Agent-Native Virtual Lab Vision

Status: Product vision and background; the Virtual Lab PRD defines implementation requirements.

Updated: 2026-07-30 Owner: RIGOR platform Scope: Product narrative, user experience, virtual/physical boundaries, and phased adoption

1. Why This Product Shape Matters

Many self-driving laboratory projects demonstrate value through physical instrument count, automation scale, and accumulated experiment data. RIGOR also needs a product shape that makes its less visible strengths—governed Agent planning, semantic devices, evidence, and safe execution—easy to understand.

The virtual lab provides that shape:

Agents rehearse experiments in an observable virtual laboratory. Researchers guide important decisions. A separate execution laboratory runs the selected plan under deterministic policy, safety, and evidence controls.

This is not a visual substitute for a physical laboratory. It separates low-cost plan discovery from formal execution. Current execution can use constrained simulated devices; physical instruments and robots can replace them incrementally behind the same semantic capability contracts.

2. One-Sentence Definition

Let the Agent rehearse the experiment before the execution lab makes it real.

A longer public description is:

Researchers watch AI agents rehearse experiments in a virtual lab, review important decisions, and send the validated workflow to an execution lab— simulated today and increasingly physical over time.

3. The Product Experience

research goal and constraints
  -> Agents plan, divide work, and test workflows in isolated rehearsal branches
  -> the researcher sees the process, risks, conflicts, and evidence gaps
  -> policy permits, asks about, or denies promotion
  -> only the selected PlanRevision enters formal execution
  -> constrained surrogate devices run it today
  -> physical instruments and robots can run it later
  -> verified execution evidence returns for the next scientific iteration

The Agents are virtual laboratory operators. The researcher remains the owner of scientific intent, material decisions, and genuinely manual work. The game-like spatial interface exists for comprehension, not entertainment: it turns abstract plans, device state, sample movement, and execution risk into a process a person can inspect.

4. Two Virtual Layers with Different Authority

The two virtual layers may reuse reviewed Device Models, but their state and evidence authority must remain separate.

4.1 Rehearsal Lab

The Rehearsal Lab helps Agents discover and repair plans. Multiple isolated branches can start from one immutable execution snapshot. A branch may:

• fork, reset, rewind, or advance simulated time;
• try plans in parallel;
• inject availability, resource, spatial, service-port, or custody failures;
• estimate duration, conflicts, failure causes, and candidate outcomes.

Rehearsal output is predicted. It cannot become scientific evidence or change the execution laboratory.

4.2 Execution Lab

The Execution Lab formally runs the selected plan. It currently uses SimFleet surrogate devices and can later bind the same semantic requirements to physical controllers.

Execution remains constrained even when the device is simulated:

• each device has one durable authoritative state;
• committed actions cannot be erased by rewinding;
• locks, idempotency, timeout, cancellation, safe-stop, and verification apply;
• Operations, artifacts, evidence, and audit cannot be overwritten by rehearsal;
• results are explicitly labeled simulated and never presented as physical measurements.

4.3 Promotion Boundary

Execution Lab state
  -> immutable ExecutionSnapshot
  -> isolated Rehearsal branches
  -> selected PlanRevision and explicit resource proposal
  -> current policy, schema, capability, and resource validation
  -> new immutable RunSpec
  -> formal Execution Lab
  -> verified simulated or physical evidence
  -> next snapshot and scientific iteration

Promotion copies a plan and explicit bindings—not device state, sample location, predicted measurements, or terminal results. Rehearsal and execution instances should come from shared reviewed Device Models rather than duplicated device definitions.

5. What the Virtual Layer Models

The first priority is operational reality, not a human-like animation of every hand movement. The model should answer:

• What semantic capabilities does each device provide?
• Is it online, available, occupied, or blocked?
• What inputs, outputs, duration, and prerequisites does an action have?
• Where are samples, carriers, containers, and materials?
• Which devices, spaces, or resources conflict?
• Which steps are automatic and which require a person?
• How does work stop, recover, or remain incomplete after failure?
• Can this laboratory execute the whole plan under current constraints?

The system should not animate an Agent pressing buttons when a real deployment would use an autosampler, fixed fixture, reviewed robot mode, or Human Task.

6. What “Trial and Error” Means

Early rehearsal is workflow-level experimentation:

• ordering steps;
• checking device and consumable availability;
• verifying safe sample transfer;
• finding resource and spatial conflicts;
• detecting missing preparation, calibration, or safety confirmation;
• testing cancellation, retry, compensation, and safe-stop paths;
• estimating duration and execution feasibility.

Unless a model has a declared applicability range and calibration evidence, rehearsal must not claim accurate prediction of unmeasured material properties, chemical outcomes, or scientific conclusions. “Operational simulation” and “experiment rehearsal” are more accurate early claims than “high-fidelity physical digital twin.”

7. Human Participation

Researchers do not operate every virtual step or approve routine actions. They:

• observe what the Agent is trying and why;
• understand device, sample, and workflow relationships;
• notice dangerous paths, omissions, and weak assumptions;
• change goals, constraints, or autonomy;
• perform required physical work;
• decide high-risk, irreversible, or materially ambiguous questions;
• Pause, Stop, or Take over when needed.

Policy remains the decision boundary:

ALLOW -> continue within declared authority
ASK   -> present the decision, evidence, alternatives, and impact
DENY  -> refuse work outside safety, equipment, or policy constraints

Visualization should make important decisions easier to understand, not add an approval button to every internal step.

8. Closing the Virtual-to-Physical Loop

The virtual experience is not an independent animation. Plans must map to real Capabilities, Operations, Procedures, and Evidence:

Goal + Constraints + Policy
  -> immutable execution snapshot
  -> Agent rehearsal and PlanRevision
  -> ALLOW | ASK | DENY
  -> current-state revalidation and immutable RunSpec
  -> deterministic RIGOR execution
  -> LabBridge semantic capability
  -> surrogate device, physical device, robot, or Human Task
  -> verification and evidence
  -> updated execution state and next rehearsal snapshot

Web 2D, Web 3D, and optional showcase clients never control hardware directly. Schema, policy, locks, idempotency, timeout, cancellation, safe-stop, verification, evidence, and audit remain deterministic service responsibilities.

9. Web 3D and Optional Unreal Clients

The primary 3D client is browser-native React Three Fiber/Three.js inside the ASCEND workspace, rendered by the researcher's GPU. Unreal Engine is optional for high-fidelity demonstrations or Pixel Streaming; it is not required for the product.

A renderer may show laboratory space, equipment, samples, Agent activity, resource conflicts, risk, waiting, and replay. It must not become:

• the source of device or workflow state;
• the only scientific record;
• a policy, lock, or approval engine;
• a path around LabBridge;
• a scientific predictor without calibration evidence.

Animation completion is never proof that an experiment succeeded.

10. Why RIGOR Is Different

Semantic communication

Virtual devices, physical devices, algorithms, and human-assisted stations use the same Capability, Operation, and Evidence language. Agents do not depend on GPIO, MQTT topics, serial protocols, robot joints, or vendor SDKs.

Human comprehension

Spatial presentation can show which Agent is doing what, where a sample is, which device is occupied, what happens next, and where risk or conflict exists. This makes an autonomous laboratory understandable beyond logs and chat.

Governed specialist collaboration

RIGOR does not treat Agent count as an advantage by itself. Scientist, Planner, and Analyst roles share goals and evidence through typed, durable boundaries; free-form Agent conversation cannot change the laboratory.

Virtual assets before hardware scale

Before every physical integration exists, RIGOR can accumulate reviewed Device Models, reusable procedures, failure scenarios, collaboration patterns, decision interfaces, execution contracts, and physical acceptance criteria. New hardware then joins an existing laboratory system instead of creating an isolated integration.

11. Relationship to the Current Architecture

This vision extends rather than replaces the Agent-first system:

• ASCEND owns goals, policy, coordination, Attention, and decisions.
• Composer constructs plans and immutable packages.
• PACE executes iterations and archives evidence.
• LabFlow schedules deterministic multi-step work.
• LabBridge owns capabilities, devices, Operations, artifacts, and fault boundaries.
• SimFleet provides reviewed surrogate behavior and failure injection.
• PRISM commits isolated analysis for the next decision.
• Spatial models own stations, service ports, routes, custody, and reviewed robot modes.

Rehearsal reuses reviewed profiles and bounded handlers in isolated instances. It does not create another user workflow, approval system, or execution state machine.

12. Phased Adoption

Phase 1: isolated rehearsal and surrogate execution

Create isolated branches from an execution snapshot, prove they cannot change surrogate state, compare them, promote only the selected PlanRevision, and run it formally with evidence labeled as simulated.

Phase 2: human-observable Web 2D

Aggregate ASCEND, LabBridge, and Rehearsal facts into one read-only scene with stable business identities, ordered replay, findings, Plan diffs, promotion, and Attention.

Phase 3: browser-native Web 3D

Render the same scene and event stream with reviewed DeviceModel assets, selection, replay, provenance, responsive layout, and automatic 2D fallback. The browser renders visuals; the server needs no GPU.

Phase 4: physical substitution and calibration

Bind a qualified plan to a small physical device chain, compare predicted, surrogate, and physical outcomes, and record applicability, uncertainty, failures, recovery, and calibration evidence.

13. A Useful Demonstration

1. A researcher submits an experimental goal.
2. Specialist roles build and rehearse a workflow.
3. The first branch finds a conflict, missing preparation step, or failed transfer.
4. The Agent revises the plan and rehearses again.
5. The researcher can understand branch differences and any required decision.
6. Only the selected plan enters the Surrogate Execution Lab.
7. Formal execution produces immutable simulated evidence.
8. The Agent proposes the next scientific iteration.

The audience should understand that the Agents advance a real experiment, the researcher can see and govern their work, and a promoted plan becomes controlled execution rather than remaining an animation.

14. Public Narrative and Claim Boundaries

Recommended summary:

RIGOR is an Agent-native virtual laboratory where Agents rehearse experiments, researchers observe and guide important decisions, and an execution laboratory runs validated plans. Constrained simulated devices provide execution today; qualified physical instruments and robots can replace them incrementally.

Avoid claims that:

• attractive 3D animation alone is a digital twin;
• an Agent or game engine can bypass device safety;
• uncalibrated simulation equals a physical measurement;
• Human Tasks, safety checks, or missing evidence can be hidden for a smoother demonstration;
• every internal stage requires manual approval;
• Agent count is itself a technical moat;
• predicted rehearsal output can become simulated or physical evidence;
• Rehearsal and Surrogate Execution share one rewindable state.

RIGOR's value is a reusable planning, execution, and evidence layer that can connect more physical equipment over time without changing its scientific and safety boundaries.