Core Workflow

1. Create one project per study or research boundary.
2. Add source data in Database.
3. Inspect schema, metadata, warnings, units, and identifiers.
4. Select exact files, artifact versions, or literature.
5. Ask in Chat; ask for a plan first for consequential methods.
6. Follow durable response or background-job status.
7. Validate files, figures, code, citations, and provenance in Workspace.

Chat Modes

/standard · /profile · /feedback

Prompt Template

• Objective: decision or estimand
• Evidence: @files and exact versions
• Population: subset and exclusions
• Method: model and assumption checks
• Outputs: files, figures, notebook, report
• Guardrail: when CATO must stop and ask

Keyboard

Workspace

• Files: preview, edit, versions, provenance
• Figures: viewers, SVG editing, visual QA, export
• Notebooks: Python/R code, captured output, downloads
• Manuscripts: Markdown/LaTeX, preview, citations
• Readers: PDF/HTML/transcript annotations
• Review: findings linked to evidence/artifacts
• Runtime: the execution log behind each version

Data and Uploads

Common families:

Tables/statistics; PDF/JSON/code; images; HDF5; genomics; single-cell; molecular; medical imaging; proteomics; graph, MATLAB, and NumPy data; ZIP/TAR archives.

No per-file size limit on any plan; files above 64 MB upload directly to storage in parts.

Archives: 500 MB compressed / 2 GB extracted safety limits.

Literature

• Deep Search runs one ranked question over the scholarly literature and returns up to 100 results.
• ClinicalTrials.gov and patents are specialized connectors.
• Use study filters, deduplication, Library, collections, watches, evidence, and networks.
• Open available full text in Workspace; otherwise CATO uses available metadata/abstracts.
• Send selected works to Chat as bounded context.
• Verify every citation and full-text-supported claim.

Memory

• Project: study-specific stable context
• Global: durable personal preferences
• Edit, pin, or delete entries.
• Control recall, auto-learning, and cross-project use.
• Never treat memory as scientific evidence.

If a Run Appears Stuck

1. Reopen the same conversation and allow reconnection.
2. Check the response or background-job status card.
3. Cancel once if the task is no longer valid.
4. Check Settings → Billing for usage boundaries.
5. Retry only after confirming no active duplicate exists.
6. Send support IDs, time, and redacted errors, not data or secrets.

Validation Checklist

• Correct source and artifact versions
• Expected rows, variables, units, and exclusions
• Code, dependencies, seed, and environment
• Assumptions, diagnostics, uncertainty, sensitivity
• Figure data, labels, scales, and accessibility
• Citation identity and claim support
• Export fidelity and reproducibility