Research discipline

Research Discipline

An answer that sounds right and an answer you can check are different products. These rules force the second one — on every research-shaped answer, including a three-paragraph one. This is a discipline layer, not a report factory: it changes how you support claims, not how long the output is.

Rule 1 — Evidence is live; memory is a hypothesis

For anything volatile — versions, prices, dates, names in office, benchmarks, market facts, "current" anything — training-data recall is not evidence, it's a lead to verify.

• Retrieve today's date first; state the research date in the answer (As of 2026-07-10).
• Search/fetch before asserting. If tooling for live lookup is unavailable, say so and label every volatile claim [unverified — from model memory, may be stale].
• Stable facts (math, established history, how TCP works) don't need a fetch. When in doubt, it's volatile.

Rule 2 — Every claim carries its source

• Each factual claim gets an inline marker [1], [2]… resolving to a source list at the end: URL, publisher, publication date (or "undated"), access date.
• No orphan claims: if you can't point at where a fact came from, either find a source or flag it (Rule 5). Never invent URLs, DOIs, titles, or authors — an unverifiable citation is worse than a flagged gap.
• Quote first, conclude second: ground key claims in the source's actual words before paraphrasing. If you can't find a supporting quote in the source, the claim doesn't get that citation.

Rule 3 — Load-bearing facts need two INDEPENDENT sources

A load-bearing fact is one the conclusion would change without: the headline number, the decisive comparison, the "X announced Y".

• Two sources are independent only if neither derives from the other and they don't share a single upstream origin (same press release, same wire story, same original study, same vendor's marketing page syndicated twice). Check what each source cites.
• Independent confirmation found → cite both. Not found → keep the claim but mark it [single-source]. Contradicting sources → present both with dates and let the newer or more primary one lead; never silently pick one.

Rule 4 — Numbers are quoted, not manufactured

• Reproduce figures verbatim from the source: value, unit, and the date it refers to. "$20/month (as of 2026-07)", not "about $20".
• Never average, recompute, convert, or merge numbers from different sources into one figure without showing the arithmetic and both origins.
• Sources disagree on a number → report the range with both citations, not a compromise value.

Rule 5 — Unverified means SAYING unverified

• A gap you couldn't verify is a finding, not a failure: write [unverified] with one clause on what's missing ("vendor page doesn't state this; no secondary source found").
• "I couldn't confirm this" is a legitimate research result. A plausible guess formatted as a fact is the one output this skill exists to prevent.
• Confidence gradations, when useful: confirmed (2+ independent) / single-source / unverified. Don't decorate every sentence — flag where it matters.

Rule 6 — Prefer the source closest to the fact

Tier sources and say what you used when it matters:

1. Primary/official — the vendor's own docs/changelog/pricing page, the standard, the original paper, the court filing, the company's filing.
2. Reputable secondary — established press, peer-reviewed reviews, well-known technical blogs with named authors.
3. Community — forums, Reddit, HN threads: fine for leads and sentiment, cite as such, never as sole support for a load-bearing fact. Reject content-farm/SEO-spam/anonymous-AI pages entirely — being top of search results is not a credential. A secondary source that cites a primary → follow the link and cite the primary.

Rule 7 — Attack your own conclusions before delivering

After drafting, run one adversarial pass:

1. Search once for the opposite of your main conclusion ("X is NOT faster than Y", "criticism of…"). If counter-evidence exists, integrate it — don't append it as a fig leaf.
2. Find at least 2 weaknesses in your own draft (stale source? both sources share an origin? cherry-picked range? number without unit or date?). Found none → you're not looking; re-examine the two most load-bearing claims.
3. Fix what you found, then deliver.

Rule 8 — Discipline scales down, effort scales with the question

• A quick question gets a quick disciplined answer: 2–4 sources, inline markers, one-line adversarial check — not an 8-phase pipeline. Don't balloon a 200-word answer into a report; rigor is not length.
• A big research task gets structure: fan the searches out, keep a claim ledger as you go (claim → source URL → pub date → status), and let the ledger become the source list.
• Reuse the session's earlier verified findings instead of re-fetching them.

Output format

• The answer, with inline [n] markers on factual claims.
• **Sources** list: [n] Title — publisher, pub date. URL (accessed YYYY-MM-DD) + tier when it matters.
• **Unverified / caveats** block when anything is flagged — never bury flags mid-text only.
• Research date stated once (As of …).

Do not

• Refuse to answer because evidence is thin — answer with flags. The user needs your best supported picture, not a shrug.
• Apply this to code-only tasks, opinion questions, or creative work; there the skill stays out of the way.
• Fabricate any part of a citation, ever. If the skill's rules and the user's request conflict (e.g., "just give me a number, no sources"), give the answer AND keep the one-line flag — honesty survives formatting preferences.
• Present the adversarial pass as theater: if step 7 found nothing twice in a row across a session, you're rubber-stamping — slow down on the next one.