deep-research

name

deep-research

description

Use for web research, comparisons, explanations that depend on external evidence, and every forecast investigation. Provides the always-on core for KIQ decomposition, source grading, evidence-yield stopping, disconfirmation, forecast inputs, provenance, and phase-aware handoffs. Detailed tradecraft and the final-dossier contract live in lazy references so repeated slash activation stays token-efficient.

Deep Research — Core Contract

Produce decision-grade research: real fetched evidence, graded provenance, explicit uncertainty, and forecast-useful structure. This core is intentionally compact because /deep-research may be activated in every independent pass. Load detailed references only when the current phase needs them.

Phase type comes first

Identify the current phase from its prompt. The phase contract overrides any final-output rule:

• Working phases — opening/source-map notes, scoped worker notes, KIQ top-ups, verification notes, actor passes, correction messages, and fan-out absorption. Return concise evidence-ledger notes and gaps. They are explicitly exempt from every final word-count, section-count, and polished-dossier floor. Do not pad or synthesize the final report.
• Final dossier synthesis only — the prompt explicitly says to write/synthesize/finalize the complete research dossier. Only this phase uses the long-form output contract in references/final-dossier-contract.md.
• Structured extraction/judging — follow the requested JSON/scorecard schema, not the prose length contract.

This distinction is mandatory. A global dossier floor applied to every pass multiplies latency and tokens without adding evidence.

Core loop

1. Decompose the brief into 3–7 Key Intelligence Questions (KIQs). Mark load-bearing KIQs, likely S1/S2 source classes, priors, and what would falsify them.
2. Search for documents, not commentary. Use web_search, then web_fetch the strongest pages in full. For forecast calibration use prediction_market_search; if unavailable, use read-only web_fetch against the public Gamma API.
3. Maintain an evidence ledger: KIQ → claim → source/title/real URL/date/tier → independence → grade → open gap.
4. Prefer primary origins, verify numbers/units/as-of dates, trace repeated claims back to their origin, and seek the strongest opposing case.
5. Check convergence after every evidence-bearing call. Continue only for a named unresolved KIQ with a credible next upgrade.
6. Before final synthesis, verify the synthesis gate and load the final-dossier reference.

Evidence discipline

• S1: original/authoritative material — official statistics, filings, laws, court records, peer-reviewed papers, transcripts, primary datasets, direct first-party statements. Preferred for load-bearing facts.
• S2: accountable high-quality reporting/analysis — major wires/outlets, serious trade press, established research bodies. Use for context and as a pointer to S1.
• S3: conditional/biased/derivative — vendor research, expert blogs, preprints, Wikipedia as a map. Corroborate or attribute explicitly.
• S4: reject — SEO/affiliate slop, anonymous/undated aggregators, unsupported statistics, synthetic rewrites. Do not fetch or cite.

Only cite sources actually fetched and read. Preserve their true URL and on-page date. Never invent a title, URL, market, quote, or future-dated document. Two outlets repeating one underlying report are one origin, not triangulation.

For each quantitative fact record the metric, value/range, unit, as-of date, definition, geography/population, and source URL. Sanity-check magnitude, currency, nominal/real basis, denominators, totals, and compounding.

For detailed query operators, document targeting, source maps, grading, number checks, adversarial analysis, and obstacle routing, load:

references/source-tradecraft.md

KIQ and evidence-yield stopping

Tool counts and source counts are safety telemetry, never quotas or quality proxies.

• KIQ done: each load-bearing claim is B2-or-better and its strongest plausible opposing case was checked; or accessible S1/S2 paths are exhausted and the remaining gap is explicit.
• Evidence delta: a new independent origin, stronger tier, verified number, resolved contradiction, reference-class base rate, watchable indicator, or genuine disconfirmation.
• Yield stop: two genuinely different angles produce no evidence delta. Stop that line and record the gap.
• Continue: at least one priority KIQ remains unresolved and the next call has a named expected upgrade.
• Never issue a near-duplicate query. Change actor, mechanism, document type, jurisdiction/language, or time window—not punctuation or word order.
• Checkpoint the ledger after every pass; later phases must not re-research settled items.

Forecast-specific requirements

Research forecasting inputs, not only present-state description:

• outside-view reference classes and base rates;
• named actors, stated versus revealed behavior, incentives, capabilities, constraints, and what each gains/loses;
• directed typed relationships among key actors, each with an evidence basis;
• causal drivers, countervailing forces, catalysts, and second-order effects;
• dated timeline and forecast horizon;
• explicit competing scenarios/hypotheses and disconfirming evidence;
• watchable indicators with thresholds/dates;
• chart-ready tables/series with definitions and provenance;
• prediction-market question, market ID, P(yes), volume/liquidity, end date, URL, and fetch time for relevant liquid markets.

Market prices are calibration anchors, not ground truth. Self-filter lexical false positives. Explain material forecast/market divergence instead of copying the price.

Synthesis gate

Final synthesis may begin only when:

• every priority KIQ is resolved or explicitly flagged;
• load-bearing claims are B2-or-better or visibly labeled weak/single-origin;
• key numbers, dates, quotes, and citation origins were checked;
• circular sourcing was removed;
• disconfirmation, key assumptions, and competing hypotheses were tested;
• forecasts have base rates, actor incentives, chart-ready evidence, and indicators.

If one item fails, run one targeted gap pass. Do not restart broad research.

Final synthesis routing

When—and only when—the prompt requests the complete final research dossier, load and follow:

references/final-dossier-contract.md

That reference contains the long-form completeness floor, structured handoff fields, citations, scenarios, market section, tables, conflicts, and failure checklist. Working phases must not load or inherit its length floor unless they are explicitly preparing final synthesis.