---
title: "Authority for Autonomous Science — the authority, evidence and finality layer beside the scientific decision chain · KYE Protocol™"
description: "An autonomous science agent can design and run an experiment; only your institution can authorise it. KYE Protocol™ sits beside the scientific decision chain and proves who legitimately empowered an agent to cause a specific real-world effect — from a wet-lab run to a nucleic-acid synthesis order — with purpose-bound authority, action admissibility, suspendable Authority Finality™, and evidence verifiable from public keys alone. Bring your models and lab automation; KYE™ makes every consequential action regulator-provable."
url: https://kyeprotocol.com/authority-autonomous-science/
lang: en
source: "KYE Protocol"
---

> An autonomous science agent can design and run an experiment; only your institution can authorise it. KYE Protocol™ sits beside the scientific decision chain and proves who legitimately empowered an agent to cause a specific real-world effect — from a wet-lab run to a nucleic-acid synthesis order — with purpose-bound authority, action admissibility, suspendable Authority Finality™, and evidence verifiable from public keys alone. Bring your models and lab automation; KYE™ makes every consequential action regulator-provable.

# Autonomous science acts on the world. Authority decides who was allowed to.

When a research agent orders a wet-lab run, triggers DNA synthesis, allocates reactor time, or files a result into a regulatory dossier, it takes a **consequential action in the physical world**. [KYE Protocol™](https://kyeprotocol.com/) sits _beside_ the scientific decision chain and proves your institution legitimately authorised that action — with [Authority Finality™](https://kyeprotocol.com/authority-finality/) that can be suspended mid-experiment, and an [Evidence Pack™](https://kyeprotocol.com/evidence-pack/) a regulator can verify. If you run a lab, this cuts incident reconstruction from days to minutes.

[Start a governed pilot](https://kyeprotocol.com/poc/) [See Authority Finality™](https://kyeprotocol.com/authority-finality/)

## Four states of an autonomous experiment

If you run a lab or answer for its biosafety, an autonomous run can succeed on one axis and fail on another. The EU AI Act binds logging and human-oversight duties per action, not once a year — so KYE™ separates these 4 states, and a green pipeline never hides an illegitimate act.

- **Technically successful** — the agent completed the protocol and the instrument returned a result. _Did it run?_
- **Policy compliant** — the configured rules and guardrails — the kind the EU AI Act Article 12 logging duty assumes — were satisfied at runtime. _Did it follow the rules it was given?_
- **Authority valid** — your institution actually empowered _this_ agent for _this_ effect, via [Purpose Permission™](https://kyeprotocol.com/open-access-authority/) and **Action Admissibility™**. _Was it allowed — by whom, for what?_
- **Effect final** — the action is sealed, revocable, and provable. **Authority Finality™** means a control that fails can suspend the mandate mid-flight. _Can it be undone, and can you prove what happened?_

A configured rule check answers "did it follow the rules?". It never answers "who was allowed to set those rules, and did they have the mandate?" — that is a source-authority question, and it is the one an inquiry asks first.

## The dual-use wedge: capability is not authority

If you lead research security, your sharpest case is dual-use. An agent that can design a toxin, propose a gain-of-function edit, or place a nucleic-acid synthesis order has **capability** — but capability is not a mandate, and your institution is the one an inquiry will ask to account for it.

Standards already draw the line at the action, not the model: the U.S. OSTP Framework for Nucleic Acid Synthesis Screening and the IGSC Harmonized Screening Protocol both gate the _order_, and U.S. Executive Order 14110 binds duties to the deployed system. KYE™ operationalises that line: an agent may be discovered and registered, yet remain in a **registered-but-unauthorised** state until a named principal grants scoped authority. **Visibility never silently creates permission.** A flagged agent cannot escalate itself from "seen" to "acted".

## Beside the safety layer — not on top of it

If you already run evaluation, red-teaming, or behavioural observability over your science models, keep them. Those platforms measure how a model _behaves_; KYE Protocol™ proves whether an agent was _authorised to act_. The two compose (§0.25 integrate-not-compete) — KYE™ consumes their signals as inputs to an authority decision.

| An AI-for-science safety / eval layer gives you | KYE™ adds beside it |
| --- | --- |
| A benchmark or eval score for a model's capability | Purpose Permission™ — was this agent granted authority for _this_ experiment, by whom? |
| Behavioural monitoring and drift alerts | Action Admissibility™ — admit, block, escalate or quarantine the consequential action at the moment it happens |
| A trace of what the model did | Replay-Proof™ — sealed evidence any regulator verifies from public keys alone, without trusting KYE™ |
| A red-team finding that a capability exists | Authority Finality™ — suspend the mandate mid-flight when a control fails, and evidence the revocation |
| Mappings to a handful of frameworks | Deep mapping to the EU AI Act, NIST AI RMF, ISO/IEC 42001 and 249 frameworks, each tied to the artefact that enforces it |

Boundary: KYE™ governs whether an agent may act and proves what it did; it does not run your science, your models, or your lab automation. It is the authority + evidence layer that turns their outputs into a decision your institution can defend.

## Why above any lab runtime

If your team is building a self-driving lab, the automation that executes your protocol — a cloud lab, a robotic workcell, an orchestration engine — is increasingly interchangeable to you. The evidence that satisfies your institutional review board, your funder, or your regulator is not.

- Authority is decided once and proven forever: the same sealed [Evidence Pack™](https://kyeprotocol.com/evidence-pack/) serves the incident responder today and the auditor next quarter.
- A lab runtime can change; the authority record and its public-key proof do not depend on which system produced them.
- A signed **Authority Risk Signal™** lets a research institution's leadership and its insurers price exposure from authority failures, not just model errors.

[Start a governed pilot](https://kyeprotocol.com/poc/) [See the Agent Authority Stack™](https://kyeprotocol.com/agent-authority-stack/)

Canonical KYE™ surfaces referenced on this page: [Purpose Permission](https://kyeprotocol.com/glossary/) · [Systemic Authority Risk™](https://kyeprotocol.com/governance-for-the-ai-web/).