Pipeline Overview

The research product uses a research-specific stage machine:

idle → ideation → baseline-digestion → design → implementation → training → analysis → writing → review → done

Gate Model

The runtime gate model is:

Gate	Meaning
human	Pause for explicit approval
auto	Continue automatically
default	Use the repo's default policy for that stage

There is no runtime auto-judge gate in the current product path.

Stage Summary

Stage	Main outputs	Primary agents
ideation	candidate ideas, idea-card, verification	Scout, Judge
baseline-digestion	baseline structured files	Scout
design	research-contract, experiment-plan	Planner, Judge
implementation	baseline reproduction, implementation notes	Coder, Planner
training	training status and experiment outputs	Coder
analysis	analysis-verdict, postmortem if needed	Judge
writing	paper draft, figures manifest	Writer, Coder
review	novelty search, reviewer outputs, format/reference checks	Scout, Judge, Coder

Key State Files

The pipeline depends on these files:

• .omr/research/project.json
• .omr/research/pipeline.yaml
• .omr/research/HANDOFF.md
• .omr/research/dispatch-log.jsonl
• .omr/research/log.jsonl

Detailed Stages

1. Ideation
2. Baseline Digestion
3. Design
4. Implementation
5. Training
6. Analysis
7. Writing
8. Review