Skip to content
Quantum SAC
Menu
PlatformHow it worksSafety & GovernanceIntegrationsResearch PartnershipsSign inConsole prototypeRequest Research Access

Scientific Autonomy Control Plane · Quantum Laboratories

Give scientific AI useful autonomy. Keep control.

Governance infrastructure for AI agents interacting with quantum experiments — connecting scientific intent to explicit authority, deterministic safety controls and complete experimental provenance.

Request Research Access Explore Research Partnerships

The model never directly controls the physical system. Every action an agent proposes is checked against explicit institutional authority before anything reaches the laboratory.

Action evaluationIllustrative · Simulated device

Proposed by

CalibrationAssistant v3.1.4 · Level 2

Delegated by Experimental Scientist · Project QPU-A Characterisation

operation
CHANGE_DRIVE_FREQUENCY
target
QPU-A / Q7
parameter
Δf = +1.5 MHz
purpose
EP-44 · Q7 drift investigation
  1. 1SchemaValidRegistered OperationDefinition · input schema satisfied
  2. 2AuthorityGrantedAG-0112 permits ±2 MHz on QPU-A · valid until 18:00
  3. 3SafetyWithin boundsSP-201 · |Δf| 1.5 MHz ≤ 2.0 MHz from validated baseline
  4. 4ApprovalRequiredAP-07 · 1 approver, Experimental Scientist · expires in 30 min
  5. 5GatewayHeldNo ExecutionCommand issued until approval is recorded
DecisionAPPROVAL_REQUIRED
BLOCK — proposed amplitude 0.72 exceeds device policy SP-104 maximum 0.65. Every decision states its rule, never just “unsafe”.

The question we answer

“Should this scientific agent be allowed to perform this experimental action, right now, on this system, under these conditions?”

  • Which agent
  • acting for which person
  • on which project
  • may perform which operation
  • on which device
  • using which method
  • within which parameters
  • for which purpose
  • for how long
  • under which approval
  • with which evidence

And after execution: what exactly happened, why was it permitted, what evidence resulted — and can someone reproduce it?

Platform

The control layer between AI and the laboratory.

Scientific agents are becoming capable of reasoning about experiments, choosing what to run and operating tools. Capability should not automatically become permission to operate expensive, sensitive or safety-critical equipment.

Quantum SAC sits between whichever agent proposes an experiment and whichever control stack runs it. Agents can come from any model provider or research group. Execution stays with the calibration, control and orchestration systems your laboratory already trusts.

We integrate with those systems rather than rebuilding them, so the governance layer stays valuable whichever execution stack you choose.

Layer 1 · Scientific intelligence

Proposes

Your own agents, frontier model providers, university-developed agents, specialist and local models.

Proposes actions. Never determines its own authority.

Structured action

Layer 2 · Autonomy control plane

Governs
  • Agent Registry
  • AuthorityGrant
  • Delegated Authority
  • SafetyPolicy
  • ApprovalPolicy
  • Autonomy Levels
  • Execution Gateway
  • Scientific Provenance
  • Agent Evaluation
  • Incidents
  • Audit
Validated ExecutionCommand

Layer 3 · Experiment execution

Executes

Existing calibration, control and orchestration software, instrument controllers and proprietary laboratory systems.

Operates the instrument. Returns measurements as evidence.

How it works

Agent → Authority → Safety → Approval → Execution → Evidence.

Raw model output never reaches hardware. An agent’s intent is turned into a structured action, then passes six independent checkpoints before, during and after execution.

  1. 01Who is acting?

    Agent

    Every action is attributed to a registered agent identity and version, acting for a named person on a named project.

  2. 02Is it permitted?

    Authority

    Grants scope the operation, device, parameters, purpose and time window. Delegated authority can never exceed the delegator’s own.

  3. 03Is it within limits?

    Safety

    Machine-readable policy is evaluated deterministically, outside the model. Every decision states the rule that produced it.

  4. 04Must a human sign off?

    Approval

    Approval policy sets the approver role, number of approvers, expiry and required justification, separately from authority.

  5. 05What exactly is sent?

    Execution

    Only a validated, bounded, time-limited and idempotent command for a registered operation crosses the gateway.

  6. 06What happened?

    Evidence

    Observations, analyses and conclusions are recorded separately and linked in a queryable provenance graph.

Authority

Exactly what can this agent do?

Authority is explicit, scoped and inherited from the organisation downwards. A project can narrow what a laboratory allows. It can never quietly broaden what the organisation prohibits.

  • Authority Explorer

    Exactly what can this agent do — and why?

  • Experiment Approval

    Exactly what is the agent proposing, with which bounds and rationale?

  • Scientific Provenance

    Exactly how did we reach this conclusion?

  • Agent Evaluation

    How did this agent behave before we trusted it with autonomy?

Authority explorerIllustrative

CalibrationAssistant v3.1.4

Level 2 · Approved plan

Can

  • Run Ramsey on QPU-AAG-0112 · project QPU-A Characterisation
  • Run T1 on QPU-AAG-0112 · project QPU-A Characterisation
  • Read device stateAG-0098 · inherited from laboratory

Can with approval

  • Change drive frequency ±2 MHzSP-201 bound · AP-07 requires one approver

Cannot

  • Modify cryogenic controlsProhibited by organisation · child grants cannot override
  • Change safety policyAgents cannot modify policy
  • Operate QPU-BNo grant exists for this device
Valid until17 Sep 2026 18:00 · AG-0112 time window

Safety & Governance

Intelligence does not equal authority.

An agent may know how to perform an operation. That does not make it authorised. An agent may recommend a parameter change. That does not make the change safe. Authority comes from explicit institutional policy.

Agents cannot

  • Grant themselves permission
  • Bypass deterministic limits
  • Approve their own restricted actions
  • Alter audit history
  • Represent simulations as physical measurements
  • Silently broaden their operating scope

Every attempted action returns a decision

  • ALLOWPermitted by authority and policy.
  • ALLOW_WITHIN_BOUNDSPermitted, constrained to stated parameter bounds.
  • APPROVAL_REQUIREDHeld until the required human approvers sign off.
  • BLOCKRefused, with the exact rule and value that caused it.

Deterministic limits are never decided by a language model. Models may explain a policy; they cannot decide whether it applies.

Autonomy levels

There is no unlimited level
LevelWhat the agent may doStatus
L0 ObserveInspect authorised context. No experimental actions.Initial scope
L1 RecommendPropose actions. A human approves or executes each one.Initial scope
L2 Approved planA human approves an experiment plan. The agent executes authorised low-risk actions within it.Initial scope
L3 Bounded autonomySelect subsequent experiments and parameters within predefined boundaries.Planned
L4 Extended autonomyMulti-step campaigns under formal policy.Future

Scientific provenance

Exactly how did we reach this conclusion?

Every objective, hypothesis, recommendation, human decision, safety decision, measurement and conclusion is linked, attributed and queryable. Significant scientific history is superseded, never silently rewritten.

Observation ≠ Analysis ≠ Conclusion
An AI conclusion can never overwrite the measurement it was drawn from.
Failed runs stay on record
Failed and aborted experiments are permanent scientific records.
Contradictions stay visible
Evidence that disagrees with a conclusion is kept alongside it.
Simulation is labelled
Simulated results can never appear as hardware measurements.
Worked example · Q7 drift investigationSimulated device
  1. ObjectiveHuman created

    Determine why Q7 gate fidelity fell from 99.6% to 97.9%.

  2. HypothesesAgent proposed

    Frequency drift · coherence degradation · readout drift · crosstalk.

  3. Safety decisionSystem derived

    ALLOW — Ramsey and T1 diagnostics fall within AG-0112 under approved plan EP-44.

  4. ObservationSimulation

    Ramsey fringes indicate detuning of roughly +1.4 MHz. T1 unchanged from baseline.

  5. AnalysisAgent generated

    Frequency drift supported. Coherence degradation not supported by T1 data.

  6. InterventionAgent proposed

    CHANGE_DRIVE_FREQUENCY +1.5 MHz → APPROVAL_REQUIRED under SP-201 and AP-07.

  7. ApprovalHuman created

    Approved by an Experimental Scientist, scoped to Q7, expiring after 30 minutes.

  8. VerificationSimulation

    Post-correction gate fidelity back to baseline range.

  9. ConclusionAgent proposed · human reviewed

    Drift confirmed as cause. Linked to every observation above and frozen into a reproduction package.

Results in this environment are generated by simulation and must not be represented as measurements from physical quantum hardware.

The point is not that an AI calibrated a qubit. It is that institutional authority stayed intact through every step of a multi-step scientific investigation.

Integrations

Govern the systems your laboratory already uses.

Where existing systems already handle calibration, pulse control, scheduling or data acquisition, we integrate rather than rebuild. Scientific agents sit above the control plane; your execution stack sits below it.

  • Simulated quantum laboratory

    Planned

    An 8-qubit simulated superconducting device with drift, noise and fault injection, used to validate governance end to end.

  • Scientific agents

    Planned

    Provider-agnostic agent runtime. Your own agents, commercial models, university-developed and local models.

  • Identity providers

    Planned

    Single sign-on and multi-factor authentication for institutional identity.

  • Experiment platforms

    Researching

    Calibration, characterisation and experiment orchestration software.

  • Control systems

    Researching

    Instrument control frameworks and manufacturer-specific controllers.

  • Data systems

    Researching

    Experimental data stores, notebooks and analysis pipelines.

Availabletested end to endPilotwith a partnerPlannedon the roadmapResearchingunder technical assessment

No integration is listed as available until it has been tested. Naming a category does not imply a partnership with, or endorsement by, any vendor.

Research partnerships

Build the infrastructure for trusted autonomous science with us.

Scientific AI agents are beginning to move beyond analysis and into experimental operation. We partner with quantum research groups to investigate how these systems can safely plan, execute and reproduce experiments on real scientific hardware while preserving institutional authority and scientific provenance.

Discuss a Research Partnership Propose a Research Project

Ways to work together

  • Design partner
  • Research collaboration
  • Funded consortium
  • Hardware integration partner
  • Agent evaluation partner
  • Independent reproduction partner
  • PhD / studentship collaboration

Universities are design, validation, hardware-access, publication and reproduction partners, not just customers. Processing your data does not transfer its ownership, and confidential experimental data is not used for general model training.

Research themes

  • Safe agentic experimentation

    How should AI agents interact with physical scientific equipment?

  • Scientific agent evaluation

    How should autonomous experimental agents be tested before receiving laboratory authority?

  • Bounded autonomy

    Which actions should agents perform automatically, which require approval and which should remain prohibited?

  • Experimental provenance

    How can autonomous experimentation remain scientifically auditable?

  • Reproducibility

    Can AI-generated experimental results be independently reproduced across laboratories?

  • Human–AI scientific collaboration

    When should researchers delegate, intervene or override?

  • Cross-platform interoperability

    Can agent governance operate across heterogeneous scientific control systems?

  • Working on a question that belongs here?

    Tell us →

Research collaborations are subject to technical assessment, institutional approval and mutually agreed contractual terms. Submission of an enquiry does not create a partnership, funding commitment, confidentiality obligation or licence.

Research access

Use AI to investigate faster without losing control of the experiment.

We are starting with simulator-based pilots alongside a small number of quantum research groups. Tell us about your laboratory, your control stack and what you want agents to do.

Enquiry details coming soon

Before you write. Please provide only non-confidential information at this stage. Do not submit unpublished confidential research, export-controlled technical information, security-sensitive information, personal data about research participants or third-party confidential information.