Deep Research Scoping and Execution Guide

Report body

Deep Research Scoping and Execution Guide

Executive summary

With no substantive topic specified in the request, the highest-value first step is not to begin collecting sources. It is to convert the request into a research protocol: a precise question, intended use, audience, evidence standard, geographic and temporal scope, deadline, and deliverables. Doing this first prevents the two most common forms of wasted research effort: answering the wrong question comprehensively and accumulating sources without a defensible inclusion rule.

The uploaded research brief is, in fact, a useful example of what a mature specification can look like: it identifies normative perspective, principal legal targets, primary-source starting points, questions to resolve, scenarios, overclaims to avoid, required methodology, and concrete output files. This guide treats the substantive topic as unspecified, as requested, but that brief illustrates the degree of specificity worth aiming for before a genuinely deep investigation begins. fileciteturn0file0

A rigorous workflow should then proceed from primary and official evidence outward: statutes, regulations, judgments, government datasets, original research papers, trial registrations, corporate filings, standards, and first-party records first; high-quality synthesis and commentary second. For systematic evidence reviews, PRISMA provides a structured reporting framework, while the Cochrane Handbook gives detailed methods for searching, selecting, appraising, and synthesizing evidence. citeturn0search3turn0search13

The recommended default, absent other constraints, is a medium-depth investigation of roughly 15–35 researcher-hours, followed by escalation to deep research only where the evidence landscape, decision stakes, or controversy justify it. The cost figures below are planning assumptions rather than market quotations.

Scope questions

The following questions should be answered before the research protocol is frozen. A one-sentence answer to each is enough to start.

The most important fields are topic, core question, intended decision, jurisdiction, temporal cutoff, depth, and deadline. Everything else can usually be given reasonable defaults once those are known.

A compact scope specification can therefore be written as:

Topic:
Core research question:
Purpose / decision supported:
Audience:
Jurisdictions:
Time period / information cutoff:
Depth: light / medium / deep
Deadline:
Must-use sources:
Excluded sources or topics:
Required deliverables:
Citation standard:
Languages:
Budget or maximum researcher-hours:
What finding would change the working hypothesis:

Research design and source strategy

Evidence hierarchy

Research should begin as close to the underlying fact or rule as possible. “Primary” does not mean “automatically correct”; it means the source has less interpretive distance from the event, experiment, legal instrument, or statement under investigation.

For legal work, version control is especially important. The eCFR describes itself as continuously updated but “authoritative but unofficial,” so researchers should preserve version dates and, when material, verify against the relevant official publication or issuing agency. citeturn2search4 EUR-Lex provides direct access to EU legislation and the Official Journal and is the natural starting point for EU-law research. citeturn2search8turn2search11

For corporate or securities questions, the SEC's EDGAR system provides free public access to company filings and supports searches by issuer, form, filing date, and full text. citeturn3search2 For U.S. public datasets, Data.gov is the federal government's open-data catalog, although the underlying datasets normally remain with the publishing agencies; therefore, important results should be traced from the catalog to the agency source. citeturn2search0turn2search3

For biomedical work, PubMed is maintained by NCBI/NLM and contains more than 40 million biomedical and life-science citations and abstracts, making it a strong discovery database rather than a substitute for reading the underlying papers. citeturn1search0 ClinicalTrials.gov is valuable for protocols, registrations, study status, and some results, but it explicitly warns that the U.S. government does not scientifically validate every listed study; registry presence should therefore not be treated as proof of quality. citeturn3search0

Recommended databases and tools

Crossref's REST API can be particularly useful for reproducible literature inventories because it exposes publication metadata, DOI information, funding data, licensing information, updates, and related identifiers; public access does not require signup. citeturn0search0turn0search7 OpenAlex is useful for complementary citation-network and entity discovery, with REST endpoints for works, authors, institutions, sources, publishers, and related entities. citeturn1search10

Search method

A robust search begins with a concept matrix, not one long query. Break the question into:

subject × mechanism/intervention × outcome × population/jurisdiction × time

Then generate synonyms, abbreviations, historic names, relevant identifiers, and adversarial terminology.

A generic scholarly query might look like:

("core concept" OR synonym* OR "alternate phrase")
AND
(outcome OR effect OR metric)
AND
(population OR industry OR jurisdiction)

A policy or legal query should add instrument and lifecycle terms:

("policy name" OR bill-number OR regulation-number)
AND
(text OR amendment OR implementation OR enforcement
 OR guidance OR litigation OR repeal OR injunction)

An adversarial search should deliberately seek evidence against the initial thesis:

"core claim" AND
(counterexample OR null effect OR criticism OR replication
 OR exemption OR limitation OR failure)

For every central claim, perform backward chaining from citations, forward chaining to later work, an author/institution search, and a terminology search using competing descriptions of the issue. DOI metadata systems such as Crossref and entity/citation systems such as OpenAlex make those steps more reproducible than relying on one search engine alone. citeturn0search0turn1search2

Source-quality rubric

Score each load-bearing source from zero to five on the following dimensions and retain the score in the evidence table:

A high score should not turn a source into authority beyond its actual scope. An impeccably conducted study of one population does not establish an effect in another; an official press release does not substitute for the regulation it summarizes; and a registry entry does not prove that the registered study ultimately produced reliable evidence. ClinicalTrials.gov's own disclaimer is a useful illustration of that distinction. citeturn3search0

For systematic reviews, preserve the search strings, database names, dates searched, inclusion/exclusion decisions, and flow of records. PRISMA 2020 specifically supplies a 27-item reporting checklist and review-flow templates for making that process visible. citeturn0search3

Timeline and depth levels

Research-depth comparison

These are planning targets, not universal industry standards.

“Source count” should never become the objective. Ten primary documents that resolve the question are preferable to one hundred derivative articles that repeat one another.

A sensible escalation rule is:

Light → Medium when major sources disagree, the decision is consequential, or primary evidence must be interpreted.

Medium → Deep when legal/version history matters, the question spans jurisdictions, quantitative synthesis is necessary, claims are politically or scientifically contested, or the report must withstand hostile expert review.

Illustrative deep-research schedule

The following Gantt schedule assumes a project beginning Monday, September 7, 2026. Tasks overlap intentionally: contradictory evidence should be sought while the main-source review is still underway rather than after the argument has hardened.

gantt
    title Illustrative Deep-Research Workflow
    dateFormat  YYYY-MM-DD
    axisFormat  %b %d

    section Protocol
    Scope and research question       :a1, 2026-09-07, 2d
    Search protocol and criteria      :a2, after a1, 2d

    section Evidence collection
    Primary-source discovery          :b1, 2026-09-09, 6d
    Scholarly and secondary search    :b2, 2026-09-10, 7d
    Citation chaining                 :b3, 2026-09-14, 5d
    Contrary-evidence search          :b4, 2026-09-14, 6d

    section Extraction and analysis
    Evidence extraction and scoring   :c1, 2026-09-11, 9d
    Data cleaning or calculations     :c2, 2026-09-16, 5d
    Claim-evidence synthesis          :c3, 2026-09-21, 5d

    section Production
    Draft report                      :d1, 2026-09-24, 5d
    Adversarial fact-check            :d2, 2026-09-28, 3d
    Citation and version audit        :d3, 2026-09-30, 2d
    Final delivery                    :milestone, d4, 2026-10-02, 0d

The key control point is the claim-evidence matrix before drafting. Every important conclusion should have an explicit row recording the claim, supporting evidence, contrary evidence, confidence, assumptions, and unresolved uncertainty. Writing prose first and searching for citations later is substantially more vulnerable to confirmation bias.

Deliverables and templates

A research bundle should preserve not only conclusions but enough structure for another person to audit how those conclusions were reached.

Executive research summary

The executive summary should answer the decision question without forcing the reader through the methodology first.

TITLE

Research question:
[One sentence]

Bottom line:
[Two to five sentences giving the answer]

Highest-confidence findings:
• [Finding + confidence + principal evidence]
• [Finding + confidence + principal evidence]

Important contrary evidence:
[Strongest evidence that weakens the conclusion]

Key uncertainty:
[What cannot presently be established]

Decision implications:
[What follows—and what does not follow—from the evidence]

Evidence cutoff:
[Exact date/time where relevant]

The distinction between finding, inference, and recommendation should be explicit. A recommendation may depend on values or risk tolerance even where the empirical findings themselves are strong.

Annotated bibliography

A bibliography becomes much more useful when each entry records what the source actually proves and what it does not.

Source ID:
Full citation:
Canonical URL / DOI:
Source type:
Issuer / authors:
Publication or effective date:
Version reviewed:
Primary or secondary:
Relevant pages / sections / tables:

Narrow support:
[Exact proposition this source supports]

Important limitation:
[What must not be inferred from it]

Method / evidence:
[Brief description]

Contrary evidence or caveat:
[If applicable]

Quality score:
[0–100 using project rubric]

Research notes:
[Free text]

Zotero is well suited to maintaining the bibliographic layer because it supports collections, tags, metadata, notes, PDF annotation, citations, and bibliographies; its browser connector can capture records directly from source pages. citeturn1search3turn1search15

Evidence and data table

For qualitative claims:

For quantitative evidence:

Every derived number should preserve the original value, units, denominator, transformation, and source. This prevents errors such as silently mixing nominal and real currency, annual and quarterly figures, populations with different denominators, or estimates from incompatible methodologies.

Visualization brief

Charts should be designed around a proposition, not added decoratively.

Figure ID:
Question the figure answers:
Claim illustrated:
Chart type:
Data source(s):
Unit:
Time period:
Population / denominator:
Transformations:
Missing-data treatment:
Uncertainty representation:
Caption:
Interpretive caution:
Reproduction data/file:

A visualization should be rejected when the underlying comparison is not meaningful. A well-rendered chart cannot repair incompatible denominators or weak source provenance.

Recommended default visuals are limited to those that materially improve interpretation: timelines for events/version histories, line charts for trends, bar charts for categorical comparisons, scatterplots for relationships, maps for genuinely geographic effects, and evidence maps where the important result is the distribution rather than a single estimate.

Budget, risks, and quality controls

Illustrative time and cost ranges

The monetary figures below use an explicit planning assumption of $75–$200 per researcher-hour. They are not claims about prevailing market prices and exclude paid database subscriptions, specialist legal review, translation, expert interviews, transcription, records fees, and proprietary datasets.

For a constrained budget, cut breadth before evidentiary integrity. A narrow question answered with ten carefully verified primary sources is normally more useful than a broad question addressed superficially with fifty sources.

Paid access is also not automatically necessary. Crossref offers public metadata API access without signup, PubMed is freely searchable, SEC EDGAR offers free public filing access, and government repositories such as Data.gov and EUR-Lex provide public access to substantial primary material. citeturn0search0turn1search0turn3search2turn2search8

Principal risks and mitigations

The final QA pass should audit claims rather than merely prose. For each important sentence ask: What source establishes this? Is it the correct source version? Does it establish precisely this proposition? Is contrary evidence omitted? Have observation and inference been distinguished?

For especially consequential projects, use a “red-team” pass in which the reviewer attempts to invalidate the report's five most important conclusions. The final report should preserve material defeaters rather than merely noting that “limitations exist.”

Next steps

Because the substantive topic is deliberately unspecified, the optimal next action is to produce a one-page research charter before any full search begins. The minimum viable charter is:

RESEARCH CHARTER

Topic:
[What is being investigated]

Primary question:
[The one question the report must resolve]

Purpose:
[Decision, publication, due diligence, advocacy, academic review, etc.]

Audience:
[Who will use the result]

Working hypothesis, if any:
[State it explicitly so it can be challenged]

What would falsify or materially weaken that hypothesis:
[Concrete counterevidence]

Scope:
Geographies:
Time period:
Languages:
Included populations / technologies / institutions:
Explicit exclusions:

Depth:
[Light / Medium / Deep]

Evidence hierarchy:
1. Primary / official material
2. Original research
3. High-quality synthesis
4. Specialist secondary analysis

Must-use sources:
[Names or URLs]

Deliverables:
[Executive summary]
[Full report]
[Annotated bibliography]
[Evidence/data table]
[Visualizations]
[Search log]
[Other]

Citation requirement:
[Ordinary / pinpoint / legal Bluebook / APA / Chicago / other]

Deadline:
[Date]

Budget or researcher-hour ceiling:
[Amount]

Information cutoff:
[Exact date]

Success criterion:
[A sentence describing what a completed, decision-useful answer looks like]

Once that charter is fixed, the recommended sequence is protocol → primary-source collection → broader discovery → contrary-evidence search → structured extraction → source-quality scoring → synthesis → visualization → adversarial fact-check → citation/version audit → final report. For systematic-review-style projects, PRISMA's checklist and flow structure provide a useful reporting discipline; for scholarly discovery and provenance management, Crossref/OpenAlex plus Zotero form a practical open-tool stack. citeturn0search3turn0search0turn1search10turn1search3

The single most important next step is therefore to complete the research charter above—especially the primary question, intended decision, jurisdiction, cutoff date, depth, and deadline—before substantive evidence collection begins.