verified-research

# Verified Research ## Overview Research skill that applies a RED→GREEN verification cycle to every claim before it enters the report. Inline verification phases catch 7 systematic error types before they compound into thesis-level failures. ## When to Use - Comparative analysis across repos, systems, or approaches - Investigating unfamiliar codebases or academic literature - Any research where the report will inform decisions or be cited later - When accuracy and verifiability matter more than speed ## Procedure ### Phase 1 — Scope (RED) Define the research boundary before collecting anything: 1. State the research questions explicitly. 2. List known sources (repos, papers, docs, APIs). 3. Identify what you do NOT know — produce a typed gap list. 4. State your working thesis as a **hypothesis**, not a conclusion. ### Phase 2 — Collect (GREEN) Fill gaps from **primary sources only**: 1. For each source: explore the **full directory tree** — hooks/, rules/, tooling/, tests/, config — not just skills/ or src/. 2. Every factual claim gets an exact citation (file:line, URL, or API response). 3. Every number is verified at the primary source (API for citations, directory listing for counts). 4. Check publication/deprecation status of every artifact cited. ### Phase 3 — Verify (inline adversarial) Before any claim enters the report, apply the **7-error checklist**: | # | Error type | Check | |---|-----------|-------| | 1 | **Inflated statistics** | Is this number from the primary source or from memory? Verify at API/file. | | 2 | **Misattributed framing** | Is this what the source says, or my interpretation? Label analytical frames explicitly. | | 3 | **Missing counter-examples** | For every "all X do Y" claim — have I searched for exceptions in the same dataset? | | 4 | **Omitted infrastructure** | Did I explore ALL top-level directories in each source, or just the obvious ones? | | 5 | **Status confusion** | Is this artifact published, draft, deprecated, or personal? Check markers. | | 6 | **Academic mischaracterization** | Is this a completed study or a protocol/proposal? Check tense and methodology. | | 7 | **Recommending existing features** | Does the target already implement this? Check before recommending. | ### Phase 4 — Hunt counter-examples For each comparative or universal claim: 1. Search the same source for evidence that **contradicts** the claim. 2. If counter-examples exist, either qualify the claim or remove the universality. 3. Document counter-examples in the report — they strengthen credibility. ### Phase 5 — Scan omissions Before synthesis, ask: 1. What directories/files in each source did I NOT examine? 2. What published artifacts are missing from my analysis? 3. Are there cross-references between sources I haven't checked? (attribution lines, import statements, citations) 4. Does my thesis survive the attribution check? (Are repos truly independent, or does one cite the other?) ### Phase 6 — Audit framing Review the draft for analytical honesty: 1. Separate "source says" from "I observe" — use explicit labels. 2. Mark analogies as analogies, not established knowledge. 3. State the thesis as supported-by-evidence, not proven-by-evidence. 4. Attach a confidence level (Very High / High / Medium / Low) to each major claim. ### Phase 7 — Synthesize Produce the report with: a **lineage section** (derivation vs. convergence vs. independent), **per-claim citations** (file:line or URL), a **confidence table** (Very High / High / Medium / Low per major claim), **counter-examples included**, and **analytical frames labeled** ("this report's characterization" vs. "the source's term"). ## Verification 1. Every statistic traces to a primary source (not model memory). 2. No universal claim lacks a counter-example search. 3. Recommendations have been feasibility-checked against the target. 4. The thesis survives the attribution/cross-reference check. 5. Analytical frames are labeled, not presented as source claims. ## Anti-patterns - **Single-pass research** — collect and synthesize without verification. - **Confirmation-only collection** — gathering only supporting evidence. - **Memory-sourced numbers** — counts and versions from model memory are wrong more often than right. - **Thesis-before-evidence** — concluding before checking for attribution or derivation between sources. > Doc-only workflow. Informed by adversarial validation of `docs/research/context-solicitation-skill-design-patterns.md`.