Repository navigation
CML: Machine-Verifiable Continuity and State Transitions for Stateful Agents #1847
cstolting-collab
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
CML: Machine-Verifiable Continuity and State Transitions for Stateful Agents
I’ve been independently developing CML, a continuity/state validation architecture for systems that must preserve accepted state across multi-step execution.
The core model separates:
canonical accepted state → read-only projection → candidate transition → invariant/authorization validation → explicit commit
and separately:
external observation → noncanonical evidence → PASS / FAIL / UNEVALUABLE
Before October 6, 2026, the implementation already included declarative invariants, state locks, authorized unlocks, stale-projection rejection, fail-closed transition handling, provenance, checkpoint recovery, and noncanonical observation validation.
I’ve published a small reproducible evidence bundle rather than asking anyone to accept the claim on trust:
CML evidence:
https://github.com/cstolting-collab/CML-Trust/tree/main/evidence/openai-2026-10-06
The minimal demonstration shows the same candidate transition being rejected while canonical state is locked, then accepted after an authorized unlock, while evaluation itself never mutates canonical state.
I’m not claiming CML is a theorem prover or that it invented formal verification. I’m interested in whether this state-transition verification layer is useful to current work on agent evaluations, oversight, and long-horizon stateful systems.
Technical review is welcome. My GitHub profile is linked through the repository.
All reactions