research

research: receipts or it doesn't ship

Advice built on training data is pattern-matching dressed as knowledge. This skill goes and finds what is true right now, checks it, cites it, and files it where the business can use it again. It is how every other skill earns the right to recommend anything.

  the question
       |
  [ tier ]      quick: minutes, 2-3 sources   standard: multi-angle
       |        deep (opt-in): fan-out + independent fact-check
       |
  [ sweep ]     best engines available, multiple phrasings
       |
  [ verify ]    corroborate across INDEPENDENT sources
       |
  [ file ]      full dossier -> wiki/research/    chat gets the gist + link

Before any step, read references/CONVENTIONS-core.md. This resolves from the installed skill directory to the plugin's shared contract, independent of the founder's working folder. Its receipts, filing, and dedupe rules govern this workflow.

Step 0: Preflight

Run the shared contract's Universal preflight. A Fresh, Partial, or Wrong-folder result stops and runs the co-founder-setup skill before anything else (re-sync for Partial), without researching or filing. On Ready, read the charter's path map and the relevant vault context.

Done when: the founder system is Ready and the dossier destination is known.

Step 1: Sharpen the question

Restate what is actually being decided, in one line, and confirm it. "Research newsletters" is not a question; "what open rate should a 2k-subscriber list expect, so we can judge ours" is. A question doing double duty gets split.

Done when: one decision-shaped question is agreed.

Apply the shared parameter gate to any business-specific premise needed to shape or interpret the research. Search the vault, then ask the founder or label the premise as an unverified working hypothesis. Search-tier mechanics are operational; the business inputs are not yours to fill.

Ask before you assert: never present a quantity, duration, sample size, date, or causal motive in prose before asking the founder for it or citing a receipt for it.

Step 2: Pick the tier

Quick: a fact to check, minutes of work, 2-3 sources. Default for small stakes.
Standard: a decision input, multiple angles, 5-8 sources.
Deep: bet-the-business calls only, and the founder opts in (it costs real time and tokens): full fan-out plus the independent fact-check pass in step 4.

Uncertain resolves upward. State the tier and, for deep, get the yes first.

Done when: the tier is stated.

Step 3: Sweep

Use the best engines this environment has, in this order of preference: dedicated search and scrape tools if available (for example, Firecrawl or similar), social-pulse tools (recent posts and community threads), then the harness's native web search and page fetching. The engines vary by machine; the process never does.

Sweep rules:

Multiple phrasings per question: the practitioner's wording, the vendor's wording, the skeptic's wording. One phrasing finds one bubble.
Prefer primary sources over roundups: the study over the listicle, the pricing page over the review, the practitioner's own numbers over the guru's summary.
Capture every candidate claim WITH its source as you go. An uncited claim does not exist.
Record the exact query, engine, search date, and result count inspected for every negative finding. A finite search supports "none found in this search," never "none exists."

Done when: each angle of the question has candidate claims with named sources.

Step 4: Verify

STOP. Read references/verify-checklist.md before grading any claim: it holds the three traps that make false claims look corroborated, and skipping it is how slop gets certified.

Grade each claim: verified (independent sources agree), stated (one source claims it; carried as a quote, never as truth), contested (sources disagree; both sides carried). Deep tier adds the adversarial pass: each load-bearing claim gets a fresh check whose only job is to break it, using the checklist.

Research findings may be external facts without pretending to know what this business should do. A benchmark applied to this business, a "should" answer, or any other consequential recommendation uses the shared receipts contract in references/CONVENTIONS-receipts-contract.md: one internal receipt and one verified external source. One floor missing: deliver the factual findings, then write the recommendation in the canonical PARKED: internal receipt missing or PARKED: external receipt missing shape and name recall or research as the next action. Evidence gaps never move to "Needs your call".

Done when: every claim in the synthesis carries a grade and a source.

Run the shared durable claim admission gate before filing. Reopen every cited page and confirm the passage supports the written claim. Preserve the short supporting quote beside the citation. A source that disappeared or does not contain the evidence is removed; the claim is downgraded to stated with its honest source or omitted. Negative findings carry the saved search scope rather than upgrading verified search absence into a universal fact.

Use the shared verified-absence record for every zero-result finding. It may establish the bounded gap and support a PARKED recommendation; it never supplies the external leg for a SURVIVES/DIES call or proves that the missing fact does not exist.

Step 5: File and deliver

Write the full dossier to wiki/research/<slug>.md with the universal wiki frontmatter, the base five plus sources (type: research, title, status from the overall confidence, created, updated, sources: list), then the question, the synthesis, every claim with its grade and citation, Parked recommendations for missing receipt legs, and a "Needs your call" section only for judgment forks the evidence cannot settle. Include a Search record with the exact queries, engines, dates, and inspected-result counts. Update index.md, link hub/research.md, and append ## [date] research | <question> to log.md.

In chat: the answer in 3-6 lines, the strongest receipt, the link to the dossier. Never paste the dossier into chat; next month's question starts from the file, not from a re-run.

Mandatory filing gate

Run scripts/graph-audit on every touched durable file, including the complete dossier.
Fix every violation and rerun it. The dossier and research event remain incomplete until it exits zero.

Done when: the dossier is filed, indexed, logged, and the founder has the gist with its link.