Architecture

Separate orchestration from the evidence.

Agents and assistants can interpret intent and coordinate work. Lab Loop owns the loop: the hypotheses, controls, evidence and conclusions that define the research.

Lab Loop reference architecture

Division of responsibility.

Different layers do different jobs.

Orchestration & agent layer

User interaction, intent interpretation, tool selection and process coordination.

Lab Loop

Programmes, hypotheses, designed experiments, controlled bundles, evidence, belief state and immutable conclusions.

Compute & systems of record

Training compute, remote clusters, trackers, data stores, model registries and lab robotics remain authoritative for execution and operational state.

Controlled loop operations.

Expose narrow, typed operations rather than unrestricted access to internal state.

Typical operations

open_programme
state_hypothesis
design_trial
capture_bundle
record_evidence
update_belief
close_conclusion

Control boundary

The caller requests a step of the loop. Lab Loop executes it against explicit governance and returns a reproducible, traced result.

Governance.

A distinct loop layer creates a clear place to apply controls, reproducibility and evidence requirements.

Explicit hypotheses

Claims and their refutation criteria can be inspected and versioned.

Bounded access

Agents operate through defined loop operations rather than unconstrained mutation of research state.

Traceability

Hypotheses, bundles, evidence and conclusions are retained together, so any claim can be traced to its experiments.