A contour scan of a brain behind the Axon statement. Decorative.
Axon sits between an agent’s proposal and its execution. It checks the proposal against the systems that own the facts, and only an action whose evidence is signed, complete and current receives authority to run.
Satisfied issues a receipt to the executor. Violated, Unresolved and Refused stop at the interlock.
Axon is the authorization layer for AI agents that move money, change code and control machines. An agent does not receive general permission. It receives authority for one proposal, under one policy generation, against one set of evidence. If the bank record changes, the order census is incomplete or the approval has lapsed, the same proposal is not authorized, and the receipt says which fact failed.
Verify the owner, signature, freshness and completeness before evaluating policy.
Models such as TypeSafe’s Jev now return typed, calibrated decisions at machine speed. That is a real shift, and Axon is built to work with it. A calibrated judgement is still a judgement. Before an agent moves 24,800 EUR, deploys to production or actuates hardware, someone has to check the authoritative world and be able to prove they did.
Why we’re rethinking pull requests ↗An agent proposes. Axon collects the facts from the systems that own them, verifies the evidence, evaluates the policy and, only on Satisfied, issues an action receipt the executor can honour.
Read: The Age of Software GamblingAdapters translate one authoritative system each into typed observations that carry their source identity, revision, capture time, completeness and content digest. Axon never treats a similar name as the same identity, and never treats a partial page as the whole.
The kernel verifies tenant, owner, digest, signature, completeness and freshness first, then evaluates the governing rules within a fixed work bound. Generic operators, no domain vocabulary. The same kernel has run treasury payments, facility safety and robot stopping distances.
A Satisfied verdict produces an action receipt bound to the proposal, the policy roots, the evidence roots and the witness rows. An executor accepts the receipt for that exact action and rejects it for any other.
A changed record invalidates only the receipts that depend on it. Any receipt can be replayed against its retained evidence to return the exact rows, digests and policy generations behind the decision, for an auditor, an incident review or a regulator.
The standing order changes. Axon checks each proposed action before it reaches the game.
Firing permitted only when the other checks also pass.
0fire actions executed
under HOLD YOUR FIRE
Recorded local prototype. 840 governed frames in ViZDoom with Freedoom assets and a scripted controller. FIRE AT WILL permits a proposal; it does not bypass the remaining checks.
Same inputs. Same verdict.
Concept illustration / receiptThe receipt binds the proposal, policy roots, evidence roots and witness rows. Changed facts invalidate dependent receipts; retained evidence enables replay.
Models propose. Axon authorizes.
Models propose. Axon authorizes.
Yes. Design partners use Axon to authorize real agent actions against their own systems of record, with a receipt for every decision. We are taking on a few more design partners: fixed scope, a few weeks, a receipt at the end. Get in touch with the action your agent has to take and the systems that own the facts.
An action receipt names the action, the policy roots, the evidence roots, their owners and capture times, the witness rows and the decision digest. It authorizes one proposal, not a class of them. A changed fact revokes the receipts that depend on it and no others. A receipt can be replayed later against the same bytes and re-verified.
The kernel is deterministic. The same proposal, policy generation and evidence bytes produce the same verdict and the same decision digest, so a receipt can be replayed and re-verified later by anyone who holds the retained evidence.
Satisfied means every governing condition held against complete, current evidence. Unresolved means the evidence was partial, stale or the bounded evaluation could not finish; it is never rounded up to a pass. Refused means a record no longer matches its digest. People still decide whether the policy describes the right rule.
Axon grew out of our work on specification-checked development for coding agents, previously introduced as Synapse. Code changes are now one action type among the ones Axon authorizes.
Payments, deployments, fleet commands. Design partners already run Axon in front of real action paths on their own evidence sources, and we are taking on a few more: fixed scope, a few weeks, a receipt at the end.
belic_bojan@sigmadelta.be