research

name	research
description	AnchoredAutoResearch: Phase-gated research protocol with human checkpoints, prediction discipline, knowledge sources, and autonomous execution. Manages the full research cycle: plan → execute → interpret.
version	1.0.0

AnchoredAutoResearch — Phase-Gated Research Protocol

Inspired by autoresearch. Applies autonomous iteration to research — but with human checkpoints at decision boundaries, prediction-before-execution discipline, and mandatory knowledge source citation.

Core idea: Knowledge → Plan → Predict → Execute → Interpret → Human Review → Repeat.

MANDATORY: Read Protocol Before Any Action

CRITICAL — READ THIS FIRST:

For ALL commands (/research, /research:plan, /research:execute, /research:context):

1. Read CLAUDE.md for project rules
2. Read RSD.md to determine current state (create if missing → INIT phase)
3. Read context/SOURCES.md if it exists — check for unread knowledge sources
4. Follow the appropriate protocol from the references/ directory

Subcommands

Subcommand	Purpose	Reference
/research	Main state machine — dispatches to current phase	references/phase-protocol.md
/research:adopt	Anchor the protocol to an in-progress project (existing LaTeX + git + logs) — imports as read-only Cycle 1	references/adopt-protocol.md
/research:plan	Interactive experiment design wizard (Goal→Scope→Metric→Verify→Guard)	references/phase-protocol.md (PLAN section)
/research:execute	Fast execution loop (modify→commit→verify→guard→decide→log) or manual execution	references/autonomous-loop.md
/research:context	Manage knowledge sources — add, read, index papers/notes/prior work	references/knowledge-sources.md
/research:paper	Write / edit / compile a venue-specific LaTeX paper. Side-effect only — never touches RSD.md	references/paper-protocol.md

State Machine

The main /research command reads RSD.md and dispatches based on status:

/research invoked
  │
  ├─ No RSD.md
  │     ├─ User input contains existing-project signals
  │     │  (".tex", "draft", "have code", "logs", path-like strings)
  │     │    → offer ADOPT: "You seem to have an in-progress project.
  │     │                    ADOPT it instead of starting fresh?
  │     │                    (strict / fully-auto / no)"
  │     │       ├─ User accepts → ADOPT (references/adopt-protocol.md)
  │     │       └─ User declines → INIT
  │     └─ Otherwise → INIT (ask goal, read context/, create RSD)
  │
  ├─ Phase: ADOPT → load references/adopt-protocol.md
  │
  ├─ Status: WAITING_HUMAN
  │     ├─ No explicit human approval/revision in chat or RSD → remind human, STOP
  │     └─ Approval or revision found → advance to next phase
  │
  ├─ Phase: PLAN → load references/phase-protocol.md (PLAN section)
  ├─ Phase: EXECUTE → load references/autonomous-loop.md
  ├─ Phase: INTERPRET → load references/phase-protocol.md (INTERPRET section)
  │
  └─ Status: BLOCKED → show reason, STOP

The ADOPT offer is a soft heuristic. The main state machine never runs ADOPT without explicit user consent. See references/adopt-protocol.md for detection signals, pre-flight checks, and the strict/fully-auto write-outs.

Common Rules (Apply to ALL Subcommands)

Load and follow: references/core-principles.md

Summary of non-negotiable rules:

1. RSD.md is the single source of truth — all claims written there
2. Predictions BEFORE execution — git history must show prediction commits first
3. Artifact-linked claims only — every result has a file pointer + metric
4. Append-only within cycles — past phases are immutable
5. Bounded execution only — NEVER run unbounded loops; human must review
6. No self-approval — only explicit human approval or revision in chat or RSD.md counts
7. Compile after every RSD write — run bash scripts/compile_rsd.sh
8. Git commit at every phase boundary — per-experiment commits during EXECUTE

RSD Format

Load: references/rsd-schema.md for the full RSD template and section definitions.

Dual-Format System

• AI writes RSD.md (markdown) — source of truth
• scripts/compile_rsd.sh converts RSD.md → RSD.tex → RSD.pdf deterministically
• Human reads RSD.pdf for formatted view (status badges, tables, decision boxes)
• Human can approve or revise in chat or by editing RSD.md
• AI NEVER writes LaTeX directly — only markdown

Codex Review Integration

After EXECUTE phase completes, use the code review tool to review changed files:

• Send the git diff to Codex for code review
• Write the review summary to RSD under EXECUTE section
• Human reads the Codex summary instead of reviewing code directly

Knowledge Sources

Load: references/knowledge-sources.md for the full protocol.

Before ANY experiment proposal, the agent MUST:

1. Check context/SOURCES.md for unread sources
2. Read all unread sources
3. Cite relevant sources in every proposal