Research Methodology¶
This document codifies the rules the research-theory skill follows. The goal: produce reports that are epistemically honest, calibrated, and useful — not essays that either credulously repeat claims or reflexively dismiss them.
Source tiers¶
Stronger claims require higher-tier sources. A Tier 5 source can suggest a lead but cannot establish a fact.
| Tier | Description | Examples |
|---|---|---|
| 1 | Primary documents | Declassified records, court filings, FOIA releases, contemporaneous government reports, original photos/audio/video with verifiable provenance |
| 2 | Peer-reviewed academic | Articles in reputable peer-reviewed journals; ideally not pay-to-play venues |
| 3 | Mainstream investigative journalism | NYT, WaPo, Reuters, AP, BBC, ProPublica, Guardian investigative pieces. Investigative > opinion > commentary. |
| 4 | Advocacy / partisan media | Outlets with a known editorial slant. Useful to understand what advocates argue; not sufficient to establish fact. |
| 5 | Social media / anonymous | Tweets, anonymous posts, unverified leaks. Treat as leads only. Almost never sufficient on their own. |
When citing a source, the report should note its tier (e.g. [3] NYT, 2017 (Tier 3)).
Framework: Steelman → Evidence → Verdict¶
-
Steelman the theory. Write the strongest possible version using advocates' actual framing — not a strawman. Pull from advocates' own writings, interviews, speeches where possible. This forces engagement with the theory's strongest form before any evaluation.
-
Decompose into atomic empirical claims. Distinguish factual claims ("X met Y on date Z") from interpretive ones ("This shows coordination"). Each claim is researched separately.
-
Evaluate each claim against the source tiers. Seek primary sources first. Note where evidence converges and where it conflicts.
-
Render a verdict per claim using calibrated vocabulary (below).
-
Render an overall verdict by synthesizing across claims. The theory may be partially correct: some claims supported, others contradicted, others open.
Calibrated verdict vocabulary¶
Use these terms — and only these — for per-claim and overall verdicts:
- Supported — strong evidence from high-tier sources confirms the claim
- Partially true — kernel of truth, but the theory's framing overstates or distorts what's actually established
- Contradicted — strong evidence from high-tier sources contradicts the claim
- Unfalsifiable — claim is not testable as stated (e.g. "the conspirators covered their tracks so well no evidence exists")
- Insufficient evidence — genuinely open question; evidence too thin to call
Calibrated confidence levels¶
Use these terms — and only these — for confidence:
- High — multiple independent Tier 1–2 sources converge
- Moderate — convergence among Tier 2–3 sources, or single strong Tier 1 source
- Low — sources are mostly Tier 3–4, or there's significant disagreement among credible sources
- Insufficient — too little to assign confidence
Hard rules¶
- No false certainty. Avoid "definitely", "obviously", "everyone knows", "debunked" without citing the debunking.
- No false balance. When one side has overwhelming evidence, say so. Don't manufacture symmetry where none exists.
- Distinguish "no evidence found" from "evidence of absence." These are different epistemic states.
- Steelman must precede evaluation. Always.
- "Where I could be wrong" section is required. Surface the strongest counterarguments to your own conclusion, sources you couldn't access, and assumptions you're relying on.
- Every secondary web source (Tier 3–5) needs a Wayback snapshot URL and an access date. Sources rot. Tier 1 primary sources hosted in permanent institutional archives (
*.gov, court PACER, declassified-document repositories) may recordWayback: n/aif a snapshot isn't applicable or possible. - Use the source tier annotation on every citation.