Skip to content
Archive
No. 008GovernanceProtocol

Datum: A Versioned Canon Layer for Multi-Agent Systems

Protocol specification v0.2: claims, authority tiers, and supersession as a first-class operation

Datum is a versioned canon layer for multi-agent systems: a protocol for recording what is true, since when, and on whose authority. Existing agent primitives are append-or-retrieve and cannot detect that two things are both asserted and cannot both be true. Datum adds authority tiers, declarative constraints, and supersession, the primitive that distinguishes a deliberate change to canon from an error: a retcon in narrative, an amendment in policy, and a superseded decision record in engineering. Agents may read, check, and propose; only an authorized principal may commit. The repository includes the specification, a threat model, the GREATGAME benchmark specification, and a conformance suite.

Date
JUL 2026
Read
9 min
Depth
Depth: Full immersionFull immersion
Cluster
Governance
Department
Research Engineering
Provenance
Datum protocol repository · Apache-2.0
Authors
Third Signal Labs
About 9 min left

Thesis

Agents get write access to proposals, never to truth.

Existing agent primitives either append or retrieve. None can detect that two things are both asserted and cannot both be true. Datum adds authority tiers, declarative constraints and supersession, and gives agents three tools, read, check and propose, but never commit.

01 · Argument

Supersession is the primitive nobody has.

Version control handles change. Canon needs deliberate change to what was previously true, with both states still meaningful: a retcon in narrative, an amendment in policy, a superseded decision record in engineering. A supersession marks the old claim superseded, never deletes it, and raises no conflict. That is what separates intent from error.

Evidence

A threat model names ten attacks, and three are still open.

Corroboration collusion, provenance forgery and supersession abuse remain open against v0.2 and carry into v0.3 as normative work. A conformance suite covers six levels, from read to federation, and a mock adapter is required to fail the upper levels by design.

Diagram · Four tools, one trust boundary

datum.read, datum.check and datum.propose are open to agents. datum.commit needs an authorized principal and rejects any unresolved blocking violation. Extraction from prose to claims is a pluggable adapter, and machine-extracted claims default to proposed, never binding.

Signal 01Weight 5 of 5Proposed
“Agents get write access to proposals, never to truth.”
Signal 02Weight 4 of 5Proposed
“Low drift at low coverage is not conformance. It is ignorance.”

Tension

Collapsing duplicate proposals makes corroboration look like consensus.

The spec counts fifty agents proposing one fact as one proposal with fifty corroborations, and treats the count as signal. The threat model identifies that as a weakness: colluding sources amplify each other. The planned fix is lineage-weighted corroboration in v0.3.

Implication

Rights holders and licensees can share one canon with asymmetric rights.

Federation separates the right to read from the right to supersede. That is the workflow PDF style guides and approval emails carry today, and the spec names it as the commercial surface.

Companion documents

The threat model, the GREATGAME benchmark specification and the conformance suite sit in the same public repository, under Apache-2.0 and CC-BY.