TRACELET
Check before you act. Every step reviewed, every session on the record.
TRACELET sits between a proposed action and the systems it would touch. It reviews the action, and it cannot change anything on its own. Those limits are built into how it works, not written into a policy. It does not think up answers and it is not an autonomous AI. You pay per action reviewed, not per word processed, it runs on modest hardware, and every session leaves a record of what was checked.
Three editions
The TRACELET suite is specified in three editions. They share one governance core, the turn boundary and its sealed record, and differ in how they are deployed and who operates them.
A turn-metered, substrate-agnostic middleware that sits between an orchestrating authority and the systems it runs on. It meters by turn, not by token: a turn begins with an authorized request and ends with a sealed outcome, and that boundary is the unit of governance, the unit of evidence, and the unit of metering at once. This is the runtime the live trace below demonstrates.
An enterprise-grade sovereign agent shell, deployed as documentation-safe quality assurance infrastructure. It is a trace-locked, non-LLM, non-networked, non-persistent control plane for validating work in regulated environments, and it never mutates upstream systems.
The hardened, fully offline edition: a fixed fifteen-module stack invoked in a boot order behind a sovereignty and safety barrier. It runs with no internet, no model inference, and no side effects. The trace guard loads the agent profile, confirms internet access is disabled and the guard is enabled, and halts execution if any check fails.
What it does
- Checking AI work before it goes out, and keeping the record an auditor asks for.
- Regulated-environment validation across finance, health, defense, and research.
- Checking your own internal tools: is the policy followed, where did this data come from, did the workflow do what it says.
- Human-in-the-loop enforcement nodes, with a human operator required for every invocation.
How it meters
You pay per action reviewed, not per word processed. Each review runs once, start to finish, and stops hard if a rule is not met before anything executes. It follows a fixed set of steps every time, so the same action gets the same answer, and it cannot run away with itself.
Nothing is kept between sessions. The session clears when you close it, and the record of what was checked stays with you. TRACELET does not trust what it is given, does not run what it produces, checks every step, and never changes your systems. It is the same review, done the same way, every time.
TRACELET works the same way everywhere else in the estate, keeps a record of what it did, and can run fully disconnected from the internet if your deployment needs that.
Live trace
Type an action you are about to take and TRACELET reviews it once: it reads what you are asking for, checks every step against the rules, refuses to touch anything upstream, and shows you the record with a receipt you can keep. One action per run. Nothing is remembered after you leave. It all runs in your browser, on modest hardware, exactly as it would in production.
Verification and replay
Every turn emits an Ethical Decision Graph: an evidence-and-decision record of the constraint evaluated, the evidence consulted, and the disposition reached at each step. Because the graph is emitted as a first-class, content-addressed artifact, any party holding its reference can verify the decision path independently of the runtime that produced it. The receipt from the live trace above is such a reference. Keep it, and the same review can be confirmed later without trusting this page.
The verification surface that exposes these references for independent confirmation is documented in the TRACELET 1.1+EDG paper, in its binding layer. That is the same mechanism a caller uses to replay and check a sealed outcome.
How to integrate
TRACELET is specified to deploy as a thin middleware layer that an orchestrator calls once per turn, placed inline on the path between the authority and the systems it governs so that no turn reaches an engine except through admission.
The integration surface is deliberately narrow. An orchestrator submits a turn request carrying the intended action and the evidence references that authorize it, and the kernel returns either a sealed outcome with its EDG reference or a sealed halt with the failed constraint. The two structural integration points are the constraint source, the rule set the kernel evaluates against, and the evidence store, the append-only record it reads prior decisions from and writes new seals to. An ill-formed request, a missing evidence reference, or an unreachable constraint source all resolve to a halt rather than to a permissive default.
Specification: TRACELET 1.1 + EDG SaaS specification.