte-compatible across languages, with the canonical-JSON, hash-binding, and RFC 3161 plumbing already done. The pattern from earlier posts, now installable from PyPI and NuGet.
Hardening AI evidence: input completeness, sequence integrity, and the AI event ledger
The simple agent timestamps every turn. That's necessary but not sufficient. This post is about what makes the difference between a log that timestamps things and a record an adversarial reader can't unwind.
Timestamping every release: a GitHub Actions workflow for tamper-evident builds
Stamp every artifact your CI produces with an RFC 3161 timestamp before it leaves the build runner — so a year from now you can prove what shipped, and when, without trusting your own logs.
Timestamping every prompt: a 200-line agent that seals what your AI says
A working ~150-line Python agent that cryptographically timestamps every prompt and every model response, producing tamper-evident records you can verify with one openssl command. For teams whose AI systems will eventually face an auditor, regulator, or counterparty asking 'prove what your model said on day X' — this is the minimal pattern that answers the question.