A mechanical world model for agents: decide, act, report, in the other order
A reader's summary of our preprint: give a proposed change a canonical identity, store its inverse before applying it, pass it through a three-valued gate, and sign the outcome, refusals included. Abstract, scope and a link to the DOI.
Today an agent that changes a system usually works in the order decide, act, report: it does the thing, then tells you what it did. Our preprint describes a model that inverts that order for the whole stack. This page is a short reader’s summary; the paper is the authority, and it is open access under CC BY 4.0.
Paper: A Mechanical World Model for Agents: Quotient Identity, Three-Valued Gates, and Signed Reversibility, preprint, 2026-08-30. DOI: 10.5281/zenodo.22168558 (record on Zenodo).
The idea in four steps
- A proposed change is given a canonical identity, so the same change always has the same name.
- Its inverse is constructed and stored before the change is applied.
- A gate looks at the change and returns one of three verdicts. A failure to evaluate is deliberately outside those three arms; it is not folded into a pass or a fail.
- The outcome, including refusals, becomes a signed record that a third party can re-check offline, without trusting whoever issued it.
What the paper contains
- Nine layers and six axioms, with the mathematical skeleton that holds them together.
- Canonical identity as a non-invertible quotient projection.
- A verdict codomain of exactly three arms.
- A domain-separated hash chain that binds records to a public log.
How it states its own limits
According to the abstract, 19 of 27 components are running code. Every layer and axiom carries an implemented, planned or vision tag, and falsifiers are stated before the argument. Nothing is claimed as an invention: every component has a named ancestor, and only the combination is offered as ours.
Why it matters for agent work
The practical reading is narrow. Reversibility is only useful if the inverse exists before the change does, and a record is only useful if someone who does not trust you can check it. The paper works out what those two requirements force on the design.
Read the paper at doi.org/10.5281/zenodo.22168558. The keywords on the record are agent safety, reversibility, content addressing, verifiable execution, three-valued logic, signed receipts and transparency log.