A General-Purpose Deep Research Framework

Report body

A General-Purpose Deep Research Framework

Executive summary

A rigorous deep-research project should begin before searching: first convert the vague topic into a decision-oriented question, define scope and exclusion rules, identify what kinds of evidence could change the answer, and establish a reproducible search and evidence log. For systematic evidence work, this emphasis on explicit eligibility criteria, documented searches, screening, and transparent reporting is consistent with PRISMA 2020 and the Cochrane Handbook. citeturn0search9turn0search17

For the present request, the following parameters are deliberately unspecified:

An uploaded example brief in this conversation demonstrates the opposite extreme: it specifies a normative perspective, authorities to investigate, evidence standards, scenarios, output files, and legal-status controls. That is a useful model of how an initially broad inquiry can eventually be operationalized, although this report treats the research topic itself as unspecified, as requested here. fileciteturn0file0

The recommended default workflow is:

flowchart LR
    A[Unspecified topic] --> B[Decision and audience]
    B --> C[Research questions]
    C --> D[Evidence protocol]
    D --> E[Search and collection]
    E --> F[Screen and validate]
    F --> G[Analyze]
    G --> H[Synthesize claims]
    H --> I[Adversarial review]
    I --> J[Report and decision]
    J --> K[Update / monitor]

The central principle is triangulation rather than accumulation. A hundred loosely relevant webpages are usually weaker than a smaller set combining primary law or regulation, original research, official statistics, corporate filings, patents, technical standards, and independent corroboration. Official data systems such as SEC EDGAR, World Bank APIs, USPTO Patent Public Search, ClinicalTrials.gov, and PubMed provide structured or directly attributable evidence that can often anchor a project before secondary commentary is introduced. citeturn2search0turn2search1turn1search0turn4search13turn3search0

A sound research program should also maintain three separations throughout:

Evidence vs. interpretation. What does the source directly establish, and what is the researcher inferring?

Description vs. causation. An association, trend, or predictive model does not by itself establish what caused an outcome.

Retrieval vs. permission. The technical ability to collect information does not settle privacy, copyright, contractual, research-ethics, or computer-access questions. RFC 9309, for example, standardizes robots.txt behavior for crawlers but explicitly says that the protocol is not a substitute for access-control security; human-subject research may separately trigger institutional and informed-consent requirements. citeturn2search4turn1search6

Research question framework

Turning a topic into an answerable question

A practical formulation is:

For [decision-maker], should [action] be taken regarding [object] in [population/market/jurisdiction] over [time horizon], compared with [alternative], given [outcomes/risks] and subject to [constraints]?

That produces a research question containing six elements:

A topic such as “AI in healthcare” is not yet a research question. It might become:

“For U.S. hospital systems considering ambient clinical-documentation AI through 2029, what evidence exists that deployment changes clinician administrative time, documentation quality, patient safety, and total operating cost relative to conventional documentation?”

That refinement immediately determines relevant sources, variables, inclusion criteria, and analysis methods.

Build a question tree

The primary question should be decomposed into mutually reinforcing subquestions:

Primary decision question
├── Baseline: What is happening now?
├── Mechanism: Why might the proposed effect occur?
├── Magnitude: How large is the effect?
├── Distribution: Who gains or loses?
├── Alternatives: Compared with what?
├── Constraints: Legal, technical, financial, ethical?
├── Failure modes: Under what conditions does the thesis break?
└── Future: What could materially change the answer?

This structure prevents the common failure of collecting a large amount of interesting information that never resolves the actual decision.

Pre-specify claim tests

Before research begins, create a claim ledger:

For each important claim, record at least one piece of evidence for, one plausible challenge against, and the conditions under which the conclusion changes. This turns research into falsification rather than confirmation.

Define scope before retrieval

A lightweight research protocol should specify:

Included: jurisdictions, years, languages, populations, study types, technologies, outcomes.

Excluded: irrelevant adjacent technologies, duplicate publications, unsupported commentary, obsolete legal versions, non-comparable markets.

Evidence cutoff: the latest date through which sources have been checked.

Search log: database, exact query, filters, date searched, hit count, screening decision.

Confidence vocabulary: for example, high / moderate / low confidence, with explicit criteria.

PRISMA 2020 provides a 27-item reporting framework and flow diagrams for documenting records identified, screened, excluded, and included; while developed principally for systematic reviews, its transparency principles transfer well to rigorous non-medical evidence synthesis. citeturn0search9turn0search10

Source architecture and search strategy

Evidence hierarchy

There is no universal ranking applicable to every question, but a useful default is to work outward from authoritative primary evidence.

For U.S. public companies, SEC EDGAR offers public filings and APIs, including submissions histories and XBRL financial data. USPTO Patent Public Search provides public access to issued U.S. patents and published applications. The World Bank Indicators API exposes a large collection of internationally comparable time series programmatically. citeturn2search0turn1search0turn2search1

For biomedical and life-science literature, PubMed supports Boolean operators, field tags, controlled vocabulary via MeSH, date restrictions, query history, and detailed inspection of how searches were translated. citeturn3search0

Recommended source stack by research type

Academic questions: start with field databases such as PubMed and domain-specific indexes, then use Crossref/OpenAlex for broad bibliographic discovery and citation-network expansion. Crossref's REST API exposes works, journals, funders and related metadata, while OpenAlex provides programmatic scholarly-graph querying. citeturn4search12turn4search2

Policy and legal questions: start with the legislature, official code, regulator, treaty repository or court itself; use law reviews and NGO analysis to identify arguments and obscure authorities, but return to primary text for the operative proposition.

Company and market questions: combine regulatory filings, audited financial statements, official statistics, customer/procurement records where public, industry datasets, patents and independent reporting. EDGAR's full-text search is particularly useful for longitudinal comparisons of risk factors, strategy terminology, customer concentration and disclosed metrics. citeturn2search0turn2search3

Technology questions: prioritize specifications, standards, repositories maintained by the creators, benchmark papers, patents, reproducible evaluations and deployment evidence. Marketing pages can establish what a supplier claims but not necessarily independent performance.

Patent landscapes: search keywords and classification codes, then examine families, priority dates, assignees, inventors, backward/forward citations and prosecution status. USPTO also provides related systems for assignments, file wrappers and international dossiers. citeturn1search7

Search in layers

A robust search generally proceeds through four passes.

Orientation pass. Learn vocabulary, entities, acronyms, landmark papers, agencies and competing hypotheses.

Precision pass. Query exact entities, titles, controlled vocabulary, statutory provisions, patents, identifiers and datasets.

Recall pass. Add synonyms, spelling variants, former names, broader terms and citation chaining.

Adversarial pass. Deliberately search for failures, null findings, retractions, criticism, litigation, contradictory data, competing mechanisms and evidence that would falsify the emerging answer.

A concept matrix helps construct systematic queries:

For systematic reviews, Cochrane recommends comprehensive searches, documentation of search and selection processes, appropriate source selection, and—in relevant health contexts—collaboration with information specialists. citeturn0search17turn0search20

The research log should preserve the exact query, not merely “searched PubMed.” That matters because PubMed can map terms through Automatic Term Mapping and exposes the translated query in Search Details. citeturn3search0

Data collection, tools, and governance

Collection methods

Use the least complex acquisition method capable of generating reproducible data.

Scrapy is designed for structured web crawling and extraction and provides components such as spiders, selectors and item pipelines. Playwright automates real browser engines and is useful when legitimate collection requires JavaScript-rendered pages rather than static HTTP retrieval. REDCap is a web-based system designed for research surveys and databases and supports features including authentication and data logging. citeturn6search0turn6search1turn6search2

Prefer APIs to scraping whenever an authoritative, sufficiently complete API exists. For example, the World Bank provides an Indicators API for programmatic data retrieval; OECD likewise provides API access to Data Explorer data based on SDMX. citeturn2search1turn2search9

Reproducible data pipeline

A minimum defensible pipeline is:

raw/
    immutable source snapshots or API responses
intermediate/
    parsed and normalized records
analysis/
    analysis-ready tables
outputs/
    charts, tables, models
metadata/
    data dictionary, provenance, search log

For every observation or document, preserve where feasible:

source_id, source_url_or_identifier, retrieved_at, published_at, version, issuer, raw_field, normalized_field, transformation, quality_flag, and notes.

Never silently overwrite raw evidence after cleaning. Transformations should be reproducible and documented.

Ethical and legal controls

The legal status of automated collection is fact- and jurisdiction-dependent. A research protocol should separately inspect:

1. whether information is legitimately accessible;
2. applicable website/API terms;
3. authentication or technical access controls;
4. copyright/licensing;
5. privacy and data-protection obligations;
6. confidentiality and trade-secret issues;
7. rules governing human-subject research;
8. contractual or institutional policies.

RFC 9309 requires conforming crawlers to follow applicable Allow and Disallow rules after successfully retrieving robots.txt; it also expressly states that the protocol itself is not a security mechanism. Treat it as one technical signal, not a universal legal permission system. citeturn2search4

In the United States, DOJ's CFAA charging policy distinguishes mere contractual restrictions from particular forms of unauthorized computer access for federal charging purposes, but the policy itself does not create enforceable private rights and should not be mistaken for a complete statement of every legal theory governing scraping. citeturn7search0

Copyright must likewise be assessed separately. U.S. Copyright Office guidance notes that research is among the categories of activity that may qualify as fair use, but fair use remains a case-specific four-factor inquiry rather than an automatic “research exception.” citeturn7search13

For research involving people, identify applicable institutional and statutory requirements before recruitment or data collection. HHS describes the Common Rule as establishing provisions for IRBs, informed consent and institutional assurances for covered federally conducted or supported human-subject research; its applicability depends on the specific project and sponsoring agency. citeturn1search6

A good practical rule is:

Publicly visible, technically retrievable, ethically appropriate, and legally permissible are four separate questions.

Data quality gates

Before analysis, run checks for:

• duplicated records;
• missingness by variable and subgroup;
• inconsistent units;
• invalid dates or impossible values;
• entity-resolution errors;
• changed definitions over time;
• survivorship or coverage bias;
• duplicated syndication of the same underlying source;
• incorrect joins;
• unexpected observation counts.

For high-consequence research, a second researcher should independently reproduce at least the most load-bearing extractions.

Analysis methods and when to use them

The analysis method follows from the claim, not from which software happens to be available.

Qualitative analysis

Use qualitative methods when the question concerns how, why, meaning, process, interpretation, institutional behavior or failure mechanisms and structured numerical data do not capture the construct.

A rigorous thematic workflow is:

familiarization → coding → candidate themes → review → definition → synthesis → negative-case search.

Braun and Clarke characterize thematic analysis as a flexible approach to identifying and analyzing patterned meaning in qualitative material; methodological literature emphasizes systematic engagement with the entire dataset rather than selecting only convenient quotations. citeturn5search3turn5search1

For codebook-style projects, document:

• code definition;
• inclusion criteria;
• exclusion criteria;
• positive example;
• borderline example;
• relationship to parent/child codes.

Where intercoder reliability is conceptually appropriate, report agreement statistics and adjudication procedures. Where using reflexive thematic analysis, do not mechanically impose reliability concepts that conflict with the chosen methodological stance; instead document researcher interpretation and reflexivity.

Statistical inference

Use statistical tests when you have a defined sampling or experimental process and the question concerns uncertainty around numerical comparisons.

Prefer reporting:

estimate + effect size + uncertainty interval + sample denominator + assumptions

rather than a binary “significant/not significant” result.

Typical choices include:

The NIST/SEMATECH Engineering Statistics Handbook is a useful authoritative reference for experimental design, exploratory analysis, probability models and statistical methods in scientific and engineering settings. citeturn5search12

Causal inference

Use causal methods only when the question is genuinely counterfactual:

“What would have happened to the same target units without the treatment/exposure?”

A useful hierarchy is:

Randomization, when ethically and operationally feasible.

Natural/quasi-experiments, including regression discontinuity, instrumental variables or policy discontinuities, when assignment creates defensible quasi-random variation.

Difference-in-differences, when treatment timing and untreated comparison units permit a credible parallel-trends design.

Matching, weighting or covariate adjustment, when conditional exchangeability is defensible and important confounders are measured.

Synthetic control, often useful for aggregate interventions with limited treated units and strong pre-treatment data.

The assumptions are part of the result. Modern difference-in-differences work, for example, emphasizes that identification depends on appropriate parallel-trends assumptions and that alternative assumptions may imply different estimands or robustness-efficiency tradeoffs. citeturn5search0turn5search10

A causal report should therefore explicitly state:

estimand → treatment → outcome → comparison → identification assumption → diagnostics → sensitivity analysis.

Machine learning

Use ML primarily when the objective is prediction, classification, ranking, clustering, representation learning, or extracting structure from high-dimensional data.

A proper evaluation separates:

training → validation/model selection → locked test evaluation.

For classification, potentially useful metrics include:

• sensitivity/recall;
• specificity;
• precision;
• F1;
• AUROC;
• AUPRC;
• calibration;
• subgroup performance.

For regression:

• MAE;
• RMSE;
• R² where meaningful;
• calibration/residual diagnostics;
• subgroup errors.

For operational models, performance metrics should be accompanied by data-provenance, robustness, privacy, interpretability and risk controls appropriate to the use case. NIST's AI RMF describes a lifecycle-oriented approach organized around Govern, Map, Measure and Manage and treats risk as extending beyond simple predictive accuracy. citeturn0search0turn0search5

Do not convert feature importance, attention scores, model correlations or prediction accuracy into causal explanations without a separate identification argument.

Mixed-method designs

The strongest answer frequently combines methods:

quantitative pattern → qualitative mechanism → primary-document verification → counterevidence → integrated conclusion.

For example, an adoption dataset could identify where a technology grew rapidly; interviews could explain procurement dynamics; filings could verify vendor economics; causal analysis could investigate effects; patents could reveal technical trajectories. Each method answers a different question rather than serving as redundant confirmation.

Synthesis, reporting, and example deliverables

Evidence synthesis

The final unit of research should be a claim, not a document.

Build an evidence matrix such as:

Then write the report from the matrix. This prevents the common “source-by-source book report” structure.

For systematic-review-style work, PRISMA's flow diagram explicitly tracks records identified, screened, excluded and included, helping a reader distinguish an evidence corpus from an arbitrary collection of citations. citeturn0search10

Suggested report template

A high-quality final report can follow:

Executive summary
Decision, answer, three to five key findings, material uncertainty, recommendation.

Research question and scope
Population, geography, period, definitions, exclusions.

Method
Searches, sources, data, screening, assumptions, cutoff date.

Current state / baseline
What is directly observable.

Core findings
Organized by question, not by source.

Counterevidence and uncertainty
Strongest rival interpretation, missing evidence, contradictory findings.

Scenario or sensitivity analysis
What changes the answer.

Implications and recommendation
Who should do what and under what conditions.

Appendices
Queries, detailed methods, data dictionary, models, additional tables, bibliography.

Visualization guide

Choose graphics according to the inference:

Do not use charts merely to decorate a report. Every visualization should answer an explicit question, show denominators and units where relevant, and make uncertainty visible where it materially affects interpretation.

Research-quality dashboard

A compact internal dashboard can track:

Corpus metrics: records found, deduplicated, screened, included.

Authority mix: share of central claims supported by primary evidence.

Freshness: age of evidence relative to the research cutoff.

Triangulation: number of independent source classes supporting each load-bearing claim.

Contradiction rate: material claims for which credible contrary evidence exists.

Data quality: missingness, duplicates, unresolved entities, anomalous observations.

Statistical quality: effect sizes, uncertainty intervals, sample sizes, robustness tests.

Predictive quality: test-set and subgroup metrics plus calibration where relevant.

Traceability: percentage of important claims that can be traced to exact source locations.

These should be treated as diagnostic metrics, not as a mechanical “research score.”

Example executive summary

Illustrative only; no topic has been supplied.

Evidence from primary records, original research and independent industry data supports the conclusion that [intervention X] probably improves [outcome Y] under [conditions A and B], but evidence is insufficient to support universal adoption. The strongest studies suggest [direction/magnitude], while implementation data indicate that benefits depend heavily on [constraint]. The principal uncertainty is [uncertainty], and the strongest counterargument is [counterargument].

Recommendation: proceed with [bounded action], contingent on [metric/threshold], while postponing [irreversible action] until [missing evidence] is obtained.

Confidence: moderate, because the direction of evidence is consistent across several source classes but causal evidence remains limited.

Example slide-deck outline

Example technical appendix

The technical appendix should normally contain:

research protocol → complete query log → inclusion/exclusion rules → source register → data provenance → cleaning rules → variable derivations → statistical specification → model parameters → sensitivity analyses → robustness checks → missing-data treatment → limitations → reproducibility notes.

This is where complexity belongs so the main report can remain decision-oriented.

Example data dictionary

Timelines and resource estimates

The following are illustrative planning envelopes, not market prices or universal benchmarks. They assume English-language research, reasonable source availability, no unusually slow records requests, and no regulated clinical intervention.

The budget ranges above are intentionally broad and exclude ordinary staff compensation. Paid databases, participant incentives, travel, transcription, specialist legal review, licensed datasets or large survey samples can dominate costs.

A small project should usually avoid bespoke data collection unless it is indispensable. A medium project can justify structured extraction, limited interviews, or secondary-data analysis. A large project becomes appropriate when the decision is consequential enough to require original data, multi-person validation, causal inference, longitudinal monitoring or multiple jurisdictions.

Example medium-project timeline

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

    section Framing
    Decision and question design       :a1, 2026-09-07, 4d
    Protocol and evidence model        :a2, after a1, 3d

    section Collection
    Primary-source retrieval           :b1, after a2, 10d
    Academic and industry search       :b2, after a2, 10d
    Structured data collection         :b3, after a2, 14d

    section Analysis
    Screening and extraction           :c1, after b1, 8d
    Quantitative / qualitative analysis:c2, after b3, 10d
    Counterevidence review             :c3, after c1, 5d

    section Synthesis
    Claim-evidence synthesis           :d1, after c2, 5d
    Draft report and visuals           :d2, after d1, 5d
    Adversarial review                 :d3, after d2, 3d
    Finalization                       :d4, after d3, 3d

Activities should overlap. Waiting until “all research is finished” before beginning synthesis is inefficient; maintain a live claim ledger and evidence matrix throughout.

A useful staffing model for a medium project is:

Research lead: question framing, synthesis, final judgment.

Domain researcher: specialist literature and primary authorities.

Data analyst: structured extraction, statistics, visualization.

Reviewer: independent challenge to methods and conclusions.

One person can hold several roles, but the final reviewer should ideally not be the sole author of the central analysis.

Annotated bibliography and actionable plan

Representative annotated bibliography

The sources below are intentionally cross-domain methodological examples, not a bibliography for any particular substantive topic.

These eight sources illustrate the deeper principle: methodology, literature discovery, official corporate records, structured statistics, patents, research ethics and technical collection rules are separate evidence domains. A real project should replace or supplement them with authoritative sources specific to its substantive question.

Actionable research plan

The most important rule in Step 6 is to match the method to the question: qualitative coding for meaning and process, conventional statistical inference for sampled numerical differences, causal inference for counterfactual effects, and ML for prediction or pattern recognition. Difference-in-differences research, for example, is credible only to the extent that its identification assumptions such as appropriate parallel trends can be defended; predictive model quality similarly requires evaluation beyond training fit. citeturn5search0turn0search0

The most important rule in Step 7 is to distinguish what is known, what is inferred, what remains disputed, and what would change the recommendation. A good deep-research product should leave a skeptical reader able to reconstruct not only the conclusion, but also the search universe, exclusions, source hierarchy, transformations, assumptions, contrary evidence and uncertainty that produced it. PRISMA and Cochrane embody this emphasis on transparent, reproducible evidence handling, while official APIs and repositories make the underlying evidence increasingly auditable at the record level. citeturn0search9turn0search17turn2search0turn2search1