Research Operating System

Investigate the internet as a mixed evidence environment, not as a pile of search results. Every task follows the same core cycle; domain-specific tactics come from reference packs loaded on demand.

Core Cycle

FRAME → MAP → FRONTIER → DEEPEN → CHALLENGE → SYNTHESIZE

1. FRAME

Define the task before searching:

• Decision question or judgment target
• Key claims that may need confirmation or falsification
• Important stakeholders and their incentives
• Source classes likely to matter
• What would change the answer

Load reference packs by task type:

Task signal	Load
Industry, company, career, opportunity, operator reality	[references/business.md]
Papers, benchmarks, state of the art, technical lineage	[references/academic.md]
"Is this true?", fact check, verify claim, debunk narrative	[references/claim-verification.md]
Mixed task	Load multiple packs

When the task is a decision between options (choose A vs B), use the core cycle but frame everything around decision criteria — see the Decision Support output variant at the end of this file.

2. MAP

Run a broad first pass across multiple evidence classes using WebSearch.

Aim to surface:

• Official narrative
• Behavioral reality (what actors do, not just say)
• Operator or expert view
• Lived experience
• Adversarial or contradiction path

Budget: 4-6 searches. Stop and write a field map before going deeper.

Search in English first for most topics. When the claim is region-specific, search both English and the relevant local language before synthesizing.

3. FRONTIER

Convert the map into 2-4 active leads. For each lead record:

1. What it is
2. Why it matters
3. Which source class it touches
4. Next action
5. What claim it could confirm or weaken

Branch types:

• Claim

  — a statement needing support or falsification

• Stakeholder

  — a company, lab, recruiter, operator, or critic that matters

• Source-gap

  — a missing source class or primary document

• Artifact

  — code, docs, dataset, filing, benchmark, or legal document

• Contradiction

  — a rebuttal, complaint, failed case, or competing narrative

• Trajectory

  — a trend in hiring, pricing, adoption, or regulation

Scoring criteria (qualitative, not numeric):

• Source proximity — how close to a primary or artifact source?
• Decision relevance — does resolving this change the answer?
• Information gain — new understanding vs more of the same?
• Coverage gap — blind spot in current map?
• Contradiction value — could it weaken the thesis?

Selection rules:

• Early: keep 2-4 branches for coverage
• Middle: prune branches that no longer change the answer
• Late: converge only after both support and challenge branches were checked
• Always keep at least one branch that could weaken the working thesis

Breadth ceiling: Stop widening and start deepening when the major actors are visible, relevant source classes are known, main debates are named, and at least one contradiction path exists. Many URLs is not coverage. Repeated narratives are not independent evidence.

4. DEEPEN

For each active branch:

WebSearch / WebFetch → EXTRACT → CHALLENGE → NEXT JUMP

Use WebFetch to read substantive sources (articles, docs, reports) rather than relying on search snippets. After each step log:

• Finding
• Source and quality
• Contradiction or caveat
• Gap
• Next jump

Budget: 2-3 fetches per branch. Do not keep widening with generic searches once lead-driven work has started.

5. CHALLENGE

For every strong thesis, actively search for:

• Direct rebuttal or criticism
• Failed example or contradictory user experience
• Short report, lawsuit, or adversarial analysis
• Competing explanation
• Evidence that the signal is marketing rather than reality

Minimum challenge coverage before synthesis:

• What source class most disagrees with the current view?
• Which stakeholder has incentives to overstate this thesis?
• Which stakeholder has incentives to attack this thesis?
• Were both operator reality and official narrative checked?
• Was behavioral evidence checked?

Budget: 2-3 searches.

6. SYNTHESIZE

Use Write to create

research-{topic}.md

in the current directory. Update it throughout the run so reasoning is visible.

Separate findings into:

• Confirmed — supported by 2+ source classes
• Likely — supported but with gaps
• Contested — sources disagree
• Unresolved — insufficient evidence
• What could change the view

Mark source quality:

✓

primary/artifact

?

secondary/incomplete

✗

contradictory/problematic

Source Classes

Use multiple classes, not a single ranking:

Class	What it is	Best for
Official	What institutions, companies, regulators say	Stated intent
Behavioral	What they actually do: hiring, pricing, commits, launches	Reality checks
Operator	What experienced practitioners say in long posts, talks	Expert view
Lived experience	Reddit, X, LinkedIn, forums, Blind, Glassdoor	Friction, day-to-day
Adversarial	Short reports, critics, lawsuits, complaints	Challenge paths
Market proxy	Fundraising, layoffs, salary, adoption, search demand	Economic weight
Artifact	Papers, repos, docs, datasets, filings, benchmarks	Hard evidence

Default triangulation for important conclusions — try to combine:

• One official or artifact source
• One behavioral or market source
• One human or adversarial source

If triangulation is impossible, say so explicitly.

Saturation Rules

Research is done when:

• Main claims are attached to high-value sources
• Support and challenge paths were both explored
• Relevant source classes were sampled, not just one ecosystem
• New searches mostly repeat known nodes
• Remaining frontier items are peripheral

Research is not done when:

• Only the supporting story was explored
• Community and official narratives sharply disagree without resolution
• Synthesis still depends on snippets or summaries
• No opposing incentives were considered

Budget

Phase	Tool	Limit
MAP	WebSearch	4-6
DEEPEN	WebSearch + WebFetch	2-3 per branch
CHALLENGE	WebSearch	2-3
SYNTHESIZE	Write	1
Total	all	~20 tool calls

If approaching budget, prioritize synthesis over more searching.

Non-Negotiable Rules

• Broad first pass means covering evidence classes, not hoarding URLs.
• Once a high-value lead appears, fetch and read it before widening again.
• After the map phase, new searches must be lead-driven, not generic rephrases.
• Social posts and community threads are valuable lead and challenge sources, not standalone truth.
• Every important conclusion needs support from 2+ source classes or an explicit note that it remains unresolved.
• Every strong thesis must run a challenge path before final synthesis.

Output Template

# Research: {topic}

## Frame
[decision question, target claims, stakeholders, evidence classes]

## Field Map
[main branches, actors, debates, source ecosystem, blind spots]

## Active Frontier
[2-4 active leads, why they matter, next action]

## Branch Log
### Branch A
[finding, source, contradiction, gap, next jump]

## Challenge Log
[what was used to attack the current thesis]

## Synthesis
[confirmed, likely, contested, unresolved, what could change the view]

## References
[important sources with URLs and source quality markers]

Decision Support Variant

When the task is choosing between options, replace the output template with:

# Decision: {question}

## Decision Frame
[options, constraints, time horizon, what matters most]

## Evidence By Criterion
[evidence gathered per decision criterion, not per random topic]

## Tradeoffs
[what you gain and lose with each option]

## Recommendation
[preferred path with explicit confidence level]

## What Would Change The Call
[conditions that would flip the recommendation]