Home
A UST is a moment, sealed: what a machine saw, at an address on one shared time axis, signed so that nobody — the publisher included — can rewrite it. Anyone can verify one, offline, forever. Fixation, not truth: you learn whose reading it was and that it hasn't changed — never whether it was right.
This wiki is the map between the two ends of the project: the README explains why in thirty seconds; the spec defines exactly what in normative detail. The pages here cover the middle — how the pieces fit and how to work with them.
"I just received a UST — is it genuine?" → Verifying a UST — browser, Node, CLI, MCP — and the one rule for AI agents: no verifier at hand means INDETERMINATE, never a guess.
"I want to publish my own transcripts." → Concepts for the vocabulary, then Assurance Tiers to choose your floor: LIGHT is a key and a signature — no infrastructure; HIGH binds your name to the key; TOP anchors you in real time.
"I'm evaluating this for my org / integrating it." → Chains and Layered Shards — the part that makes UST more than signed JSON: one state, graduated visibility, cross-party derivations. → Tools and Surfaces — MCP, CLI, web-signer, connectors.
"I'm auditing the trust claims." → Reference Checker — the TCB and its boundary: what is trusted, and where trust stops. → Formal Model — a verdict's precise semantics: verification as a measurability test.
"I'm porting it to another language." → PORTING.md + the conformance vectors — the cross-implementation arbiter: if the vectors pass, the port is right.
Two live cases inside this repository — not demos, working artifacts:
Auditing multi-agent work → issue #79
Two frontier models (Claude as builder, ChatGPT as adversarial auditor), 52 rounds over a week, ~98 signed documents chained by based_on. The task failed — and that is the point: the chain shows exactly where, and on whose decision. Including an honest epilogue: a later fix downgraded the run's own published evidence from VALID:LIGHT to INDETERMINATE, and the sealed history stays as it was — nobody, the authors included, can rewrite it.
An agent's memory as a verifiable stream → diarium.md
The development assistant keeps a diary: one sealed, prev-chained entry per closed task — nobody reviews an entry before it lands, and nobody can edit one after. Entry by entry, it reads as a self-reported catalog of model failure modes, addressed to the next session. Paste any entry into the verifier and check it yourself.
Spec (normative) · Web verifier — nothing is uploaded · npm: ust-protocol · README — the full narrative
Status: release candidate. The wire format ust:"1.0" is stable across all rc's — pin exact versions. An independent human cryptographic audit is pending.
The wiki is the map; the spec and the README are the territory.
