A founder's letter.

Ibuilt Verdict because I kept watching the same conversation play out, almost word for word. A general counsel would ask their AI team, "Can you prove what the agent did?" The AI team would point to a trace, a dashboard, a log. The GC would push: "Will that hold up in a deposition?" And the room would go quiet.

That silence is the gap. It's where the first billion-dollar AI verdict gets decided. The question in that courtroom won't be whetherthe agent did it — by the time you're in front of a judge, the action is established. The question will be what evidenceyou have to prove what it did, and whether that evidence survives cross-examination. LangSmith traces, Datadog dashboards, and Sentry logs were built for engineers debugging at 2am. They were not built for opposing counsel, regulators, or underwriters. They don't survive Daubert. They don't satisfy FRE 902(14). They don't clear a chain-of-custody review.

So we built the thing that does. Verdict captures every action your AI agent takes and seals it into an immutable Sealed Evidence Record — content-addressed hash, Merkle root, anchored to the Sigstore public transparency log within seconds. One sealed record renders into a SOC 2 dossier, an EU AI Act Article 12 log, an FRE 902(14) court exhibit, and an insurer underwriting packet. Same record. Five audiences. Zero rework.

A note on Anthropic and Claude. When Anthropic shipped Claude for Legal — twelve plugins, eighty-plus agents, twenty connectors going live in a single drop — the gap they did not close was forensic authentication of the AI output itself. Source attribution and conservative privilege defaults are in their README. Cryptographic sealing, chain of custody, and FRE 902(14)-ready custodian packets are not. That is by design — Anthropic ships the workflow surface and trusts the ecosystem to ship the specialty layers. Verdict is the specialty layer that closes the evidence gap underneath Claude for Legal. Our remote MCP endpoint speaks the same protocol Claude already understands, and every sealed record carries a public Sigstore Rekor anchor verifiable in seconds. We are building so that any firm running Claude for legal workflows in Q3 has a real evidence layer to point to when the underwriter, the regulator, or opposing counsel asks for one.

The infrastructure is intentionally open. The Sealed Evidence Record specification is Apache 2.0, forever. We hold three USPTO patent filings — one non-provisional on the foundational sealing system, two provisionals on adjacent agent-governance primitives — but the licensing posture is RAND-Z for any conformant implementation. The patents exist to keep the standard from being captured or forked, not to extract rent from anyone using it.

What I tell every customer, every partner, every investor: the seal is the evidence, the standard stays open, the category gets locked in eighteen months — by us, or by someone else.If you're building autonomous AI right now, you already have the problem we built Verdict to solve. The only question is whether you want the evidence ready before the first lawsuit lands, or after.

If any of this matters to you — as a builder, a buyer, a partner, or an Anthropic collaborator — write to me directly. Fastest reply you'll get from any founder this week.

— Shayne