Control Room and intervention

Research Control Room is the operating and audit center for one run. SSE carries invalidation hints, after which the UI refetches authoritative Runtime views. Reconnection is automatic, and a manual refresh remains available.

Lifecycle and header actions

Buttons are enabled only when Runtime permits them. Cancel terminates the active executor process group, preserves committed events and artifacts, and ends as CANCELLED, not FAILED. Recover replays a deterministic committed checkpoint before continuing.

Research Timeline

The event-backed timeline offers four filters:

• All — every readable projected event;
• Evidence — datasets, models, evaluation, metrics, experiments, analysis, knowledge objects, verification, and reports;
• Decisions — scientific decisions, user interventions, and verification requests;
• Failures — warning/error events.

Expand a card for a concise reasoning summary, duration, metrics, action/job/artifact IDs, and sanitized details. Hidden chain-of-thought is not displayed, and animation is never a substitute for Runtime state.

The nine inspector tabs

State

Current step, failure/recovery state, lifecycle reason, event and scientific-object counts, Runtime contract versions, current action, criteria progress, trusted metrics, active models, executor, authorized capabilities, and intervention count.

When the run is not RUNNING, Edit acceptance criteria creates an auditable revision and reassesses existing metrics without deleting prior criteria history.

Knowledge

Workflow switches, artifact-completion state, Evidence, Claims, Active/Tested Hypotheses, and Verification plans/results. Items preserve confidence, limitations, versions, and source actions.

Verification

Hypothesis, objective, experiment type, parameters, success conditions, result, condition outcomes, and version chain. A PLANNED item can be Approved; an APPROVED item can be Requested. Commands still pass contract, policy, budget, and lineage validation.

Reports

Report artifacts, size, and Artifact ID. Desktop can open the containing directory. Use Lineage to read immutable Markdown and inspect figures/manifests. The complete structured report read model is also available through Control API.

Model calls

Role, provider/model, route/fallback, latency, token usage, retries, schema/code repairs, failure codes, parent action, and produced artifacts. Raw prompts and hidden reasoning are not returned.

Metrics

Current values and time series, with reported/trusted classification. A single point is shown as the first observation; multiple points render as a chart.

Resources

Wall time, CPU/GPU hours, memory, training samples, model/Reasoner/code calls and tokens, repair attempts, jobs, and active jobs. Unavailable observations appear as —.

Budget

Used/limit values, percentage, and whether each resource is enforced. Input, output, and total tokens are also grouped by provider/model. Runtime does not infer monetary prices.

Jobs

Executor, command summary, state, duration, exit code, failure, source/output artifacts, and stdout/stderr links. Diagnostic streams are not a scientific metric transport.

Ask: read-only questions

Useful examples:

Which evidence supports the current decision?
Why is the run blocked?
Which metric satisfied the criterion?

Answers cite timeline, artifact, metric, job, or model-call records. Exchanges persist in a separate interactions.jsonl; they survive refresh/restart but do not consume action budget, invoke a capability, change Criteria, or mutate ResearchState.

Intervene: finite writes

Intervene accepts exactly four structured directives:

• Adjust budget — change one limit; Experiments may be unlimited. For BUDGET_EXHAUSTED, the UI suggests the exhausted field and performs increase-and-resume.
• Add research note — add a visible note of up to 4,000 characters.
• Pause run — available only for RUNNING.
• Resume run — available for PAUSED, BLOCKED, or CAPABILITY_MISSING.

Runtime validates each operation and records it permanently as a local_user intervention. Free text is never silently interpreted as a command.

Budget-change rules

• A new value must differ from the current value. Recovery from exhaustion requires raising or removing that limit.
• A non-exhausted run may be tightened; the new limit takes effect at the next budget check.
• Unlimited max_experiments is JSON null, not an arbitrarily large integer.
• Budget changes cannot authorize capabilities, grant trusted-evaluator authority, or rewrite old metrics.

Scientific package export

After SUCCESS, select Export Scientific Package. It creates a human-facing directory such as:

<name>_export/
├── README.md
├── report.md
├── model/          # model, inference.py, metadata when present
├── examples/       # sample input and invocation
├── figures/
└── docs/           # limitations and supporting notes

A scientific package is for use and sharing. It may omit the complete event stream, internal checkpoints, and full provenance. For disaster recovery use Settings → Storage → Export Backup.