Research State

The research product treats disk as the long-term source of truth. Context is just working memory.

All persistent project state lives under:

.omr/research/

Canonical Structure

.omr/research/
├── project.json            # current stage and metadata
├── pipeline.yaml           # gate configuration
├── HANDOFF.md              # resumable narrative handoff
├── dispatch-log.jsonl      # dispatch history
├── log.jsonl               # decisions / transitions / verdicts
├── infrastructure.yaml     # optional symlink to shared infra
├── ideas/
│   ├── candidates.md
│   ├── screening-1a.md
│   ├── idea-card.md
│   └── verification-1c.md
├── baseline/
│   ├── summary.md
│   ├── experiments.md
│   ├── method-details.md
│   ├── code-notes.md
│   └── reproduction-checklist.md
├── results/
│   ├── baseline-reproduction.md
│   ├── implementation-notes.md
│   ├── training-status.md
│   ├── analysis-verdict.md
│   ├── figures-manifest.md
│   └── postmortem.md
├── review/
│   ├── novelty-search.md
│   ├── r1-methodology.md
│   ├── r2-experiments.md
│   ├── r3-novelty.md
│   ├── reference-check.md
│   └── format-check.md
└── paper/
    ├── main.tex
    └── figures/

What Matters Most

File	Why it matters
project.json	The external stage mirror for status and recovery
pipeline.yaml	Gate policy per stage
HANDOFF.md	Human-readable resume context
dispatch-log.jsonl	What got dispatched
log.jsonl	Why the system made a decision

Resume Model

When a session ends, the next session should be able to recover from:

1. project.json
2. pipeline.yaml
3. HANDOFF.md
4. specific files listed in HANDOFF.md -> read_files

That is the intended recovery contract. A research session should never depend on old chat context alone.

Trace IDs

Each batch of related work should carry a trace ID such as:

run-a3f8

It should appear in:

• dispatched tasks
• log.jsonl
• outputs produced by downstream agents

This is what makes debugging and recovery tractable.

Next

• Commands & Interaction
• Pipeline Overview