Deep Research Project Plan for an Unspecified Topic

Full report

On this page

Executive summary

This is a reusable two- to four-week deep-research framework designed to turn an initially broad topic into a defensible, source-rich analytical product. The project should proceed through five linked stages: scope definition, systematic evidence discovery, structured extraction, analysis and adversarial checking, and publication-ready synthesis. Primary and official evidence should carry the greatest weight; secondary literature should be used for interpretation, controversy mapping, and discovery rather than as a substitute for original sources.

The method is deliberately topic-agnostic. It can support technology, health, environmental, economic, or policy research and can scale from a focused 10–20 page assessment to a larger evidence review with datasets, sensitivity analysis, and a slide deck. The attached commissioning brief provides one concrete way this framework could be instantiated: it specifies a legal-policy investigation of automated expression, bot disclosure, synthetic-media labeling, political communication, and machine association, with unusually strong requirements for primary legal texts, status verification, adversarial analysis, and structured research artifacts. For present purposes, however, the subject remains open.

A strong final research product should answer four things clearly: What is known? How strong is the evidence? Where does it disagree or remain uncertain? What follows for decisions or policy? Every important conclusion should be traceable to evidence, and observations should be kept separate from inference, forecasting, and recommendation.

Scope and research questions

The first deliverable should be a one-page research charter that converts the eventual topic into a bounded question using the pattern:

For [population/system/jurisdiction], how does [intervention, technology, condition, or policy] affect [specified outcomes] over [time period], compared with [baseline/alternative], and with what uncertainty, distributional effects, costs, and implementation constraints?

Possible starting topics include:

The policy option above is especially close to the uploaded brief, which asks whether automated or nonhuman expression is sometimes treated as presumptively deceptive and calls for close analysis of disclosure, impersonation, political-communication, and association rules.

Once a topic is selected, the project should define a small hierarchy of questions rather than attempting to research everything.

Primary question: What is the best-supported answer to the decision problem?

Evidence questions: What high-quality primary evidence exists? How comparable are studies or datasets? Where are important gaps, contradictory results, or measurement problems?

Causal questions: What mechanisms plausibly connect the intervention or condition to the observed outcome? Which links are demonstrated versus inferred?

Comparative questions: Which alternative technologies, policies, interventions, jurisdictions, or baselines produce different results?

Distributional questions: Who receives benefits, costs, risks, or decision-making power?

Forward-looking question: Which conclusions are robust to plausible future scenarios, and which depend on assumptions likely to change?

Three scope levels make the project easy to resize:

The default should be standard scope, narrowed to one primary question, approximately three to five subsidiary questions, a defined geography, a defined time period, and explicit exclusions.

Methodology

Literature and document review

The review should use a documented search protocol rather than an informal collection of interesting sources.

First, translate the research questions into search concepts, synonyms, technical terms, acronyms, and controlled vocabulary where appropriate. Searches should combine a precision strategy for the exact question with a recall strategy designed to expose unexpected terminology.

Second, search in layers:

1. Primary sources: statutes, regulations, agency data, trial registrations, original datasets, standards, technical documentation, court decisions, company filings, original experiments, and peer-reviewed empirical papers.
2. High-quality syntheses: systematic reviews, meta-analyses, national-academy reports, major institutional reviews, and authoritative assessments.
3. Interpretive and adversarial material: scholarly critiques, methodological disputes, expert testimony, civil-society research, industry analyses, and strong arguments against the emerging conclusion.

Third, apply explicit inclusion criteria. A source should ordinarily be retained because it bears directly on a research question, supplies primary evidence, materially changes confidence, explains methodology, or presents a serious counterargument. Search-result snippets and publicity materials should lead to underlying documents rather than substitute for them.

Fourth, record exclusions. Examples include superseded versions, duplicate datasets, studies without relevant outcomes, uncontrolled anecdotes presented as causal evidence, or material whose provenance cannot be established.

Data collection

A structured extraction sheet should be created before detailed reading. Each retained item should receive a stable source ID and fields for provenance, method, sample, jurisdiction, dates, variables, results, limitations, conflicts of interest, and the precise claim it supports.

For quantitative research, preserve:

raw source → transformation → analytical dataset → calculation → chart/table → claim

That lineage makes later verification substantially easier.

The appropriate data source depends on the eventual topic. For biomedical work, PubMed currently indexes more than 40 million biomedical citations, while ClinicalTrials.gov provides searchable and downloadable trial records and an API. For international socioeconomic work, World Development Indicators offers large cross-country time series with metadata and programmatic access. For U.S. environmental work, EPA provides an enterprise data catalog, Envirofacts, and EnviroAtlas; EnviroAtlas alone exposes hundreds of environmental and ecosystem-service data layers.

Analysis

Analysis should be selected according to the evidence rather than imposed in advance.

For qualitative or legal-policy topics, use a claim-evidence matrix, jurisdictional comparison, chronology, doctrinal or institutional analysis, stakeholder mapping, and explicit causal chains.

For observational quantitative topics, begin with descriptive distributions and trends, then use appropriate comparisons, regression, interrupted time-series, panel, quasi-experimental, or matching methods only where the underlying identification assumptions are defensible.

For experimental or clinical evidence, compare effect sizes, confidence intervals, attrition, protocol deviations, endpoint selection, study power, and risk of bias. Registered trials should be checked against published reports where feasible; ClinicalTrials.gov explicitly supports structured downloading of study information for evidence synthesis.

For economic analysis, distinguish nominal from real values, levels from rates, correlation from causal identification, and aggregate effects from distributional effects. International indicators should be accompanied by metadata because comparability can depend on definitions and country reporting practices. World Bank WDI provides both data and metadata for this purpose.

For technology assessments, evaluate performance under representative conditions rather than relying exclusively on benchmark averages. Where AI risk is involved, a useful organizing framework is governance, mapping, measurement, and management; NIST's AI Risk Management Framework is a voluntary, use-case-agnostic framework organized around those functions, although the framework itself should not be mistaken for empirical proof that any particular control works.

Every major conclusion should receive a simple confidence judgment:

The final analytical pass should be deliberately adversarial. For every important conclusion, ask: What is the strongest contrary evidence? What assumption is necessary for this conclusion? What observation would defeat it? Is there another causal explanation?

Source strategy

The source hierarchy should prioritize originality, authority, methodological transparency, and proximity to the underlying event or measurement.

For technology, begin with original papers, conference proceedings, official model or system documentation, relevant standards bodies, NIST, government evaluations, benchmark repositories, and documented incident evidence. NIST's AI RMF remains a useful governance reference, but its official site notes that AI RMF 1.0 is undergoing revision, making version control important for any current project.

For health, prioritize PubMed/MEDLINE, ClinicalTrials.gov, regulator documents, public-health agencies, original trial papers, systematic reviews, and recognized clinical guidelines. PubMed covers MEDLINE and other biomedical literature resources, while ClinicalTrials.gov offers machine-readable trial data suitable for structured evidence collection.

For environment, prioritize EPA and equivalent environmental agencies, meteorological and geological agencies, satellite/remote-sensing sources where appropriate, original studies, and IPCC or comparable assessments. EPA's current data resources include its Enterprise Data Catalog, Envirofacts, and EnviroAtlas, with downloadable and geospatial data.

For economics, prioritize national statistical agencies, central banks, World Bank, IMF, OECD, official administrative datasets, and original empirical economics papers. World Development Indicators currently supports downloadable data, visualization, metadata, and API access across a broad set of development measures.

For law and policy, retrieve the operative instrument itself first: current statute or regulation, amendments, commencement provisions, official legislative history, judicial opinions, administrative guidance, and enforcement records. The uploaded brief illustrates a particularly rigorous version of this method: it requires each instrument to be tracked by jurisdiction, version, provision, affected actor, territorial trigger, exceptions, enforcement mechanism, sanction, and review route, while distinguishing proposed, enacted, applicable, enjoined, repealed, and expired measures. That approach is worth preserving for any legal-policy project.

Searches should normally be conducted in English first, while jurisdiction-specific work should also search authoritative materials in the relevant official language. Translations should be labeled as official or unofficial rather than silently treated as equivalent to the controlling text.

Timeline and milestones

A three-week standard schedule is the recommended default; the same workflow compresses into two weeks by overlapping stages or expands to four weeks for deeper data work and independent checking.

gantt
    title Standard Deep-Research Schedule
    dateFormat  YYYY-MM-DD
    axisFormat  %b %d

    section Framing
    Research charter and scope          :a1, 2026-09-07, 2d
    Search protocol and inclusion rules :a2, after a1, 1d

    section Evidence
    Primary-source collection           :b1, 2026-09-09, 5d
    Literature and counterevidence      :b2, 2026-09-10, 6d
    Structured extraction               :b3, 2026-09-11, 6d

    section Analysis
    Quantitative or comparative analysis:c1, 2026-09-15, 4d
    Adversarial and sensitivity review  :c2, after c1, 2d

    section Synthesis
    Draft report                        :d1, 2026-09-19, 3d
    Tables, charts, and slides           :d2, 2026-09-20, 3d
    Source audit and final revision      :d3, 2026-09-22, 2d

The milestones should be decision gates rather than merely dates:

A two-week version should reduce breadth rather than sacrifice verification: one geography, fewer outcomes, fewer comparisons, and only the most decision-relevant analyses. A four-week project is appropriate when original data cleaning, cross-country legal research, meta-analysis, or substantial sensitivity testing is required.

The underlying workflow is:

flowchart TD
    A[Define decision and scope] --> B[Translate questions into searchable concepts]
    B --> C[Retrieve primary evidence]
    B --> D[Retrieve scholarly synthesis]
    C --> E[Structured evidence register]
    D --> E
    E --> F[Quality and relevance screening]
    F --> G[Analysis]
    G --> H[Search for counterevidence]
    H --> I[Robustness and uncertainty assessment]
    I --> J[Draft conclusions]
    J --> K[Claim-to-source audit]
    K --> L[Report, slides, data tables and visuals]

Deliverables and evidence templates

The core deliverable should be a detailed analytical report organized around findings rather than around the order in which sources were discovered. A typical structure is executive findings, research question and scope, methodology, evidence base, principal findings, competing explanations, uncertainty and limitations, implications, options or recommendations, and technical appendices.

A companion slide deck of roughly 10–20 slides should tell the same story more selectively: question, why it matters, evidence base, three to six principal findings, one or two counterintuitive results, uncertainties, decision options, and next actions.

The data package should contain cleaned analytical tables, a codebook/data dictionary, provenance notes, calculations underlying headline numbers, and reusable chart data. Where licensing or confidentiality prevents redistribution, the package should contain locators and transformation instructions rather than improperly copying data.

Source-comparison template

Key-findings template

Budget template

The figures are planning placeholders, not market-rate quotations; actual effort depends heavily on the chosen topic and evidence burden.

Recommended visuals

A strong project should favor analytical figures over decorative graphics. Useful candidates include:

Every quantitative chart should state the unit, denominator, date range, data source, and relevant exclusions. Whenever uncertainty is measurable, the figure should display it rather than presenting a single point estimate as certain.

The slide deck should generally contain fewer, stronger visuals than the long report. The detailed report can hold methodological charts, secondary analyses, source matrices, and robustness figures in appendices.

Assumptions and narrowing decisions

This plan assumes that the topic has not yet been chosen, that the final output is intended for an analytically sophisticated English-language audience, and that public or lawfully accessible evidence will be sufficient. It also assumes a research task rather than original laboratory experimentation, human-subject recruitment, or confidential corporate due diligence.

There is one notable ambiguity: the uploaded material is not actually topic-neutral. It describes a very specific prospective project on laws governing automated expression, bot disclosures, synthetic-media labeling, political communication, and machine association, including California, EU, U.S. election-law, and foreign comparators. I have treated that document as a possible concrete instantiation rather than silently replacing your explicit request for a reusable plan. Were that subject selected, the generic workflow above would be tightened substantially around official statutory text, legislative status, judicial precedent, territorial applicability, compelled-speech analysis, structured legal registers, and worked scenarios.

The most effective narrowing decisions, in descending order of value, are:

Decision use. Identify the actual choice the research is supposed to inform. “Understand AI” is much weaker than “decide whether a specified disclosure mandate should be adopted.”

Primary outcome. Select one or two outcomes that determine success.

Unit of analysis. Specify people, firms, technologies, jurisdictions, studies, policies, ecosystems, or another clearly defined unit.

Geography. Prefer one jurisdiction or a deliberately selected comparison set over an undefined global review.

Time window. Define both the evidence cutoff and the historical period relevant to the question.

Comparator. Specify the counterfactual: no intervention, prior policy, alternative technology, peer jurisdiction, standard care, or another baseline.

Evidence standard. Decide whether the project is principally descriptive, causal, legal/doctrinal, forecasting-oriented, or decision-analytic. Different standards of proof should not be blended without explanation.

Follow-up checklist

Before substantive research begins, the user should answer only the questions necessary to freeze scope:

• Topic: What precise phenomenon, technology, condition, market, or policy should be investigated?
• Decision: What decision should the research help make?
• Audience: Who will read or act on the findings?
• Geography: Which countries, states, markets, or populations matter?
• Time horizon: Current-state assessment, historical analysis, or future scenarios?
• Comparison: What alternative or baseline should be evaluated?
• Depth: Focused, standard, or extended scope?
• Outputs: Report only, or also slides, reusable data, calculations, and publication-quality graphics?
• Perspective: Neutral synthesis, decision recommendation, adversarial stress test, or a specified normative framework?
• Constraints: Any sources, jurisdictions, methods, claims, or confidential materials that must be included or excluded?