AI proposes. Independent checks constrain. A human decides.

How agents and humans work together

1. The agent proposes. An AI agent or planner suggests an action, such as a stage move, along with its rationale. Tetheract treats the agent as an untrusted proposer.
2. Tetheract checks, independently of the agent. Authority checks confirm the proposal is backed by a valid, unexpired, unrevoked and non-replayed grant. Constraint checks test it against position, velocity, force and temperature limits, sensor freshness, units, coordinate frames and interlocks.
3. The operator sees everything in one place. The proposal, the check results, the current machine state and the predicted outcome appear together, with a plain-language explanation of anything that was denied.
4. The operator decides. Approve and execute, edit and re-check, reject with a reason, or take manual control, which pauses agent proposals.
5. Only approved commands run. Execution goes through a typed machine adapter. The agent cannot approve its own proposals.
6. Every handoff is recorded. Each step lands on a shared evidence timeline as a signed, hash-chained receipt showing who acted, when and why.

The result is a working arrangement where the agent does the planning and the human keeps the authority. An attention queue tells the operator what needs a decision, and incident review is built from the same evidence receipts.

Who it is for

• Lab-automation teams exploring AI-assisted operation of equipment such as motorized stages, who want a person to stay in charge of every physical action.
• Lab-automation integrators who want to evaluate a typed adapter contract and an operator approval flow in simulation.
• Qualified independent reviewers. An independent review is the open requirement before the first supervised real-hardware test.

Tetheract is early-stage software for evaluation in simulation. It is not ready for production use or for unsupervised operation of any machine.

What exists today, in simulation

• A detailed simulator of a motorized linear stage, including homing, limit switches, encoder tracking against commanded position, stalls and missed steps.
• A lab protocol demo in which an AI planner proposes a multi-step stage scan and the operator approves each motion, with predicted and measured results shown side by side.
• An operator console with an attention queue, denial explanations, manual takeover and incident review.
• Passkey-only sign-in with viewer, operator and admin roles, and a fresh passkey check for each consequential action.
• Signed, hash-chained evidence receipts that can be exported and verified for tampering.
• Fault injection (stale, spoofed or missing sensors, unit and frame mismatches, partial actuation, timeouts, network partitions) and adversarial tests (forged, expired and replayed grants, prompt-injected proposals, attempts to rewrite limits, tampered updates).
• Additional simulated fixtures on the same adapter contract: a two-axis gantry and in-process Modbus TCP and OPC UA simulators.

What Tetheract is not

• It does not control real hardware. A real-hardware adapter exists only as an interface with actuation disabled.
• It has no safety rating or certification, it has not had an independent safety review, and it makes no regulatory compliance claims.
• It is not a hosted service. It runs on your own computer and listens only on the loopback address by default.
• It is not suitable for autonomous operation without human supervision.