# How Tetheract works

Tetheract separates three roles that are often blurred together: the agent that plans, the checks that constrain, and the person who decides. Everything below describes the current simulation-only software.

> Status: simulation only. Tetheract does not control real hardware today and makes no safety, certification or compliance claims. A supervised test on a real motorized linear stage is planned only after a qualification checklist and an independent review, and no independent reviewer has been named yet.

## The layers

1. **AI agent (untrusted proposer).** Proposes intents with a rationale through a Python SDK that can only propose. The description and rationale are shown to the operator but are never used as enforcement input; enforcement reads typed fields only (actuator, target, unit, frame, velocity).
2. **Authority and constraint checks.** Run independently of the agent and never consult it.
3. **Operator console.** Shows the proposal, check results, live state and predicted outcome, and records the operator's decision.
4. **Machine adapter.** A typed contract between Tetheract and a machine. Today every adapter targets a simulator.
5. **Simulated machine.** The motorized linear stage simulator is the first lab-automation target.

## The independent checks

**Authority.** A proposal must be backed by a grant that is correctly signed, in scope, unexpired, unrevoked and not replayed. In this release the authority service is a development stub for testing the validation logic, not a production authority service.

**Constraints.** The proposed command is tested against position, velocity, force and temperature limits, required sensors, sensor freshness, expected units and coordinate frames, and the state of the interlock. Limits are frozen against changes from the agent.

**Enforcement process.** A separate local enforcement process applies the checks with a watchdog and deadlines, does not import the planner, and fails closed when it is not sure.

## Operator decisions

- **Approve and execute** the proposal as checked.
- **Edit** the command, which sends it back through the checks.
- **Reject** it with a reason that is recorded.
- **Take manual control**, which pauses agent proposals until the operator hands control back.

Each consequential action, such as approving a move, taking manual control or changing a policy or key, requires a fresh passkey check tied to that specific action. The check is copied onto the evidence receipt. Recovery uses a second registered security key, never email or SMS. Policy changes and key rotation can require approval from a second admin once one exists.

## Evidence

Every request, check, decision, execution and sensor reading produces a receipt. Receipts are signed and hash-chained, so an exported bundle can be verified for tampering. Receipts record the actor (agent, operator, system, constraint enforcer or authority validator), what happened, when, and the reason for any denial. Exports are labeled simulation only and are not authority to act.

## The lab protocol demo

In the lab demo, an AI planner proposes a multi-step scan on the simulated linear stage. The stage must home before it moves. For each step the operator sees a card with the predicted result next to the measured one and approves the motion or confirms an observation. The planner never approves. Results export as an experiment-result record.

## Before any real hardware

Simulation of a machine class does not qualify that class. The planned progression is:

1. Simulation (current stage).
2. An independent reviewer is named (open; not yet started).
3. Hardware-in-the-loop testing, with actuation still disabled.
4. Supervised, low-energy bench validation of the motorized linear stage.
5. Integration with a real authority service.
6. Pilot deployment.
7. Production qualification.

## Learn more

- [Overview](https://tetheract.com/): What Tetheract is, how the agent-human loop works, who it is for, what exists today and what it is not.
- [FAQ](https://tetheract.com/faq): Direct answers to common questions about status, hardware, approvals, sign-in, certification, pricing and source code.

Canonical page: https://tetheract.com/how-it-works