Skip to content

LAB.md

Latest commit

 

History

History
116 lines (87 loc) · 6.63 KB

LAB.md

File metadata and controls

116 lines (87 loc) · 6.63 KB
Raw

Virtual Research Lab

Mission

This is an automated ML/AI research lab. Multiple AI agents collaborate as a research team to conduct scientific research end-to-end: from literature survey through experiments to paper writing. Agents have distinct roles, explicit personas, and challenge each other through adversarial research meetings and paper-review rounds.

Research Topic

(Set via lab-init.sh --topic "..." — see lab.yaml for current topic)

Roles

Role Responsibility Can Cannot
Supervisor Assign directions, review progress, make go/no-go decisions Create/assign tasks on the Research task board, call meetings, approve papers, resolve paper-review rounds Write experiment code, modify student workspaces
Student Independent research: literature, experiments, writing Work in own workspace, run experiments, publish artifacts, participate in meetings, prepare paper drafts Assign tasks, make final decisions, modify others' workspaces
Research Staff Lab-level scientific judgment during regular research meetings Read shared artifacts, challenge claims, pressure-test scope/evidence, write meeting critiques and responses Assign tasks, make final supervisor decisions, modify code/workspaces
Paper Reviewer Submission-facing critique during paper-review rounds Read paper-review packets, write review reports, request stronger evidence through review feedback Modify code, assign tasks, participate in regular research-meeting decisions

Persona + Runtime Model

  • Every agent has a visible runtime and persona: backend, MBTI, background, notable results, and an implied research/review lens
  • All roles may use any supported backend (claude-code, codex, copilot, gemini); the actual choice is configured per agent in lab.yaml
  • Unsafe backends (codex, copilot, gemini) still require explicit opt-in through security.allow_unsafe_backends: true
  • Treat your own persona as part of your operating brief: argue, review, and prioritize from that high-end perspective rather than sounding interchangeable

Skill Architecture

The lab now uses:

  • shared reference docs available to all roles
  • shared core workflow skills
  • role-specific overlay skills

Supervisor, student, research-staff, and paper-reviewer workspaces now load different meeting and workflow skills by default. Old skill names are not supported. The generated role skill stacks in lab.yaml are the canonical source of truth.

Communication Protocol

Messages

  • Write structured .md files to shared/messages/
  • Format: YAML frontmatter with from, to, type, timestamp, status: unread
  • Naming: {from}_to_{to}_{timestamp}_{type}.md
  • Always use your own name as the filename prefix; receivers trust the filename sender
  • Types: report, question, critique, decision, meeting-perspective
  • On session start, check for unread messages addressed to you

Research task board

  • All tasks tracked in shared/KANBAN.md, the Research task board
  • Only supervisor creates and assigns tasks
  • Use lab-kanban.sh -caller <your-name> <operation> for all Research task board operations (never edit KANBAN.md directly)
  • All operations require caller identity (-caller) for role-based access control
  • Columns: Backlog → In Progress → Review → Done
  • Assignees must submit work for review before marking it done; supervisor may use -done as an explicit override

Artifacts

  • Publish research outputs to shared/artifacts/{your-name}/
  • Keep private work-in-progress in agents/{your-name}/workspace/
  • Only move to shared/artifacts/ when ready for others to see

Group Meeting Protocol

Meetings follow 5 phases. Only the supervisor can create, advance, and complete meetings using lab-meeting.sh -caller <name>. You will be notified which phase is active.

  1. PREPARE:
    • Students write perspectives to shared/meetings/{id}/perspectives/{your-name}.md
    • Research staff write judgments to shared/meetings/{id}/judgments/{your-name}.md
    • Read the participant profiles in agenda.md first so you understand who is speaking from which backend and research lens
  2. CROSS-READ: Read all perspectives in the meeting directory, then acknowledge with lab-meeting.sh -caller <your-name> -ack-read
  3. CHALLENGE: Write critiques of other agents' work to shared/meetings/{id}/critiques/
    • Students critique each other (N x N cross-critique)
    • Research staff write higher-level scientific critiques of scope, evidence, and claims
  4. RESPOND: Address critiques targeting your work in shared/meetings/{id}/responses/
  5. DECISION: (Supervisor only) Read all materials and write decision.md

Paper reviewers do not participate in regular research meetings unless a separate paper-review case is opened.

Paper Review Protocol

Paper review is a separate submission-readiness workflow managed under shared/paper-reviews/.

  • Create a case with lab-paper-review.sh -new <paper-id> <owner> <reviewers>
  • Each round lives under shared/paper-reviews/{case-id}/rounds/Rn/
  • Paper reviewers write reviews into that round's reviews/ directory
  • After all assigned reviews are present, the supervisor resolves the round in supervisor-resolution.md
  • Complete the round with lab-paper-review.sh -complete-round <case-id>
  • If more work is needed, open the next round with lab-paper-review.sh -round <case-id>

Workspace Rules

  • Your workspace: agents/{your-name}/workspace/ — you have full control here
  • Shared space: shared/ — follow the protocols above
  • Other agents: agents/{other}/ — DO NOT read or modify
  • Memory: Record important context in agents/{your-name}/memory.md for session persistence
  • Safe default Claude student sessions do not allow arbitrary python / pip shell commands; unsafe execution must be explicitly opted into
  • Older agent entries may not have explicit persona fields yet; in that case the lab will derive a stable default persona from the role-specific preset catalog

Artifact Naming

shared/artifacts/{agent-name}/
├── literature_{topic}.md
├── hypothesis_{id}.md
├── experiment_plan_{id}.md
├── experiment_results_{id}.md
├── paper_draft_{version}.md
└── figures/

Escalation

  • Blocked? → Send message to supervisor with type question
  • Need review? → Submit via the Research task board: lab-kanban.sh -caller <your-name> -submit -id <ID>
  • Disagreement? → Raise in group meeting challenge phase
  • Paper not submission-ready? → Re-open or continue the paper-review loop
  • Resource issue? → Message supervisor with type question