1. Artifact-driven progress: Research is organized into 5 stages, each producing required deliverables
2. Quality gates: Each stage has explicit quality criteria that must be met before proceeding
3. Traceable evidence: Every claim must be traceable to sources with proper citation
4. Structured analysis: Use structured analytic techniques to mitigate bias and improve rigor
5. Uncertainty expression: Clearly distinguish facts, judgments, and speculation with likelihood and confidence
6. Decision-ready outputs: Deliverables are structured for immediate use by decision-makers

1. Information Collection (Stage 3):
   • Navigate to official sources, company websites, regulatory filings
   • Capture screenshots with timestamps for audit trail
   • Extract structured data from web pages
   • Archive web pages before they change or disappear
2. OSINT Verification (Stage 4):
   • Reverse image search on multiple platforms (Google Images, TinEye, Yandex)
   • Verify social media posts and UGC authenticity
   • Check geolocation using map services
   • Capture verification evidence (screenshots, page snapshots)
3. Source Archiving:
   • Take full-page screenshots of key sources
   • Capture page snapshots with accessibility tree for later analysis
   • Document page state at time of access (for reproducibility)
4. Cross-Platform Verification:
   • Navigate between multiple sources to verify consistency
   • Check multiple language versions of same content
   • Verify across different platforms (official site, news, social media)

• Always capture screenshots/snapshots when accessing sources (for audit trail)
• Use browser navigation to verify links are still active
• Take snapshots before archiving (captures full page state)
• Use browser console to check for dynamic content or hidden information

• collection-strategist: Use browser to access and archive sources during collection
• verification-expert: Use browser for reverse image search, geolocation verification, UGC checking
• evidence-librarian: Use browser to verify citations and capture source snapshots

• Task Contract: Goal, audience, time window, scope, non-goals, deliverable format
• KIQs (Key Intelligence Questions): 3-7 must-answer questions
• Success criteria: Definition of "good enough" and "unobtainable"

• Issue tree / Hypothesis set (MECE decomposition + initial hypotheses)
• Collection plan: Source map, retrieval routes, priorities, verification strategy
• Risk log: Data gaps, timeliness, compliance boundaries, conflicting evidence expectations

• Source Register: Type, time, reliability, bias risk, usable scope for each source
• Evidence Table: Claim→Evidence→Strength→Conflict→Notes
• Collection log: Query strings, timestamps, exclusion reasons (for reproducibility and audit)

• Structured analysis workbook: ACH/Key Assumptions Check/Red Team (at least 1-2 techniques)
• Key Judgments (3-7): Each with likelihood + confidence + evidence anchors
• Alternative explanations and flip conditions: What would change my mind / signposts

• Deliverable (brief/memo/table/appendices package)
• QA checklist record: Fact-checking, traceable citations, consistent uncertainty expression
• Follow-up actions: Gap list, next collection suggestions, monitoring indicators

⚠️ MANDATORY FIRST STEP: Before starting any DeepResearch project, create

AGENTS.md

in project root using

@cursor-agents-md

. This file defines project-specific instructions that all research work must follow.

1. Create AGENTS.md (MANDATORY):

   @cursor-agents-md Create an AGENTS.md file for this DeepResearch project

   • Must include project-specific research guidelines
   • Must include reminder to read

     cursor-agents-md

     before updates
   • Must be in project root directory
2. Create Task Contract: Define research question, KIQs, scope, and success criteria
3. Build Issue Tree: MECE decomposition with initial hypotheses
4. Develop Collection Plan: Source map with cross-validation strategy
5. Collect & Register: Build Source Register and Evidence Table as you collect
6. Analyze: Apply structured analytic techniques (see STRUCTURED_ANALYSIS.md)
7. Synthesize: Generate Key Judgments with likelihood and confidence
8. Deliver: Create deliverable following REPORT_TEMPLATE.md and run QA_CHECKLIST.md

research-lead

⚠️ MANDATORY PREREQUISITE: Before starting Stage 0, ensure

AGENTS.md

exists in project root. If not, create it using

@cursor-agents-md

.

• AGENTS.md exists (created using

  @cursor-agents-md

  )
• Research question is falsifiable and testable
• Non-goals are explicit
• Time window defined
• Success criteria clear
• Deliverable format specified

methodologist

research-lead

methodologist

• research-lead

  : Defines KIQs and priority

• methodologist

  : Ensures claims are falsifiable, suggests hypothesis structure

collection-strategist

• KIQs ≤ 7
• Each KIQ has corresponding claims
• Claims are falsifiable
• Priority order established

collection-strategist

verification-expert

• Each claim mapped to ≥2 source types
• Cross-validation design exists
• Retrieval routes specified
• Archiving strategy defined
• Language variants considered

collection-strategist

• collection-strategist

  : Monitors coverage, adjusts strategy
• Collectors: Execute retrieval, archive materials

• Navigate to sources and capture screenshots/snapshots immediately
• Archive web pages before they change or disappear
• Verify links are still active
• Extract structured data from web pages

verification-expert

• Original sources archived
• Screenshots/snapshots captured (browser evidence)
• Timestamps recorded
• Archive links created
• Metadata captured
• Coverage threshold met (e.g., 70% of core claims)

verification-expert

• Perform reverse image/video search across multiple platforms
• Verify geolocation using map services (Google Maps/Earth)
• Check social media accounts and capture snapshots
• Verify chronolocation using weather/timezone services
• Capture verification evidence (screenshots, snapshots) at each step

evidence-librarian

• UGC verified (source, time, location, originality)
• Images/videos geolocated (if applicable)
• Chronolocation verified
• Consistency checks passed
• Verification log complete
• Browser-captured evidence included (screenshots, snapshots)

evidence-librarian

methodologist

quant-analyst

• Source Register complete
• Evidence Table built
• Each claim linked to evidence
• Conflicts explicitly recorded
• Citations traceable

methodologist

• May consult

  domain-expert

  for mechanism plausibility
• May consult

  quant-analyst

  for data consistency

quant-analyst

devils-advocate

• ACH matrix complete
• At least 2 competing hypotheses
• Discriminating evidence identified
• Key assumptions checked
• Alternative explanations evaluated

quant-analyst

devils-advocate

domain-expert

• Metrics unified
• Consistency checks passed
• Sensitivity analysis complete
• Error ranges specified
• Timeline closed (if applicable)

devils-advocate

1. Construct 2-4 counterfactual worlds
2. Identify Kill Points (evidence that if falsified, conclusion fails)
3. Create Fragility Map (which judgments sensitive to which assumptions)
4. Write Adversarial Review Memo
5. Design Decision-Failure Simulation

• At least 2 counterfactual worlds
• Kill points identified
• Fragility map complete
• Adversarial review challenges main conclusion
• Failure scenarios designed

domain-expert

methodologist

domain-expert

editor

• Mechanisms are plausible
• Industry patterns respected
• Anomalies flagged
• Context provided
• Common sense boundaries checked

editor

1. Structure: Conclusion-first pyramid
2. Language: Precise uncertainty expression
3. Format: Scannable, forwardable
4. Risk narrative: Clear and actionable

qa-gatekeeper

• Conclusion-first structure
• Key Judgments clear
• Uncertainty expressed consistently
• 3-minute grasp test passed
• Evidence anchors present

qa-gatekeeper

1. Tradecraft QA (all 8 dimensions from rubric)
2. Compliance check (ethics, privacy, permissions)
3. Risk assessment
4. Final Go/No-Go decision

• All quality gates passed
• Compliance verified
• Ethics boundaries respected
• Risk acceptable
• Ready for delivery

• PROJECT_MANAGEMENT.md: Complete project management template with all stages, gates, and tracking tables
• REPORT_TEMPLATE.md: Standard report template with all required sections
• QA_CHECKLIST.md: Pre-delivery quality assurance checklist

• STRUCTURED_ANALYSIS.md: Structured analytic techniques (ACH, Key Assumptions Check, Red Team, etc.)
• OSINT_VERIFICATION.md: OSINT verification techniques for UGC, images, videos, geolocation
• UNCERTAINTY_EXPRESSION.md: How to express likelihood, confidence, and distinguish facts/judgments/speculation

• RUBRIC.md: Evaluation rubric with 8 dimensions and scoring criteria
• ETHICS_GUARDRAILS.md: Compliance and ethics boundaries

• Subagents System Setup: Initialization and setup instructions
• Evidence-Chain Production Line: Complete 12-stage workflow with subagent assignments
• Weekly Research Rituals: Fixed weekly ceremonies
• Subagent Role Definitions: Detailed role descriptions for all 10 subagents

• Type (official/company/media/academic/UGC/database)
• Provenance (who produced, when, version)
• Access path (how obtained/paid/scraped)
• Bias risks (stakeholder interests, propaganda tendency, method limitations)
• Reliability (High/Medium/Low + rationale)
• Use limits (what it can prove)

• Claim (falsifiable assertion)
• Evidence (citation + summary)
• Supports/Contradicts (which hypothesis)
• Strength (Strong/Medium/Weak: based on method and independence)
• Alternative explanations
• Notes (gaps, next verification steps)

• Each Key Judgment: At least 2 independent sources (or 1 primary authoritative + explanation why sufficient)
• Each key number/timeline node: Must be traceable to original source or clear derivation chain
• Conflicting evidence: Must be explicitly presented with explanation of choice and remaining uncertainty

• Causal attribution / Who did it → ACH + prioritize disconfirming evidence
• Future prediction / Risk → Scenario planning + indicator framework
• Strong consensus → Devil's Advocacy / Team A-Team B
• Unstable key premises → Key Assumptions Check
• Adversary intent/behavior → Red Team (avoid mirror imaging)

• Fact: Observable, verifiable with high certainty
• Judgment/Estimate: Evidence sufficient but still probabilistic
• Inference/Speculation: Limited evidence, more logic-based

• Almost impossible
• Unlikely
• Possible
• Likely
• Almost certain

• High: Multiple independent sources, consistent, high-quality methods
• Medium: Some independent verification, partial consistency
• Low: Single source, high uncertainty, limited verification

1. Will not doing it affect the decision?
2. Is there an evidence path? If unobtainable, move to "future work", not current scope

• Register sources during collection (Source Register)
• Pull evidence from Evidence Table when writing conclusions (not from memory)

• Day 1: Task Contract + Issue tree + Collection plan (Gate 1/2)
• Day 2-4: Collection + registration + initial Evidence Table (daily Gate 3)
• Day 5: Structured analysis (ACH/Key Assumptions Check) + initial Key Judgments (Gate 4)
• Day 6: Fill gaps, handle conflicts, update confidence
• Day 7: Deliverable writing + Red Team + QA (Gate 5)

• What new "usable evidence" was added yesterday (not "what was read")
• What hypothesis/gap will be verified today
• Blockers: Can't get data? Conflicting evidence? Scope change?

• Are claims covered? Are conflicts recorded?
• Is there a tendency toward "feeling-based convergence"?
• Does collection strategy need adjustment (change source types, languages, timeline)?

• What is the most vulnerable point of your conclusion?
• Which evidence is weakest? What if it's wrong?
• Have you "missed alternative explanations"?

research-lead

methodologist

collection-strategist

domain-expert

1. Review Current Claims (10 min)
   • List all active claims from Claim Tree
   • Check: Are they still falsifiable?
   • Check: Do they still answer KIQs?
2. Scope Check (10 min)
   • Compare claims to Task Contract
   • Identify scope drift
   • Decide: Keep, modify, or remove claims
3. Priority Update (10 min)
   • Re-rank claims by decision impact
   • Identify which claims need evidence first
   • Update collection priorities

• Updated Claim Tree
• Priority Matrix
• Scope Change Log (if any)

• All claims traceable to KIQs
• No scope drift beyond Task Contract
• Priorities reflect decision needs

devils-advocate

research-lead

methodologist

evidence-librarian

1. Identify Kill Points (15 min)
   • Review current Key Judgments
   • For each judgment: What evidence, if falsified, would kill it?
   • List all Kill Points
2. Assess Fragility (10 min)
   • Which judgments have single-point failures?
   • Which assumptions are most vulnerable?
   • Create Fragility Map
3. Action Plan (5 min)
   • Which Kill Points need additional evidence?
   • Which need re-verification?
   • Assign follow-up tasks

• Kill-Point List (updated)
• Fragility Map
• Action Items

• All Key Judgments have identified Kill Points
• Single-point failures flagged
• Action plan for strengthening weak points

evidence-librarian

1. New Conflicts (5 min)

   • evidence-librarian

     reports new conflicting evidence
   • Brief description: What conflicts, why
2. Status Update (5 min)
   • Each subagent: Any conflicts discovered in their work?
   • Quick round: "I found X conflicting with Y"
3. Next Steps (5 min)
   • Which conflicts need investigation?
   • Assign to appropriate subagent
   • Update Conflict Evidence Log

• Updated Conflict Evidence Log
• Action Items for conflict resolution

• All conflicts explicitly recorded
• No conflicts "hidden" or ignored
• Action plan for each conflict

devils-advocate

1. Current State Review (20 min)

   • research-lead

     : Current Key Judgments

   • methodologist

     : Current hypothesis status

   • evidence-librarian

     : Evidence summary
2. Counter-World Construction (40 min)

   • devils-advocate

     presents 2-3 alternative explanations
   • Group discussion: Can these worlds explain the evidence?
   • Identify what evidence would distinguish worlds
3. Adversarial Attack (40 min)

   • devils-advocate

     attacks evidence chain
   • Each subagent defends their work
   • Identify weaknesses and gaps
4. Action Plan (20 min)
   • What needs re-verification?
   • What evidence is missing?
   • Update Fragility Map and Kill Points

• Adversarial Review Memo
• Updated Kill-Point List
• Re-verification Plan
• Updated Fragility Map

• At least 2 counterfactual worlds constructed
• Evidence chain weaknesses identified
• Action plan for strengthening conclusions

• devils-advocate

  must be systematic, not emotional
• All participants must engage, not just listen
• Focus on structure, not personalities

qa-gatekeeper

editor

research-lead

devils-advocate

1. Pre-Flight Check (15 min)

   • editor

     : Deliverable status

   • research-lead

     : Key Judgments summary

   • qa-gatekeeper

     : QA checklist preview
2. Rubric Review (30 min)
   • Go through 8-dimension rubric
   • Score each dimension (1-5)
   • Identify any dimensions ≤2 (must fix)
3. Failure Simulation (10 min)

   • qa-gatekeeper

     : "What if this is wrong?"
   • Identify worst-case scenarios
   • Check: Are risks acceptable?
4. Go/No-Go Decision (5 min)

   • qa-gatekeeper

     : Final decision
   • If Go: Approval for delivery
   • If No-Go: Specific feedback and return stage

• QA Report
• Rubric Scores
• Go/No-Go Decision
• Action Items (if No-Go)

• All dimensions ≥3 (ideally ≥4)
• No critical issues
• Risks acceptable
• Ready for delivery

• Monday: Morning: Claim Review (30 min); Rest of day: Normal workflow
• Wednesday: Morning: Kill Point Review (30 min); Rest of day: Normal workflow
• Friday: Morning: Conflict Evidence Stand-up (15 min); Afternoon (bi-weekly): Devil's Day (2 hours)
• Before Delivery: 24-48 hours: QA Pre-Gate (1 hour)

• Claim Review → Updates Stage 1 (KIQ & Claim Draft)
• Kill Point Review → Updates Stage 8 (Counter-World Attack)
• Conflict Evidence Stand-up → Updates Stage 5 (Evidence Registration)
• Devil's Day → May trigger return to Stage 6 (Structured Reasoning) or Stage 2 (Source Map)
• QA Pre-Gate → Final check before Stage 11 (QA Gate)

• Smaller teams: Combine roles, reduce frequency
• Tighter timelines: Shorten durations, focus on critical rituals
• Larger teams: Add sub-rituals, more detailed agendas

1. Skipping rituals: "We don't have time" → Leads to scope drift, missed conflicts
2. Rituals become formalities: No real engagement → No value
3. Wrong participants: Missing key roles → Incomplete reviews
4. No follow-up: Rituals identify issues but no action → Problems persist

⚠️ MANDATORY FIRST STEP: Before initializing subagents, you MUST create an

AGENTS.md

file in your project root using the

cursor-agents-md

skill. This file defines project-specific instructions for all DeepResearch work.

1. Create AGENTS.md (MANDATORY):

   @cursor-agents-md Create an AGENTS.md file for this DeepResearch project

   The AGENTS.md file must include:
   • Project-specific research guidelines
   • Code style and documentation standards (if applicable)
   • File structure and organization rules
   • Boundaries and constraints
   • Important: Add a reminder to read

     cursor-agents-md

     skill before updating AGENTS.md
   Required content in AGENTS.md:
   markdown

   > **⚠️ Important:** You must read `cursor-agents-md` skills every time before write or update this `AGENTS.md`.
   
   # DeepResearch Project Instructions
   
   [Project-specific guidelines for DeepResearch work]

2. Create workspace agents directory (if it doesn't exist):

   mkdir -p .cursor/agents

3. Copy subagent definitions from the skill:
   • Copy all

     .md

     files from

     @deepresearch/subagents/.cursor/agents/

     to your workspace

     .cursor/agents/

   • Required files:

     • research-lead.md

     • collection-strategist.md

     • verification-expert.md

     • evidence-librarian.md

     • methodologist.md

     • quant-analyst.md

     • domain-expert.md

     • editor.md

     • qa-gatekeeper.md

     • devils-advocate.md

4. Verify Cursor settings:
   • Ensure Cursor is in Nightly mode (Settings > Cursor Settings > Beta > Nightly)
   • Subagents feature should be enabled (Settings > Cursor Settings > Subagents)
5. Test subagent invocation:

   @research-lead Create a Task Contract for researching [topic]

• Stage 0-1:

  research-lead

  +

  methodologist

• Stage 2-3:

  collection-strategist

• Stage 4:

  verification-expert

• Stage 5:

  evidence-librarian

• Stage 6:

  methodologist

• Stage 7:

  quant-analyst

• Stage 8:

  devils-advocate

• Stage 9:

  domain-expert

• Stage 10:

  editor

• Stage 11:

  qa-gatekeeper

⚠️ MANDATORY FIRST STEP: Before initializing subagents, you MUST create an

AGENTS.md

file in your project root using the

cursor-agents-md

skill.

1. Invoke cursor-agents-md skill:

   @cursor-agents-md Create an AGENTS.md file for this DeepResearch project

2. Required content in AGENTS.md:
   markdown

   > **⚠️ Important:** You must read `cursor-agents-md` skills every time before write or update this `AGENTS.md`.
   
   # DeepResearch Project Instructions
   
   ## Project Overview
   [Describe your research project]
   
   ## Research Guidelines
   [Project-specific research guidelines]
   
   ## File Structure
   [How research files should be organized]
   
   ## Boundaries
   [What should never be done]

3. Verify AGENTS.md exists:
   bash

   ls AGENTS.md  # Should exist in project root

1. Locate the skill's subagent directory:
   • Path:

     @deepresearch/subagents/.cursor/agents/

   • Or:

     [skill-install-path]/deepresearch/subagents/.cursor/agents/

2. Create workspace agents directory (if it doesn't exist):
   bash

   mkdir -p .cursor/agents

3. Copy all subagent files:
   bash

   # Windows (PowerShell)
   Copy-Item "@deepresearch/subagents/.cursor/agents/*.md" -Destination ".cursor/agents/" -Exclude "README.md"
   
   # Linux/Mac
   cp @deepresearch/subagents/.cursor/agents/*.md .cursor/agents/ --exclude README.md

4. Required files to copy:

   • research-lead.md

   • collection-strategist.md

   • verification-expert.md

   • evidence-librarian.md

   • methodologist.md

   • quant-analyst.md

   • domain-expert.md

   • editor.md

   • qa-gatekeeper.md

   • devils-advocate.md

1. Open Cursor IDE
2. Navigate to

   @deepresearch/subagents/.cursor/agents/

3. Select all

   .md

   files (except

   README.md

   )
4. Copy to your workspace

   .cursor/agents/

   directory

1. Check AGENTS.md exists (MANDATORY):
   bash

   ls AGENTS.md  # Must exist in project root

2. Check files exist:
   bash

   ls .cursor/agents/*.md

3. Test subagent invocation:

   @research-lead Create a Task Contract for researching [your topic]

4. Verify Cursor settings:
   • Settings > Cursor Settings > Beta > Nightly (enabled)
   • Settings > Cursor Settings > Subagents (should show your subagents)

• Check Cursor version: Must be Nightly build
• Check file location: Must be in

  .cursor/agents/

  in workspace root
• Check file format: Each file must have valid YAML frontmatter
• Restart Cursor: After copying files, restart Cursor IDE

• Check file name: Must match subagent name exactly (e.g.,

  research-lead.md

  )
• Check YAML frontmatter: Must have

  name

  ,

  description

  ,

  model

  fields
• Check file encoding: Must be UTF-8

1. Edit subagent files in

   .cursor/agents/

2. Modify descriptions to match your workflow
3. Add project-specific instructions to system prompts
4. Restart Cursor to load changes

1. Compare versions: Check if skill subagents have changed
2. Backup customizations: Save your custom changes
3. Re-copy files: Copy updated files from skill
4. Re-apply customizations: Merge your custom changes back

• Define Task Contract: research question, boundaries, time window, deliverable format, success criteria
• Set priorities and rhythm: prioritize KIQs that most impact decisions
• Make final judgments: Key Judgments, confidence levels, risk narrative

• "Decision-backward": Ask "what will readers decide" first, then determine what to research
• "Claim-first": Write research as falsifiable claims, not topic summaries
• "Stop rules": Define what evidence is sufficient, when to stop

• Create Source Map: source type combinations for each sub-question (official/academic/industry/primary data/media/UGC)
• Design retrieval routes: keywords, language variants, exclusion terms, archiving strategy
• Control costs: prioritize "most discriminating" information

• "Source type complementarity": Each key judgment has at least 2 independent source types for cross-validation
• "High signal first": Get high-credibility/auditable primary materials first, then supplement with secondary
• "Coverage thinking": Focus on claims coverage, not link count

• Verify UGC, event materials, images, videos: source/location/time/editing traces/reuse
• Output auditable evidence chains: coordinates, screenshot comparisons, exclusion rationale, timelines

• "Falsify first, then verify": Try to prove it's fake first (reused old images, out-of-context, fake accounts)
• "Chain inference": Only advance to "auditable" degree at each step, no jumps
• "Consistency constraints": Time, geography, physical details must be mutually consistent

• Source Register (source registry), Evidence Table (claim-evidence table)
• Citation standards: each key judgment traceable to sources (with version/timestamp/location)
• Conflict evidence management: conflict points, methods, bias risks, why chosen

• "Audit perspective": Assume someone will verify line by line, can you reproduce?
• "Claim-evidence alignment": Each key sentence points to evidence; without evidence, downgrade to hypothesis
• "Conflicts don't disappear": Conflicting evidence is an asset, must be explicitly presented

• Select and facilitate structured analysis: ACH, Key Assumptions Check, Red Team, scenario planning, indicator framework
• Make reasoning process "replayable": why this explanation is better, what are key discriminating evidence

• "Competing explanations": Default at least 2 sets of hypotheses explaining the world in parallel
• "Maximize discrimination": Prioritize evidence that best distinguishes hypotheses, not most evidence
• "Bias immunity": Design processes to counter confirmation bias, premature convergence, mirror imaging

• Data cleaning, metric unification, comparable system building (e.g., market size, share, financial metrics)
• Sensitivity analysis/scenario analysis: how sensitive conclusions are to assumption changes
• Consistency checks: do numbers conflict, are timelines closed

• "Metrics before numbers": Unify definitions first, then discuss conclusions
• "Range and error": Output intervals, confidence and error sources
• "Explainable modeling": Models reveal driving factors, not "calculate a precise number"

• Provide "industry common sense boundaries": which conclusions cannot hold in reality (common mechanisms, regulation, business logic)
• Guide evidence priorities: which sources/indicators are more critical in this domain
• Help team identify "seemingly reasonable but mechanistically wrong" inferences

• "Mechanism testing": Not just "is there evidence", but "does it make sense mechanistically"
• "Anomaly identification": When seeing signals violating industry patterns, trigger re-verification

• Package research into usable deliverables: conclusion-first, clear hierarchy, scannable, forwardable
• Control language quality: qualifiers, risk warnings, consistent uncertainty expression
• "Reader experience": readers can grasp key judgments in 3 minutes, understand basis in 10 minutes

• "Reader bandwidth": Information density designed for decision-maker's time
• "Conclusion-reason-evidence pyramid": Each layer stands independently
• "Semantic precision": Treat "possible/likely/almost certain" as engineering specs, not rhetoric

• Tradecraft QA: objectivity, source transparency, alternative explanations, uncertainty, logical consistency
• Compliance and ethics boundaries: privacy, permissions, gray data, misinformation risk
• Failure plans: how to downgrade delivery when evidence insufficient, how to declare gaps

• "Prevent disasters before adding points": Research's most expensive cost is wrong conclusions causing decision losses
• "Gatekeeper not debater": Doesn't make conclusions for you, but ensures you're qualified to make conclusions
• "Revocability": Any conclusion must allow future updates with new evidence

• Construct counterfactual worlds: 2-4 "completely different but still self-consistent" world versions for current main conclusion
• Attack evidence chain: identify single-point failures, sensitivity points, unverifiable reasoning steps
• Trigger critical re-verification: specify "Kill Points" where if evidence is falsified, entire conclusion must restart
• Failure simulation: design 3-5 "research failure decision disaster" scenarios

• "Adversarial intelligence assumption": Assume you're seeing selectively exposed information, systematically manipulated
• "Disconfirming evidence priority": A conclusion doesn't need more supporting evidence, it needs disconfirming attempts that still can't kill it
• "Failure backward": Ask "three years later, where is this project most likely to fail on which assumption"
• "Single-point failure sensitivity": Particularly dislikes key judgments supported by only one piece of evidence

• Stage 4 (Verification) + Stage 5 (Evidence Registration): As evidence is verified, librarian can start registering
• Stage 6 (Structured Reasoning) + Stage 7 (Data Consistency): Can run in parallel, then reconcile
• Stage 8 (Devil's Advocate) + Stage 9 (Domain Expert): Can run in parallel, then reconcile

• Stage 6 → Stage 2: If evidence insufficient, return to collection
• Stage 8 → Stage 6: If Kill Points identified, may need re-analysis
• Stage 11 → Stage 10: If QA fails, return to editor
• Stage 11 → Any Stage: If major issues found, return to appropriate stage

• Solo Researcher: research-lead + methodologist + editor
• Two-Person Team:
  • Person A: research-lead + collection-strategist + evidence-librarian
  • Person B: methodologist + verification-expert + editor + qa-gatekeeper
• Three-Person Team: Distribute roles more evenly