Continuity Format v1: signing handoffs between agents
Ed25519 source signatures, one-time HMAC destination binding, and auth-before-parse.
32 MCP tools · 11 platform surfaces · shared registry
Every article on this page links to its source evidence — an ADR, a test run, or a design decision. No fabricated scores. No invented certifications. No AI slop.
The rule is that a post has to be adapted from something already in the repository — an architecture decision record, a test run, a design document. That constraint is what keeps the writing narrow, and it is why there are five posts here rather than fifty. If a claim in one of them sounds surprising, the source link at the bottom of the post is the place to check it rather than take our word.
How the store is built: the bi-temporal model, and the test suite that gates it.
What the system refuses to do — abstain rather than guess, and bind a scope it cannot later widen.
Moving a session between tools without losing the record of how it got there.
Illustrative example — not live data
Each post is adapted from a source document in the repo. If a claim sounds surprising, the source link at the bottom of the post is the place to check.
Ed25519 source signatures, one-time HMAC destination binding, and auth-before-parse.
Valid-time, transaction-time, and why verification is a cascade rather than a score.
Trust over fluency: when evidence is missing or contradictory, the system stops rather than fabricate.
Project, repository, worktree — admitted once and immutable for the lifetime of the connection.
Deterministic scenarios, a single measured feature, and an honest list of what we have not evaluated yet.
git clone https://github.com/DigitalSoftDistribution/zephr.gitThe documentation covers the architecture, tool contract, TrustBench methodology, and trust model in detail. Every claim links to its source.