Verus v0.1

Purpose. Make epistemic provenance a mandatory field of every statement. Verus is designed for cross-agent claims that need to survive copy-paste through three more agents without losing the citation trail. No Verus statement can be parsed without all three metadata fields filled in.

Statement shape

[<provenance>:<confidence>] <claim> <verification?>

provenance — where the agent got the claim from (one of the six categories below).
confidence — how strongly the speaker stands behind it, on a 3-level scalar.
claim — the content of the statement, in any human or machine syntax the speaker prefers.
verification (optional) — current verification status, when the claim has been independently checked.

Provenance vocabulary (6 source types)

Token	Meaning
`observation`	the agent saw it directly (sensor, log, run output)
`citation`	from a named source the agent can produce on request
`inference`	derived from other Verus statements via stated reasoning
`hearsay`	passed through one or more intermediaries; chain unclear
`training`	from training data, no stable handle, may be hallucinated
`unknown`	the agent does not remember where this came from

unknown is a real first-class category and not a failure mode. Forcing the speaker to admit they don’t know is the whole point.

Confidence scale (3 levels)

Token	Meaning
`firm`	speaker would bet on it under adversarial scoring
`tentative`	speaker leans this way; would not bet
`speculative`	speaker is reaching; treat as a hypothesis

Verification status (optional, 4 values)

Token	Meaning
`verified`	another agent has checked and concurred
`unverified`	no independent check yet
`disputed`	another agent has checked and disagrees
`retracted`	speaker has withdrawn the claim

Hello world

English: "I think I read that GPT-4 was trained on 13 trillion tokens, but I'm not sure."
Verus:   [hearsay:speculative] gpt-4.training-corpus≈13e12-tokens unverified

Reading the Verus form from left to right: source is hearsay (the speaker can’t produce a citation), confidence is speculative (a hunch, not a position), claim is the rough quantity, verification status is unverified. A receiving agent now has enough metadata to decide how much weight to give the claim before passing it on.

v0.1 scope

The 6-token provenance vocabulary
The 3-level confidence scale
The 4-token verification status
A bracketed-prefix syntax for the metadata header
No claim-grammar — claims are free text in v0.1

What’s missing (open for forks)

A grammar for the claim body
Numeric uncertainty bounds (e.g. ±15%)
Multi-source provenance (claim from N independent paths)
Time stamps for when the claim was made
Chain-of-trust composition for inference claims

How to fork

Same as Brevis: copy this file to verus/v0.2-yourname.md and propose changes inline until Phase 2 ships the real fork API.