How do you research with Super Intelligence? Start by turning a broad topic into a research question, deciding what kind of evidence could answer it, finding and evaluating sources, extracting relevant evidence, comparing disagreements and producing a conclusion whose strength does not exceed the evidence.
SI can accelerate research because it can help generate sub-questions, search terms, source maps, comparison tables and cited syntheses. Current research systems from major providers can also plan multi-step investigations across web sources and selected files. But research quality still depends on source authority, freshness, scope, methods and human verification.
This eduKateSG guide introduces Stage 3 of the How to Learn Super Intelligence Quickly curriculum. It shows how to build a repeatable SI research workflow before moving into web search, source reliability and long-article reading.
Terminology: SI is our editorial term for practical contemporary AI learning. Research with SI means using intelligent tools to assist the research process; it does not mean treating generated text as evidence by itself.
The First Principle: Research Is a Chain From Question to Evidence to Conclusion
Strong research has a visible chain. A question defines what must be known. Evidence supplies information relevant to that question. Analysis compares and interprets the evidence. A conclusion states what the evidence supports.
Weak SI research often skips one of those links. It begins with a topic and jumps directly to a fluent answer. The answer may contain useful information, but the reader cannot tell which claims came from which source or whether the sources actually answer the question.
The central research skill is therefore not asking SI for more facts. It is building a traceable path from question to source to evidence to conclusion.
Research Step 1 — Define the Research Question
A topic is not yet a research question. “Homework” is a topic. “What does recent evidence say about the relationship between homework amount and mathematics achievement for lower-secondary students?” is a research question.
The stronger question contains population, outcome and scope. It is easier to search, easier to evaluate and less likely to produce a universal answer to a context-dependent problem.
Before researching, ask what decision or understanding the question should support. That keeps the project from expanding indefinitely.
Research Step 2 — Define the Evidence Needed
Different claims require different evidence. A question about an organisation’s current eligibility rules is best answered by the organisation’s current official material. A question about treatment effects may require peer-reviewed studies or systematic reviews. A question about market price may require current market data.
SI can help identify candidate source types, but the user should decide what evidence is capable of supporting the claim.
This prevents a common error: using an authoritative source that is authoritative about the wrong thing.
Research Step 3 — Break the Question Into Sub-Questions
Complex research questions often contain several smaller questions. Separate definitions, mechanisms, current state, evidence, disagreement and implications.
Example: “How is generative AI changing secondary education?” can split into adoption, teaching use, student use, assessment, teacher workload, policy and evidence of learning effects.
Sub-questions improve search because each can use different source types and search terms.
Research Step 4 — Build a Search Plan
A search plan identifies likely sources, keywords, synonyms, dates and domains. It also decides when to use broad discovery and when to restrict to authoritative sources.
Current OpenAI guidance describes deep research as appropriate for multi-step questions that require combining and analysing many sources, while ordinary search is better for quick facts. Google’s current Gemini documentation similarly describes Deep Research as a multi-step research agent working across selected sources.
The tool choice should follow the research depth, not the prestige of the feature.
Research Step 5 — Gather Candidate Sources
At this stage, breadth matters. Collect enough candidate sources to understand the landscape. Do not treat search ranking as evidence quality.
Record source title, publisher, date, source type and why it may matter. A candidate list is a discovery artifact, not yet the final evidence set.
SI can help cluster sources by topic or viewpoint, which makes later selection easier.
Research Step 6 — Evaluate Source Authority
Authority depends on the claim. Official rules should come from the responsible organisation. Scientific findings should be traced to research methods and original papers when possible. Company claims about their own products can establish what the company says, but may not independently establish comparative superiority.
Ask who produced the source, for what purpose, using what method and with what access to the underlying information.
Article 23 in this series develops source evaluation in detail.
Research Step 7 — Check Freshness
Research can be wrong because the evidence is old rather than false. Software capabilities, laws, prices, policies and leadership change.
For time-sensitive questions, record publication or update date. Distinguish historical evidence from current-state evidence.
Do not use an older source as current merely because it ranks highly or is easier to read.
Research Step 8 — Extract Evidence, Not Just Summaries
A summary can hide the exact wording needed to support a claim. Extract the relevant passage, table, statistic or methodological detail.
A useful evidence note can contain Claim, Source, Locator, Evidence, Population, Date and Limitation.
This creates a research ledger that can be checked later without rereading every source from the beginning.
Research Step 9 — Separate Source Fact From Interpretation
Sources contain both observations and interpretations. A study may report a measured association and then discuss possible explanations. An organisation may state a rule and separately describe its rationale.
Your synthesis should preserve those roles. “The study found X” is different from “X causes Y” unless the methodology supports that causal conclusion.
SI can help label evidence status, but the classification must be checked against the source.
Research Step 10 — Compare Sources on the Same Question
Sources should be compared using common dimensions. If one source measures test scores and another measures student satisfaction, they do not directly answer the same outcome question.
Create a comparison table with Population, Time, Method, Outcome, Finding and Limitation.
This often reveals that apparent disagreement comes from different definitions or populations rather than contradictory evidence.
Research Step 11 — Preserve Genuine Disagreement
When credible sources disagree, do not force consensus. Identify what each source claims, how the methods differ and what evidence could resolve the disagreement.
A bounded conclusion may be: “Evidence is mixed under these conditions.” That can be more accurate than a confident synthesis that hides conflict.
SI is useful for mapping disagreements, but the user should not ask it to choose a winner without criteria.
Research Step 12 — Write a Bounded Conclusion
The conclusion should be no stronger than the evidence. If evidence comes from one population, say so. If results are correlational, do not present them as causal. If a source is historical, do not imply that it describes the current state.
A bounded conclusion can still be useful. Precision is stronger than exaggerated certainty.
Research Step 13 — Record What Remains Unknown
Every research project has edges. List unanswered questions, missing data and evidence that would change the conclusion.
Unknowns help the reader understand the operating envelope of the research. They also create a future research agenda.
Research is not weakened by explicit limits; it is made more trustworthy.
Research Step 14 — Convert Findings Into a Deliverable
The final deliverable may be a report, annotated bibliography, decision memo, lesson, presentation or source register.
Choose a format that preserves traceability. A decision memo may need concise conclusions plus source links. A literature review may need methods and disagreements. A student report may require direct citations and explanation.
Receiver needs determine the final structure.
The SI Research Stack
- Question.
- Scope.
- Evidence requirements.
- Search plan.
- Candidate sources.
- Source evaluation.
- Evidence extraction.
- Comparison.
- Contradiction handling.
- Synthesis.
- Verification.
- Bounded conclusion.
- Unknowns.
- Deliverable.
The stack turns research into a controllable process rather than a one-prompt answer.
Quick Search Versus Deep Research
Use quick search when the question is narrow, current and answerable from a small number of sources. Examples include a current official deadline or one product specification.
Use deeper research when the question requires multiple sources, competing evidence, source selection, synthesis or a structured report.
Current ChatGPT documentation describes search as appropriate for quick facts and deep research for more complex multi-step investigation. The distinction is useful as a workflow principle even when another product is used.
Research Plans: Why Planning Before Search Helps
A research plan prevents the first few search results from defining the entire project. It identifies sub-questions and source types before discovery begins.
The plan can change when evidence reveals a better question. Research is iterative, not rigid.
Current deep-research tools often expose a plan or allow the user to refine focus. Human review of the plan remains valuable because the research goal belongs to the user.
The Source Register
Create a source register for non-trivial research. Fields can include Source, Type, Publisher, Date, Authority, Sub-Question, Status and Notes.
Mark sources as Candidate, Accepted, Background, Superseded or Rejected. A rejected source can remain in the register with a short reason if the exclusion matters.
This prevents the final report from citing a source simply because it was discovered early.
The Evidence Ledger
The source register tracks documents. The evidence ledger tracks claims. A single source may support several claims; one claim may require several sources.
Fields can include Claim, Source, Evidence, Scope, Limitation and Verification Status.
The ledger creates traceability from final prose back to evidence.
Research Notes Should Preserve Locators
Save page number, heading, table number, paragraph or another useful locator when possible. A source URL alone may not be enough for a long document.
Locators reduce later review time and make collaborative checking easier.
For changing webpages, record access date when freshness matters.
Source Hierarchy
A source hierarchy ranks source types according to the research question. Official source may control current policy. Original study may control methodology. Independent review may be stronger for synthesis across studies.
The hierarchy should be explicit enough that SI does not treat every webpage equally.
Different sub-questions may use different hierarchies.
Primary and Secondary Sources
Primary sources provide direct evidence or first-party records: original studies, legislation, official announcements, datasets, transcripts or company filings.
Secondary sources analyse, summarise or interpret primary material. They can be valuable for context, but should not silently replace primary evidence when the original is accessible and important.
Source type alone does not guarantee quality. Evaluate methods and relevance.
Researching a Current Product or Company
Use official documentation for features, supported behaviour and policies. Use independent testing or reviews for comparative performance where relevant.
Date every capability claim. Products change quickly.
Avoid turning a vendor’s own marketing claim into an independent comparative conclusion.
Researching Science
Define outcome, population and study design. Prefer original research and high-quality reviews when the question requires scientific evidence.
Read methods and limitations, not only abstracts or press coverage.
SI can accelerate extraction, but scientific interpretation still requires understanding what was actually measured.
Researching Education
Education evidence varies by age, subject, setting, intervention and outcome. Avoid universal conclusions based on one population.
Separate attainment, motivation, retention and other outcomes. A study showing improved engagement does not automatically show improved examination performance.
For Singapore-specific questions, distinguish international evidence from local policy or population evidence.
Researching Business and Markets
Separate company-reported figures, regulatory filings, market data, analyst interpretation and community sentiment.
Financial periods and definitions matter. Revenue, bookings and cash flow are not interchangeable.
Current market questions require current data rather than historical summaries.
Researching Law and Policy
Use the current authoritative legal or government source for the rule. Commentary can explain application, but should be attributed.
Jurisdiction and effective date are essential. Similar rules in another country or older period are not direct evidence of the current local rule.
For consequential legal questions, qualified professional advice may be required.
Researching History
Historical research benefits from primary records and scholarly interpretation. Source date is historical context rather than freshness in the same sense as current policy.
Ask what perspective the source represents, what records survive and how historians disagree.
Avoid judging historical evidence through one modern summary alone.
Researching News and Fast-Moving Events
Publication time and event time both matter. An older article may remain online after circumstances change.
Use recent authoritative sources and distinguish confirmed events from live claims or speculation.
When facts are still emerging, state uncertainty rather than forcing a stable narrative too early.
Researching Online Community Experience
Community sources can reveal lived experience, failure modes and practical tips that official sources may not contain.
They are weaker for establishing general factual claims. Treat anecdotes as experience evidence and look for patterns across multiple contributors.
Do not convert community consensus into measured population data.
Researching With Uploaded Files
When the research corpus is supplied by the user, first inventory the files. Record title, date, version and role.
Search within the corpus for the research question. Preserve locators and identify conflicts between files.
Do not assume a filename proves the system actually read the file. Verify access and source references.
Researching With Connected Apps
Current research systems may be able to use selected connected data sources subject to account and workspace permissions. Treat those connections as source access, not permission to rewrite or alter the source.
For research, read-only access is usually sufficient. Respect provider and organisational permissions.
Keep connected internal sources distinguishable from public web evidence in the final report.
Research Prompt Design
A strong research prompt states question, purpose, scope, time period, source priorities, required comparisons and output format.
Example: “Research current approaches to AI literacy for secondary students. Prioritise official education frameworks and recent peer-reviewed evidence. Separate policy recommendations from measured outcomes. Focus on 2024–2026 where current evidence is needed. Produce a source table plus bounded synthesis.”
The prompt guides the investigation without predetermining the conclusion.
Research Prompt Anti-Pattern: “Find Evidence That I Am Right”
A prompt designed to support a preferred conclusion encourages confirmation bias.
Use neutral framing: “Find evidence that supports and weakens this hypothesis. Identify methodological differences.”
If the project is advocacy, evidence standards still matter; persuasion should not be confused with research.
Research Prompt Anti-Pattern: Undefined “Best”
“What is the best teaching method?” hides population, subject, outcome and constraints.
Define the criteria before searching. Best for retention may differ from best for speed or motivation.
A well-defined question prevents fake universal rankings.
Research Prompt Anti-Pattern: Unlimited Scope
“Research AI” cannot close. Define the decision, topic boundary and time horizon.
Unlimited scope creates long reports with weak prioritisation.
A smaller question can be researched more deeply and honestly.
Research Prompt Anti-Pattern: Citation Theater
Requesting citations does not guarantee source quality. A source can be real but irrelevant. A link can open but fail to support the attached claim.
Verify the key citations manually or through source extraction.
Citation quality matters more than citation count.
A Worked Research Example: School AI Policy
Question: “What should a school consider when designing student AI-use guidance?”
Sub-questions: learning integrity, privacy, age/account rules, teacher practice, assessment, verification and permitted use.
Sources: official education guidance, provider age/account documentation, relevant research and local school requirements. Synthesis should distinguish rules from evidence and recommendations.
Conclusion should be a framework, not a claim that one universal policy fits every school.
A Worked Research Example: Does a Study Support the Headline?
Headline says an intervention “doubles learning”. Research task: inspect the original study.
Extract population, sample size, outcome, comparison, duration and effect measure. Compare headline wording with study result.
If “double” refers to a relative metric rather than absolute learning gain, explain that distinction.
The research answer becomes a claim audit, not a repetition of the headline.
A Worked Research Example: Compare Two AI Products
Define criteria before searching: supported file types, current web access, citations, tool integrations and price as of a specific date.
Use official product documentation for capabilities and pricing. Use independent tests for quality comparisons if relevant.
Keep missing or plan-dependent features visible rather than forcing a universal winner.
A Worked Research Example: Local Education Question
Question: “How does a current Singapore secondary-school policy work?”
Use Singapore official sources for the current policy. Historical or overseas sources may provide background but should not be presented as current local authority.
Record effective dates and terminology. Local specificity is part of research validity.
A Worked Research Example: Personal Purchase
Question: “Which laptop suits this workflow?” Research begins with requirements rather than products: software, battery, weight, budget, ports and display.
Current product specifications and prices require current sources. Community reviews can add real-world experience.
The final answer should explain trade-offs instead of ranking products by popularity alone.
Research Failure Mode 1 — Search Result Bias
The first results define the research narrative. Repair with planned source categories and alternative queries.
Failure Mode 2 — Authority Confusion
A respected source is treated as authoritative outside its domain. Repair by matching source authority to claim.
Failure Mode 3 — Freshness Failure
Old evidence is presented as current. Repair by date-scoping and current-source checks.
Failure Mode 4 — Population Drift
Evidence from one group is generalised to another. Repair by recording population and scope.
Failure Mode 5 — Outcome Drift
One measured outcome is presented as another. Repair by naming the exact outcome.
Failure Mode 6 — Method Blindness
The research cites results without understanding how they were produced. Repair by extracting method and limitation.
Failure Mode 7 — Consensus Laundering
Several secondary articles repeat one original claim, creating the appearance of many independent sources. Repair by tracing citation chains to primary evidence.
Failure Mode 8 — Source Quantity Over Quality
A report cites many weak sources. Repair by prioritising evidence quality and relevance.
Failure Mode 9 — Unsupported Synthesis
The final conclusion combines sources into a claim none of them supports. Repair with an evidence ledger linking each claim.
Failure Mode 10 — Unknowns Disappear
The report fills gaps with fluent speculation. Repair by maintaining an unresolved-question section.
A Research Quality Checklist
- Question is specific enough to close.
- Population and time period are defined when relevant.
- Source types match the claims.
- Current claims use current evidence.
- Important primary sources are checked where appropriate.
- Methods and limitations are understood.
- Evidence is extracted with locators.
- Disagreements remain visible.
- Claims do not exceed evidence.
- Unknowns are recorded.
- Key citations are verified.
- Final deliverable serves a real receiver.
A Practice Lab: Research One Claim
Choose one factual claim from an article or conversation. Write the exact statement and what source type could support it.
Find the strongest accessible source. Extract the relevant evidence and limitation. Rewrite the claim so its scope matches the source.
This small exercise trains the full research chain without requiring a long report.
A Practice Lab: Build a Source Register
Research one narrow question and collect eight candidate sources. Classify each by type, authority, date and relevance.
Accept only the sources that genuinely help answer the question. Record why others were rejected or background-only.
The exercise teaches selection rather than accumulation.
A Practice Lab: Compare Two Sources
Choose two sources that appear to disagree. Compare population, time, definitions, methods and outcome.
Decide whether the disagreement is real, methodological or merely apparent.
Write a synthesis that preserves whatever disagreement remains.
A Practice Lab: Research With a Stop Condition
Define a small research task and write the closure criteria before searching: two authoritative current sources, one independent synthesis and a list of unresolved questions.
Stop when the criteria are met. More searching is not automatically better.
This trains research efficiency and prevents endless browsing.
Research Maintenance
Long-lived reports need refresh cycles. Mark current claims that are likely to change and record the source date.
When refreshing, recheck only the parts whose evidence may have changed rather than rewriting stable conceptual sections unnecessarily.
A maintained source register makes updates much faster.
Frequently Asked Questions
Can SI do research for me automatically?
Current systems can automate substantial search and synthesis, but users should still define the question, review source choices and verify important claims.
How many sources do I need?
There is no universal number. Use enough high-quality sources to answer the question, represent relevant disagreement and support the conclusion.
Are citations enough to trust the answer?
No. Check whether the cited source supports the claim and whether it is appropriate for the population, date and question.
Should I always use primary sources?
Use them when they are important and accessible, especially for official rules, original research and first-party records. Secondary synthesis can add valuable context.
How do I research a topic I know nothing about?
Begin with orientation and vocabulary, then refine the question. Use reliable overview sources before moving into specialised evidence.
When should I stop researching?
When the defined closure criteria are met or when the next progress depends on evidence that is unavailable.
What comes next?
Continue with How to Search the Web With Super Intelligence, then How to Find Reliable Sources With Super Intelligence.
Triangulation: Do Independent Sources Point in the Same Direction?
Triangulation means checking a claim through more than one genuinely independent evidence path. Two articles that repeat the same press release are not two independent sources. They are one source echoed twice.
A stronger research design might combine an official record, an independent dataset and a scholarly analysis. Agreement across different source types can increase confidence when they measure the same underlying phenomenon.
Do not count source quantity without checking dependence. Citation chains often create the appearance of consensus when many pages trace back to one original claim.
The Claim Graph
A claim graph maps which evidence supports which conclusion. Start with the final claim, then list the subclaims required for it to be true. Attach evidence to each subclaim.
Example: “This programme expanded nationally in 2026.” Subclaims might include programme existence, geographic scope and effective date. Each may require a different source.
Claim graphs are useful because they expose unsupported links. An elegant final sentence can depend on one weak hidden assumption.
Evidence Grading
Not all evidence in a report deserves equal weight. Grade evidence qualitatively according to directness, authority, method, freshness and relevance.
A direct current official record may strongly establish a policy date. A small anecdotal survey may weakly suggest user sentiment. A random social post may reveal one person’s experience but cannot establish population prevalence.
Evidence grading helps synthesis because weak evidence is not allowed to carry the same argumentative weight as stronger evidence.
Direct Evidence Versus Proxy Evidence
Sometimes the ideal measurement is unavailable and researchers use a proxy. Website traffic may be used as a proxy for interest; self-reported study hours may be used as a proxy for effort.
Proxies can be useful, but name them. Do not write as if the proxy is identical to the underlying concept.
SI can help identify proxy relationships, but the user should decide whether the proxy is valid enough for the research question.
Correlation, Causation and Mechanism
A correlation shows that two variables move together under the observed data. Causation requires stronger evidence that changing one produces change in the other.
A plausible mechanism can support interpretation, but mechanism alone does not prove causal effect. Keep observational association, experimental evidence and theory distinct.
When SI summarises research, ask it to label which claims are correlational and which are supported by causal designs.
Sample Size Is Not the Whole Method
Large samples can still be biased. Small samples can still be informative for some qualitative questions. Research evaluation should examine selection, measurement, controls, missing data and analytic method.
Do not reduce study quality to one number. SI can extract sample size quickly, but human evaluation should consider how the sample was constructed and what population it represents.
The right methodological question depends on the claim being made.
Population Validity
Always ask who was studied. Evidence from university students may not generalise to primary-school children. Evidence from one country may not directly describe Singapore policy or behaviour.
Record age, geography, setting and inclusion criteria when those dimensions affect interpretation.
Population mismatch is one of the most common ways research becomes overgeneralised in secondary summaries.
Time Validity
Some evidence ages slowly; some ages quickly. A mathematical theorem remains stable. A software feature or pricing plan may change in weeks.
Match evidence age to topic volatility. For current product, policy and market research, old sources may be background only.
Research reports should distinguish historical trends from current-state claims rather than mixing them under one tense.
Method Extraction
When a source matters, extract how the result was produced. For a survey: sampling, questions and response rate. For an experiment: groups, intervention, outcome and duration. For a dataset: source, coverage and definitions.
This prevents conclusions from floating free of methodology.
A method note can be short, but it should be sufficient to understand the evidence’s main limitations.
Citation Verification
For every decision-critical claim, open the cited source and locate the passage, table or data that supports it. Do not rely only on the generated citation title.
Check three things: the source exists, the source says what the answer claims and the scope matches the claim.
This process catches fabricated citations, misread sources and citations that support only part of a sentence.
Citation Granularity
A single citation at the end of a long paragraph may not reveal which sentence it supports. Where precision matters, attach evidence closer to the relevant claim.
If one sentence contains two distinct claims, consider whether they need different sources.
Granular citation improves both reader trust and later maintenance.
Citation Chasing
Secondary sources often cite original research. Follow the citation chain backward when the original evidence is important.
Then move forward when useful: find later reviews, replications or updates. This shows whether the original claim remained stable, changed or was challenged.
SI can accelerate citation discovery, but bibliographic identity and source content still need checking.
Research Diversity Versus False Balance
Good research considers relevant competing interpretations. This does not mean giving equal weight to positions with very unequal evidence.
Represent credible disagreement proportionately. If one view is supported by strong replicated evidence and another by weak speculation, explain the asymmetry instead of creating artificial balance.
Neutral research means fair treatment of evidence, not equal treatment of every claim.
Research Bias Checks
- Did the question assume the answer?
- Did the search terms favour one conclusion?
- Did source selection overrepresent one institution or viewpoint?
- Were contradictory findings excluded without reason?
- Did the synthesis turn weak evidence into certainty?
- Did current evidence get mixed with historical evidence?
- Did population differences disappear?
- Did a vendor claim become an independent conclusion?
These checks help reduce confirmation bias in both human and AI-assisted research.
Negative Evidence and Absence of Evidence
Failure to find evidence is not always evidence that something does not exist. Search coverage may be incomplete or the phenomenon may be poorly measured.
Write carefully: “We did not identify current official evidence in the sources reviewed” is different from “There is no evidence.”
SI should preserve the scope of the search when reporting absence.
Research Saturation
Research reaches saturation when additional sources stop changing the important conclusions, source categories are covered and remaining uncertainty is explicit.
Saturation is one practical stop condition for qualitative and broad evidence gathering. It is not a universal statistical rule.
The researcher should still stop earlier when the task’s predefined closure criteria are already satisfied.
Research Reproducibility
A reproducible research workflow records the question, search terms or source plan, inclusion logic, accepted sources and evidence extraction method.
Another researcher should be able to understand why the source set was chosen and how the conclusion was derived.
SI can automate parts of the process, but reproducibility requires preserving enough of the process outside ephemeral generated prose.
Research Logs
Keep a dated research log for non-trivial projects. Record major query changes, source decisions, rejected evidence and scope changes.
The log prevents later confusion about why a source was excluded or why the research question narrowed.
It also makes handoffs easier when more than one person participates.
Inclusion and Exclusion Criteria
Define what qualifies a source for the final evidence set. Criteria might include date range, geography, study type, official status or direct relevance to the outcome.
Exclusion should be principled rather than based on whether the finding agrees with the preferred conclusion.
When criteria change, record the reason and re-evaluate affected sources.
Research Scope Changes
Evidence can reveal that the original question was too broad. Narrowing scope is often a sign of stronger research, not failure.
Example: a global question becomes Singapore-specific after policy differences prove decisive. Update the source hierarchy and conclusion accordingly.
Do not continue combining evidence from the old and new scopes without labels.
Research With Multiple Languages
Relevant evidence may exist in different languages. Machine translation can assist discovery and reading, but important quotations and legal or technical wording deserve careful verification.
Record the original language and source. If interpretation depends on a nuanced term, seek qualified translation or authoritative bilingual material.
Translation is another transformation layer, not invisible access to identical meaning.
Research With PDFs and Tables
PDF research should preserve page locators and inspect tables or figures directly. Text extraction can lose column relationships or footnotes.
When a numerical claim comes from a table, check headers, units, notes and denominators rather than relying on a prose extraction.
SI can accelerate navigation, but document layout remains part of the evidence.
Research With Images and Charts
Charts can summarise evidence but also mislead through truncated axes, aggregation or omitted denominators.
Extract underlying values when possible. Identify source and time period. Do not infer unlabelled quantities from visual appearance alone.
Generated descriptions of charts should be checked against the image and data.
Research With Interviews or Transcripts
Transcripts are primary records of what participants said, not automatic evidence that their claims are factually correct.
Separate participant experience, reported fact and researcher interpretation.
Preserve speaker identity or anonymised role consistently when attribution matters.
Research With Internal Company Data
Internal data can answer questions unavailable publicly, but it may have collection biases, missing definitions or changing schemas.
Document how metrics are defined and whether the data covers the relevant population.
Respect organisational permissions and confidentiality. Research access does not imply authority to publish or redistribute.
Research for Decisions
Decision-oriented research should end with implications, not only facts. Connect evidence to the decision criteria explicitly.
A useful structure is Decision Question, Evidence, Options, Trade-Offs, Unknowns and What Would Change the Decision.
The human decision-maker retains responsibility for value judgments and risk tolerance.
Research for Learning
Students can use SI research to build source literacy. Ask them to identify source type, authority, evidence and limitation rather than merely collecting links.
Require them to explain why one source is stronger for a particular claim than another.
This builds transferable research judgment rather than dependence on generated summaries.
Research for Publishing
Public articles need claim-level verification, current links and stable source provenance. Keep a source packet during drafting.
When updating an article later, refresh volatile claims rather than rewriting evergreen mechanisms unnecessarily.
Internal links should point to published canonical pages rather than planned destinations.
A Claim-to-Source Matrix
- Claim text.
- Claim type: fact, inference, recommendation or forecast.
- Best source type.
- Accepted source.
- Locator.
- Date.
- Population or scope.
- Limitation.
- Verification status.
This matrix is one of the most powerful tools for long research reports because it exposes unsupported prose immediately.
A Research Decision Memo
For practical work, the final report may be shorter than the research process. A decision memo can contain Question, Answer, Evidence, Trade-Off, Risks, Unknowns and Recommendation or Decision Needed.
Keep the evidence packet behind the memo so claims remain traceable.
Compression is safe when the evidence architecture survives behind the concise surface.
A Research Handoff Package
When research passes to another person, provide the question, scope, source register, evidence ledger, current conclusion and unresolved questions.
Do not hand over only the final prose. The next researcher needs to know why the conclusion was reached and where to verify it.
A good handoff makes the research resumable.
Research Maintenance and Refresh
Mark claims by volatility. High-volatility claims may need frequent refresh; stable mechanisms may remain evergreen.
When refreshing, search for updated authoritative sources, check whether definitions changed and identify which conclusions depend on old evidence.
Version major reports when the changes matter to downstream users.
A Final Research Governance Gate
Before accepting an SI-assisted research report, ask whether another informed reader could trace the central claims to evidence, understand the source-selection logic and see the important limitations.
Check that current claims are current, disagreements are not hidden, unknowns remain visible and generated synthesis does not exceed the source material.
If the report will drive a consequential decision, increase independent review proportionately.
Research is complete when the evidence chain is strong enough for the intended use—not when SI has run out of things to say.
The Research Reproducibility Gate
Before calling the research complete, ask whether another informed person could reproduce the logic of the investigation. They do not need to find exactly the same search ranking, but they should be able to understand the question, source criteria, evidence set and path to the conclusion.
Preserve the research prompt or plan, the main search terms, the accepted source register, important exclusions and the claim-to-source matrix. This is enough for most practical research to remain inspectable without keeping every exploratory click.
Reproducibility is particularly important when the report will be updated later. A future researcher can see which source categories were searched and which claims need refreshing instead of beginning from zero.
The Receiver Gate
Research quality is not complete until the output serves its receiver. A student needs a source-backed explanation they can understand. A manager needs the evidence that changes a decision. A public reader needs clear claims, limits and citations. A researcher may need a fuller methods record.
Ask what the receiver must be able to do after reading. If the answer is “make a decision”, foreground the decision-relevant evidence. If the answer is “verify the analysis”, preserve locators and methods. If the answer is “learn the topic”, include conceptual structure as well as facts.
The same evidence set can support different deliverables without changing the facts.
The Research Retirement Rule
A research workflow can accumulate outdated queries, duplicate source lists and obsolete templates. Periodically retire steps that no longer add unique value.
Keep the evidence standards and regression cases that still matter. Remove search rituals that do not change source quality or conclusions.
A mature SI research system becomes more selective over time: better questions, cleaner source sets, clearer evidence and less unnecessary browsing.
A Final Research Exit Checklist
- Question and scope are explicit.
- Evidence requirements are matched to the claim.
- Source selection is documented.
- Key claims have traceable evidence.
- Important citations have been opened and checked.
- Population, date and method are preserved where relevant.
- Credible disagreement remains visible.
- Unknowns are listed.
- Conclusion strength matches evidence strength.
- Current claims are fresh enough for the task.
- Receiver can use the output.
- Research can be resumed or reproduced from the saved record.
When these conditions are met, additional searching should have a reason. Research is not measured by the number of sources collected. It is measured by whether the evidence architecture is strong enough for the conclusion and the decision it supports.
The Final Question: What Would Change the Conclusion?
Every strong research report should name the evidence that could materially change its conclusion. This prevents the report from becoming frozen certainty and gives future updates a clear trigger.
If a new official rule, larger study, corrected dataset or contradictory primary source would change the answer, record that condition. Research then remains open to revision without remaining permanently unfinished.
Research With SI Is Evidence Engineering
The strongest SI research is not the longest report. It is the clearest evidence chain from question to source to finding to bounded conclusion.
Use SI to accelerate discovery, extraction, comparison and synthesis. Keep source authority, human judgment and verification visible.
Return to the complete SI learning hub as Stage 3 continues into web search, source reliability and long-form reading.