Deep Research Method

When to invoke

Answering any question that requires evidence from the web; writing a research brief; auditing a source; deciding between two conflicting facts. Do NOT use for reading a single document you already trust, or for summarising a file you have locally.

When to invoke (concrete triggers)

Trigger	Load?
"research this", "find sources for", "what is the evidence for"	YES
"two sources disagree", "which is right", "contradiction"	YES
"is this source credible", "score this source"	YES
"summarise this PDF/article" (single doc, already in hand)	NO
"what does the code do" (local repo)	NO

Procedure (numbered — follow in order)

1. Decompose the question into sub-questions before fetching anything. A sub-question that cannot be answered by a source is a scope gap, say so.
2. Score every source 0-10 before using it (table below). Only sources >=6 are kept; a source scoring 5 is discarded unless it is the only source for a sub-question, in which case it is flagged low confidence.
3. Fetch the primary, not the summary. A secondary source never carries a retrieval date that proves you read the primary. Record URL + retrieval date
   • pinned version for every claim.
4. Quote under 15 words. Anything longer is a summary with a citation, not a quote. Paraphrase bait is how citations drift.
5. Resolve contradictions by quality weight, never by majority vote and never by averaging (see references/contradiction-handling.md).
6. State confidence: high (primary + >=2 corroborating), medium (primary only), low (secondary only). Never claim high without corroboration.

Decision tables

• Quality score: authority 0-4, recency 0-3, corroboration 0-3. >=6 keep.
• Authority: official body/standards org = 4; major vendor = 3; academic = 3; established media = 2; blog/unknown = 0-1.
• Recency: <1 year for fast-changing = 3; 1-3 years = 2; 3-5 years = 1;

  5 years = 0 unless the fact is stable.

• Corroboration: 3+ independent = 3; 2 = 2; 1 = 1; none = 0. "Independent" means distinct upstreams, not distinct URLs: two pages that copy the same primary count as one. Verify with scripts/score_sources.py corroboration.
• Confidence: high = primary + >=2 corroborating; medium = primary only; low = secondary only; none = no source.
• Citation rules: every claim needs URL + title + retrieval date; a claim missing any one is downgraded one rung. Quotations <=15 words are fine with a citation; longer ones need a page/section reference. A paraphrase must change wording, argument order, or level of abstraction; otherwise it is a quote. Version-sensitive facts record the exact version seen and are re-fetched if older than 7 days. Contradictions are recorded in references/incidents.md with both sources and the resolution (see references/citation-discipline.md).

Near-miss triggers (stop and re-check before proceeding)

• A source with no retrieval date -> not evidence, re-fetch and record.
• A secondary source used as the sole evidence -> fetch the primary.
• A contradiction resolved by majority vote -> re-weight by quality.
• A fact stated without a source -> mark it as unverified.
• Two sources disagree and the answer averages the numbers -> fake precision; list, weight, state the range.
• A 2024 fact used for a 2026 question without re-checking -> re-fetch.
• A kept source with score <6 -> it was never scored.
• Two URLs that agree exactly on a value but share an upstream -> the fake middle; they are one source, not two. Count upstreams, not URLs.
• A kept source with no retrieval date or no pinned version -> it is memory, not evidence; run scripts/score_sources.py provenance to catch it.

Pitfalls from research

• P1: Source quality is judged before reading; a good-looking page can be wrong.
• P2: Retrieval dates are mandatory; memory is not evidence.
• P3: Contradictions must be surfaced, not averaged away.
• P4: Secondary sources masquerading as primary are the common failure.
• P5: Stale facts are reused because they were once verified.
• P6: A source that scores 5 is not "close enough" — it is discarded, unless it is the only source for a sub-question, in which case it is flagged low.
• P7: Averaging two sources into a middle value produces a number no source claimed; it reads as precision and carries no evidence.

Verification checklist

• Question decomposed into sub-questions.
• Every source scored 0-10 before use; only >=6 kept.
• Every source has URL + retrieval date + pinned version.
• Every claim cites a source; no unsourced claims.
• Contradictions listed and resolved by quality weight.
• Confidence stated (high/medium/low) with the reason.
• Quotations under 15 words.
• scripts/score_sources.py corroboration passes: every kept URL traces to a distinct upstream. Two URLs sharing an upstream count as one source.
• scripts/score_sources.py provenance passes: every kept source carries URL + retrieval date + pinned version.

Worked examples

Good — scored, cited, confidence-stated answer

Question: "What is the retention period for GDPR erasure requests?"

1. Decompose: (a) the statutory deadline, (b) whether it has changed, (c) any member-state derogation.
2. Score candidates before reading: the EU legal-summary page scores 10 (authority 4, recency 3, corroboration 3); a blog scores 2 and is discarded.
3. Fetch the primary legal text first; record URL + retrieval date + version.
4. Cite by short paraphrase under 15 words, with the URL next to each claim.
5. Two independent upstreams corroborate -> confidence high, stated with the reason.

Result: every claim carries URL + retrieval date + version; confidence is high only because two upstreams corroborate. Detection signal: no claim lacks a citation, and the confidence word is followed by "because ...".

Bad — unsourced claim dressed as a finding

"The model is 3x faster than the baseline." No URL, no retrieval date, no version.

Why it fails: the claim cannot be verified or re-checked. The rule is URL + retrieval date + pinned version for every claim; a claim without a source is a confidence of low and must be flagged, not asserted. Detection signal: a sentence with a number and no citation.

Bad — contradiction averaged into a fake middle

Source A (university preprint, score 8) says "one month". Source B (industry report, score 6) says "two months".

Why it fails: averaging to "one and a half months" produces a number no source claimed; it reads as precision and carries no evidence. Resolve by quality weight: cite A as primary, note B as a lower-weight disagreement, and state the range (one to two months). Never majority-vote and never average. Detection signal: the summary names both sources and gives a range, not a single number.

Bad — quoting 22 words to "prove" a point

"the results show a statistically significant improvement in latency, throughput, and accuracy across all benchmarks".

Why it fails: long quotations are paraphrase bait and easy to misquote. Keep quotations under 15 words; anything longer is a summary with a citation, not a quote. Detection signal: a blockquote longer than 15 words that is not labelled a summary.

More worked examples, each ending with the check that would have caught it, live in references/examples.md.

Related skills

• verifiable-environments — the SPEC a verifier is derived from is a research product with cited sources; a verifier's bypass audit is itself a research claim.
• skill-authoring-evals — the source-scoring rubric and the trigger-eval methodology are research products; cite them and state confidence.

References

• procedures: references/procedures.md
• citation discipline: references/citation-discipline.md
• scoring: references/source-scoring.md
• contradiction handling: references/contradiction-handling.md
• confidence ladder: references/confidence-ladder.md
• staleness: references/staleness.md
• source scoring: references/source-scoring.md
• cite fetch: references/cite-fetch.md
• examples: references/examples.md
• incidents: references/incidents.md
• SOURCES: references/SOURCES.md