autolab-research-loop

name	Autolab Research Loop
description	This skill should be used when the user asks to "run research", "start research loop", "autolab loop", "autonomous research", "research marathon", "run experiments autonomously". It provides the protocol for autonomous research iteration using Autolab.
version	0.1.0

Execute the Autolab autonomous research loop. Each iteration follows this exact cycle:

Step 0: Orient

Read the current state before doing anything:

• Read JOURNAL.md for past iterations' learnings
• Read DISCOVERIES.md for verified novel findings
• Run PYTHONPATH=src python3 -m autolab.cli status for experiment counts
• Read research_plan.yaml for current question decomposition
• Read .autolab/state.json for machine-readable progress
• Run git log --oneline -20 for recent commits

Step 1: Hypothesize

• Review research_plan.yaml for active questions
• If no active questions, decompose the research directive (from autolab.yaml) into 3-5 testable questions and write them to research_plan.yaml
• Form a testable hypothesis targeting the highest-priority runnable question
• Write the hypothesis in JOURNAL.md BEFORE proceeding to design

Step 2: Design Campaign

Choose a path:

• Path A: New campaign YAML in campaigns/ (parameter exploration)
• Path B: Code change in experiments/ + campaign (algorithmic exploration)
• Path C: Fundamentally different approach (moonshot)

Check the moonshot ratio in .autolab/state.json. If below the target in autolab.yaml (default 50%), design a moonshot campaign.

Campaign YAML format:

version: 1
name: descriptive_name
hypothesis: "What I expect to find"
question: q1
moonshot: false
runner:
  backend: local
  command: "python experiments/script.py --param {param}"
  working_dir: .
  timeout_seconds: 3600
defaults:
  fixed_param: value
grid:
  variable_param: [val1, val2, val3]
metrics:
  primary: metric_name
  direction: maximize
  collect:
    - name: metric_name
      pattern: "Metric: ([\\d.]+)"
stopping:
  window: 3
  threshold: 0.05
  max_failures: 3

Step 3: Execute

PYTHONPATH=src python3 -m autolab.cli run campaigns/<name>.yaml

If a campaign fails twice, skip it and try a different approach.

Step 4: Analyze & Record

Query results:

PYTHONPATH=src python3 -m autolab.cli results --campaign <name> --metric <primary> --top 5

• Confirm or refute the hypothesis
• Update JOURNAL.md with iteration number, hypothesis, results, key learnings
• If a novel finding: add to DISCOVERIES.md using the template (include prior art search)
• Update research_plan.yaml — answer questions, spawn sub-questions
• Update .autolab/state.json — increment iteration, update moonshot counts
• Commit: git add -A && git commit -m "research: <description>"

Escape Protocol

If consecutive_no_improvement in .autolab/state.json reaches 3:

1. Search the web for alternative approaches
2. Write a "devil's advocate" journal entry
3. Add a new research question to research_plan.yaml
4. Design a moonshot campaign testing the opposite assumption

Completion

Continue iterating until the targets in autolab.yaml are met (check with cat .autolab/state.json).