research.md

The decision forge. An MCP server that enforces evidence-graded, phase-gated, peer-reviewed research workflows so AI agents cannot skip rigor under time pressure.

What it enforces

Evidence gates (upstream)

Gate	Trigger
2+ sources for CONFIRMED evidence	finding_create / finding_update fail if upgrading to CONFIRMED with < 2 sources
Disconfirmation search required	finding_create / finding_update fail if upgrading to CONFIRMED without documenting what you searched to disprove the claim
Content hash for REASONED+	Source must include content_hash: proving the agent fetched and read the material
Web research nudge	Tool returns advisory when findings have no sources
Vendor-only warning	Advisory when all sources are VENDOR tier
Landscape scan advisory	Nudge on first candidate to document the full option landscape

Process gates (downstream)

Gate	Trigger
Criteria locked before scoring	candidate_score fails if decision-criteria.md not locked
No TBD on scored candidates	candidate_score fails if candidate has _TBD_ claims
Peer review before scoring	candidate_score fails if no evaluations/peer-review.md

Install

pip install research-md

Or install from source:

pip install -e ".[dev]"

MCP configuration

Add to your Claude Code config:

claude mcp add research-md --scope user -- research-md

Or add to .mcp.json:

{
  "mcpServers": {
    "research-md": {
      "command": "research-md"
    }
  }
}

Agent workflow

A typical research session follows this path:

project_set          Register project, get research_id
    |
finding_create       Record claims with evidence grades (UNVERIFIED -> LOW -> REASONED -> CONFIRMED)
    |                Tool nudges: "Use WebSearch to find sources"
finding_update       Add sources, disconfirmation search, upgrade evidence grade
    |                Gate: CONFIRMED requires 2+ sources + disconfirmation
candidate_create     Define options to evaluate
    |                Advisory: "Document the full landscape before narrowing"
criteria_lock        Freeze decision criteria weights
    |
peer_review_log      Log reviewer assessment
    |
candidate_score      Score candidates (gated on criteria + peer review + no TBD)
    |
project_decide       Record the decision with rationale

Evidence grade ladder

Grade	Meaning	Requirements
UNVERIFIED	Claim recorded, not yet investigated	None -- tool nudges toward web research
LOW	Single source or anecdotal	At least a coherent argument
REASONED	Credible source, verified consultation	1+ source with content_hash: proof
CONFIRMED	Confirmed -- validated by evidence	2+ independent sources + disconfirmation search

Source quality tiers

Each source is tagged with a tier for awareness (no hard gate):

Tier	Examples
PRIMARY	Census data, RFC specs, user study results
EXPERT	PMC papers, RAND reports, Thoughtworks Radar
SECONDARY	Blog roundups, tutorials, comparison guides
VENDOR	Company blog comparing itself to competitors

Trilogy conventions

research.md follows shared conventions with ike.md and visionlog. See CONVENTIONS.md for the full standard.

• research.md -- decide with evidence (this tool)
• visionlog -- record the decision as a contract
• ike.md -- execute tasks within those contracts

Config lives at .research/research.json (committed to git).

Targeting pattern: project_set + research_id

Every tool call requires a research_id -- the GUID from .research/research.json. This is an in-memory mapping that does not persist across MCP server restarts.

1. Call project_set with the project's absolute path
2. It returns the project's research_id (a UUID)
3. Pass that research_id on every subsequent tool call

If you call a tool without a valid research_id, the server tells you exactly how to fix it.

Project structure

Single project

my-research/
  .research/
    research.json              <- config with project GUID (commit this)
    findings/                  <- NNNN-slug.md
    candidates/                <- slug.md
    evaluations/
      decision-criteria.md     <- criteria table (lock before scoring)
      peer-review.md           <- reviewer log (required before scoring)
      scoring-matrix.md        <- generated from locked criteria + candidates

Multi-project root

research-root/
  .research/
    research.json              <- root config (lists subprojects)
  vendor-selection/
    .research/
      research.json            <- subproject GUID
      findings/
      candidates/
      evaluations/

Initialize: project_init { path, root: true } then project_init { path, subproject: "name" }. When you project_set a root, all subprojects are registered automatically.

Tools (20)

Session

Tool	Description
project_set	Register a project path, returns its GUID. Also registers subprojects if root.
project_get	List all registered projects and their GUIDs for this session.

Project

Tool	Description
project_init	Initialize project structure (single, root, or subproject).
status	Project health: evidence gate status, criteria locked, peer review, TBD count, finding/candidate totals.

Findings

Tool	Description
finding_create	Create finding with evidence grade, sources array, and disconfirmation. Nudges toward web research.
finding_list	List all findings with status and evidence grade.
finding_update	Update status, evidence grade, sources, disconfirmation, or claim. Gates CONFIRMED evidence.

Candidates

Tool	Description
candidate_create	Create candidate for evaluation. Landscape advisory on first candidate.
candidate_list	List all candidates with verdict status.
candidate_update	Update verdict (provisional/recommended/eliminated) or description.
candidate_add_claim	Add binary testable claim to validation checklist.
candidate_resolve_claim	Mark a claim Y or N (clears _TBD_).

Scoring

Tool	Description
criteria_lock	Lock decision criteria weights. Required before scoring.
candidate_score	Score a candidate against locked criteria. Gated on criteria lock + peer review + no TBD.
scoring_matrix_generate	Generate comparison table from locked criteria + scored candidates.

Peer Review

Tool	Description
peer_review_log	Log reviewer name and findings. Required before scoring.

Decision

Tool	Description
project_decide	Record the final decision with rationale.
project_supersede	Mark a decided project as superseded by new research.
research_brief	Generate a layered research brief from a completed project.
research_report	Generate a full untruncated research report.

Development

pip install -e ".[dev]"
pytest
ruff check .

License

MIT -- see LICENSE.