Research Blog Pipeline — Operator Guide

The Mindburn research lab operates a 7-stage content pipeline managed entirely by AI agents. Every publication passes through ideation, drafting, citation verification, peer review, editorial gate, formatting, and deployment before appearing on mindburn.org/research.

Pipeline Architecture

┌─────────────────────────────────────────────────────────────┐
│                    HUMAN LAYER                              │
│  program.md — research priorities, quality thresholds       │
│  Override hooks: manual publish, reject, skip, inject       │
└──────────────────────┬──────────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────────┐
│                    AGENT PIPELINE                           │
│                                                             │
│  Stage 1: IDEATION          → research-scout                │
│      ↓                                                      │
│  Stage 2: DRAFTING          → research-writer               │
│      ↓                  ↑ (revision loop, max 2 cycles)     │
│  Stage 3: CITATION AUDIT    → citation-verifier             │
│      ↓                  ↑ (revision loop, max 2 cycles)     │
│  Stage 4: PEER REVIEW       → peer-reviewer                 │
│      ↓                  ↑ (revision loop, max 1 cycle)      │
│  Stage 5: EDITORIAL GATE    → editor                        │
│      ↓                                                      │
│  Stage 6: FORMATTING        → formatter                     │
│      ↓                                                      │
│  Stage 7: PUBLICATION       → blog-publisher                │
│                                                             │
└──────────────────────┬──────────────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────────────┐
│                    INFRASTRUCTURE                           │
│  Postgres: research_artifacts, lab_runs, article_reviews    │
│  Next.js: /research/notes/{slug} with ISR                   │
│  DeerFlow: skills, MCP servers, memory                      │
└─────────────────────────────────────────────────────────────┘

Quality Thresholds

Gate	Metric	Minimum	Fail Action
Stage 1	Citation count	≥3 sources	Retry
Stage 2	Word count	≥1,500 words	Retry
Stage 3	Citation pass rate	≥85% verified	Return to Stage 2
Stage 4	Peer review score	≥6.0/10	Return to Stage 2
Stage 5	Editorial decision	PUBLISH	Return to Stage 2 or Reject
Stage 6	MDX compilation	Valid	Retry
Stage 7	Deployment	Success	Retry

Agents

Agent	Role	Skill File
research-scout	Intelligence gathering	deerflow/skills/custom/research-scout/SKILL.md
research-writer	Draft synthesis	deerflow/skills/custom/research-writer/SKILL.md
citation-verifier	Citation validation	deerflow/skills/custom/citation-verifier/SKILL.md
peer-reviewer	Adversarial review	deerflow/skills/custom/peer-reviewer/SKILL.md
editor	Editorial decision	deerflow/skills/custom/editor/SKILL.md
formatter	MDX formatting	deerflow/skills/custom/formatter/SKILL.md
blog-publisher	Deployment	deerflow/skills/custom/blog-publisher/SKILL.md

How to Change Research Direction

Edit deerflow/program.md — the human-editable meta-instruction file. Do not edit individual SKILL.md files to change topics.

The program.md defines:

• Current research priorities (Tier 1/2/3)
• Quality thresholds
• The autonomous loop logic
• Constraints for all agents

Human Override Hooks

Hook	When to Use	Access
Manual Publish	Force-publish a rejected draft	Admin
Manual Reject	Kill a pipeline run at any stage	Admin
Skip Peer Review	Time-sensitive publications	Admin
Inject Topic	Manually add a research topic	Researcher

Experiment Logging

Every pipeline run is logged as an experiment (Karpathy autoresearch pattern). The lab_runs table records:

• Pipeline run ID, stage, agent
• Duration and token consumption
• Quality metrics per stage
• Verdict (PASS/FAIL/REVISE/PUBLISHED/DISCARDED)
• Revision cycle count

This creates a persistent audit trail showing what was attempted, what worked, and what was discarded.

Triggering a Pipeline Run

The pipeline runs automatically every 24 hours. To trigger manually:

1. Use the API trigger or DeerFlow CLI to create a new lab run
2. Or inject a topic via the inject_topic override hook
3. Or start the DeerFlow runtime and prompt:

Have a look at program.md and let's kick off a new research pipeline run.

File Locations

File	Purpose
deerflow/program.md	Human meta-instructions
deerflow/pipeline.yaml	Pipeline DAG configuration
deerflow/config.yaml	DeerFlow runtime config
deerflow/extensions_config.json	Skill and MCP server registry
deerflow/skills/custom/	All 14 agent skill definitions
deerflow/mcp-servers/lab-publisher/	Lab publisher MCP server (14 tools)
content/research/notes/	Published research articles (MDX)
content/blog/en/posts.json	Blog index