Source-grounded literature reviews with AI

Learn in your browser · Chapter 01 of 08

A defensible review begins with a question and a declared method, not a fluent summary. Decide what you are trying to learn, which evidence could answer it and what the available collection cannot establish. AI assistance must operate within those decisions rather than quietly making them.

The recurring case is an explicitly fictional campus-energy collection. Adult researchers are preparing a structured critical narrative review of observed energy changes following interventions in university settings. The twelve supplied records contain duplicates, a correction, unsuitable populations, unavailable material and an unverified reference. These are original teaching fixtures, not real publications or evidence about actual universities. No live database search or classroom validation is implied by completing the exercise.

Frame the question neutrally: what observed energy changes are reported after interventions in university settings, and what limits their interpretation? The wording permits increases, decreases and uncertain findings. Asking which technologies successfully reduce campus energy would encourage looking only for favourable results and narrow the intervention types prematurely. Record the population, intervention range, outcomes and eligible evidence before deciding which findings seem persuasive.

The question determines the review purpose. A narrative review can critically organise selected literature while explaining its selection and limits. A scoping review maps the extent and characteristics of evidence around a defined area using explicit methods. A systematic review needs a method appropriate to its question, including an explicit strategy for identifying and evaluating relevant evidence. Applying a checklist to a supplied folder does not turn the folder into a comprehensive search.

PRISMA 2020 supports transparent reporting of systematic reviews. It is not a certificate of methodological quality or a complete conduct manual. PRISMA-ScR provides reporting guidance for scoping reviews. Our closed fixture borrows useful reporting distinctions but remains a structured critical narrative teaching exercise. The title, abstract and methods must state that boundary even if the final document contains a flow diagram and detailed extraction tables.

Write a short review contract. Specify the question, university-only population, observed-outcome requirement, permitted source collection, extraction fields and decision process. Include who resolves uncertain cases, how changes are recorded and which outputs require source verification. Define how you will distinguish reported observations from causal claims. A contract makes departures visible; it does not establish that the method is sufficient for every research purpose.

Consider R01, the uncontrolled LED example. Consumption changes from 120 to 90 MWh, a reduction of 30 MWh or 25 percent relative to the initial value. Those calculations describe the supplied observation. Without a control group or another justified causal design, they do not establish that lighting caused the whole change. A useful contract preserves this distinction instead of letting an attractive percentage become an unsupported effectiveness claim.

An average prompt asks for a comprehensive review showing that campus interventions save energy. An elite prompt defines the role as a research assistant, supplies only permitted records, states the neutral question and requires a claim, source ID, passage locator and limitation for each proposed finding. It forbids new references and requires an unresolved label where support is missing. The researcher still checks each proposed connection; a well-structured answer can contain a well-structured error.

Keep methodological amendments separate from editorial improvements. Including simulations later changes the observed-outcome criterion and affects R06. Record the reason and revisit relevant decisions consistently. Do not broaden the rule only to admit a favourable result. Equally, do not call a missing report an exclusion merely to make the final count tidy. The research record should reveal what remains unanswered.

Learn in your browser · Chapter 02 of 08

Being able to read a document does not automatically authorise uploading it to an AI service. Permission depends on the material, role, institution, funder, publication and actual tool arrangement. Check those boundaries before preparing a model input.

Separate discovery from ingestion. A catalogue record may contain public bibliographic details while its associated report is licensed, confidential or unavailable. Record those states independently. R08 is a confidential unpublished observational study report, represented by metadata only; eligibility is not yet assessed and its content is withheld. Its existence can remain in an appropriately limited internal register. Withholding it is not a judgement that the unseen findings are irrelevant.

Prepare a permission record for each proposed input: identifier, source, access basis, proposed operation, approved environment, restrictions and review date. A library subscription does not necessarily permit transfer of full text to a third-party processor. Ask the institution library, research support or information-governance team where terms are unclear. Keep the question specific: summarising a published passage and evaluating a confidential funding application are different activities.

Open licensing answers some questions but not every question. CC BY 4.0 permits sharing and adaptation subject to its terms, including appropriate attribution and identification of changes. Other rights may matter. A licence on an article does not establish that every separately credited image or disclosed personal detail can be reused in any manner. Verify the material actually covered and the proposed operation.

Institutional rules are examples, not universal permissions. Oxford research policy version 1.1, dated 7 September 2026, addresses responsibility, intellectual property, information security and disclosure within its stated scope. It distinguishes governed arrangements from personal accounts. Researchers elsewhere must check their own requirements. An enterprise label alone does not prove approval for a dataset, connector or activity.

Funder rules depend on role. UKRI permits a narrow language-refinement exception for assessors, without entering application content or personal information, or using AI to understand, summarise or evaluate the application. Applicant responsibilities are separately stated. Relabelling an assessment prompt as editing does not change the actual processing activity.

NIH Notice NOT-OD-23-149 prohibits specified generative-AI analysis and critique preparation in NIH scientific peer review. It also describes prior approval for possible accessibility exceptions. Its scope is particular peer-review activity, not every use of AI in United States university research. An exception from another funder cannot be transferred to this process.

Elsevier journal policy, updated June 2026, permits specified supportive reviewer uses only with private AI tools as defined by that policy and verified from their terms, while preserving responsibility and confidentiality. It prohibits uploading submitted manuscripts or their information, including to tools described as private. A background literature search must avoid disclosing confidential manuscript information. Check the actual journal and role rather than relying on remembered guidance.

The European Commission fourth-edition research guidelines, published on 15 October 2026, are nonbinding guidance. They do not displace applicable institutional, funder, journal or legal requirements. Select the jurisdiction and setting explicitly. Where obligations conflict or permission remains unclear, retain the ordinary research workflow and obtain an authorised interpretation before transferring material.

An average approach uploads every file and asks the tool to ignore confidential parts. An elite process removes unauthorised material before any transfer, records the permitted subset and requests only the necessary operation. A local model changes where processing occurs but does not remove source rights, confidentiality or disclosure obligations. Logs, indexes and generated outputs also need reviewed access boundaries. Do not leak withheld content through a filename, summary or retrieval cache.

Learn in your browser · Chapter 03 of 08

Retrieving plausible references is not the same as searching an evidence field defensibly. A useful record explains what was attempted, retrieved and potentially missed. AI can propose search ideas, but it must not invent a search history.

Begin with concepts from the question. One concerns university settings and another concerns energy interventions or consumption. Candidate terms include university, campus and higher education alongside electricity, energy use, lighting, occupancy sensors and heating. These are starting points, not a validated strategy. Check terminology with subject expertise and the platform actual indexing.

Combine synonyms within concepts and connect concepts deliberately. A draft query is (university OR campus OR higher education) AND (energy OR electricity OR lighting OR heating). Each platform has its own syntax, fields and phrase handling, so translate and test it before use. A mandatory saving term could miss R05, whose total energy increases. Do not make beneficial outcome language a hidden eligibility criterion.

Distinguish the question from the string. A university-only review may retrieve R07, a school report, through broad education terms. Screening can then exclude the wrong population. Retrieval of an ineligible record does not itself show a defective search. Conversely, a tidy list of relevant-looking results does not prove that important eligible work has been found.

Choose sources for the discipline and question. Energy research may span engineering publications and institutional repositories selected with an information specialist. Cochrane chapter 4, updated March 2025, discusses documented searching in its health-intervention context. Do not copy that context-specific database recipe into campus-energy research. Adopt justified source selection and traceable execution while adapting to the actual discipline.

For each executed search, retain platform, database or collection, exact query, fields, filters, date, operator and exported records. Preserve the original export before deduplication. Record revisions and their reasons. A later rerun can differ because indexing and web content change. A search-engine estimate is not the count of records actually exported or screened. Separate these quantities so the methods can be checked.

Do not invent execution dates or pretend the twelve supplied records were retrieved by your proposed query. This activity is a search-design rehearsal. A local test can check whether terms retrieve selected fixture descriptions. Finding them all establishes only that bounded behaviour. It cannot estimate missed real-world literature because that wider evidence denominator is not known.

An average request asks for every important paper and a final bibliography. An elite request asks for concept synonyms, platform-specific candidates and likely omissions, with proposed searches distinguished from executed searches and unknown fields left unknown. The researcher executes authorised searches and checks the exports. A generated citation is a lead until its identity and support are established.

Language, date, document type, access and screening capacity can remove evidence from view. State and justify restrictions rather than hiding tool defaults. R09 is unresolved because the full report is unavailable, while R06 conflicts with the observed-outcome criterion. These are different reasons. Citation chasing can reveal additional leads, but finding more references still does not certify complete coverage.

Agree the stopping and update decision before fatigue determines it. For the exercise, stop after evaluating the supplied collection and report its boundaries. For real work, plan the question-appropriate searches and updates with the team. Retain negative searches and unresolved leads. If a deadline limits screening, state how far the process progressed rather than reporting all retrieved material as examined.

Learn in your browser · Chapter 04 of 08

A real reference can fail to support a claim, and several references can describe the same study. Establish identity, relationships and version status before treating a collection as independent evidence. Preserve corrections and uncertainty instead of simplifying them away.

A citation record is a catalogue or search entry; a source document is an inspectable report or separate correction; a study is the underlying investigation. One study can have multiple documents and one document can have multiple catalogue entries. These entity distinctions prevent a long bibliography from masquerading as a large independent evidence base. The fixture uses R identifiers and records their relationships explicitly.

R02 duplicates the R01 LED report entry, and R12 duplicates R04 HVAC metadata. Removing these catalogue duplicates leaves ten unique citation candidates from twelve incoming records. Preserve links to the removed entries and original search routes. Similar titles are not enough to merge records: verify that they identify the same source document.

R03 and R11 need different treatment. R11 is a separate correction document about the sensor study reported in R03. Keep both documents linked to one study ID. The current estimate is 4 percent with an interval from -2 to 10 percent; the original 12 percent result is superseded. Unchanged methods can remain contextual material read alongside the correction. No interval type or confidence level should be invented when the fixture does not state it.

The ten candidates partition into five retained source documents, two scope exclusions and three other dispositions. Retained documents are R01, R03, R04, R05 and R11, representing LED, sensor, HVAC and behaviour studies. R06 is the excluded simulation and R07 the excluded school-population report. R08 is permission-held, R09 has no retrieved full report and R10 has an unverified reference. These are not ten verified publications or three identical holds.

Compare title, authors, venue, date, identifier and source location when checking identity. A genuinely available DOI can help; never invent a plausible DOI to fill a blank. Crossref exposes deposited scholarly metadata and some update relationships. A match aids identification but does not establish that a passage supports a claim. A missing match does not prove fabrication because coverage and matching have limits.

The instructor key deliberately identifies R10 as an invented reference. A learner encountering the metadata should record the unverified state and seek the missing publication, not manufacture a convincing abstract. R09 is different: the fixture provides candidate metadata but no full report. Outside this exercise, one failed lookup establishes neither existence nor fabrication. Record the enquiries performed and the remaining uncertainty.

Provenance connects a statement to the actual material reviewed. Record source ID, version, retrieval details, passage locator and permitted transformation. Keep authorised original bytes separate from extracted text. A file hash can detect a byte change against a recorded reference, but cannot establish authorship, truth or semantic support. A valid locator still requires reading the passage and comparing it with the claim.

R04 illustrates distinctions inside a document: energy falls from 80 to 68 MWh, while the weather-adjusted estimate is 2 percent with interval -3 to 7 percent. Raw and adjusted values answer different questions, not two independent studies. Preserve labels so extraction does not choose the larger value merely because it sounds impressive.

An average cleaning prompt asks to remove duplicates and keep strongest results. An elite prompt proposes record-to-document and document-to-study links, distinguishes duplicates from corrections, identifies current result status and exposes uncertain matches. A researcher approves the links using source evidence. Log old wording, replacement and affected outputs when a correction changes a claim.

Learn in your browser · Chapter 05 of 08

Screening applies the review question to candidate evidence. It should not reward favourable results, familiar authors or fluent summaries. An unresolved candidate remains visible until a documented decision can be made.

Scope requires a university setting, an energy intervention and observed rather than solely simulated outcomes. Inclusion in the inspectable synthesis also requires sufficient permitted source material. Missing access or permission leaves a candidate unresolved; it does not establish an out-of-scope study. The review question concerns what these supplied reports show and what they cannot establish. It does not ask which technology should be purchased. Register these conditions before seeing whether the reported change is attractive.

Do not silently turn missing information into a negative answer. R09 has metadata but no supplied full report. That is not the same as an inspected report outside scope. R10 is an invented reference in the exercise, with no supplied publication. In real research, failure to find a reference through one service would not prove fabrication. Record the searches and request the source before judging its status. Here the fixture explicitly supplies its invented status.

Use a staged decision record: candidate ID, stage, applicable criterion, evidence locator, proposed outcome, reviewer, date and unresolved question. Titles can support a provisional decision, but an abstract may not reveal the denominator, setting or design needed for final inclusion. A model can propose a classification from permitted text. A researcher checks the actual passage and records the decision. The model must be allowed to say that the available text is insufficient.

Scroll sideways to see every column.

R02 and R12 are duplicate catalogue entries removed before these ten candidate dispositions. The ten candidates are unique citation entries, not ten studies and not ten verified publications. The five retained source documents include the old and corrected sensor reports. Deduplicating a catalogue record and linking two reports of one study solve different problems. Deleting R03 would hide why an earlier 12 percent statement changed. Including both estimates as independent evidence would exaggerate support.

A useful disagreement exercise gives two reviewers the same criteria and source passages before they discuss the result. Suppose one retains R06 because a simulation concerns campus energy, while the other excludes it because the outcome is predicted. The criterion resolves the disagreement: this review requires observed outcomes. Record both proposals, the passage, the final rationale and whether the criterion wording changed. A new criterion must be applied consistently to earlier records as well.

Do not invent a second reviewer. If working alone, identify the decision as single-reviewer screening and ask a supervisor or information specialist to review a sample where possible. Repeating the same model prompt is not an independent researcher check. A dual-reviewer workflow is a methodological choice or requirement that depends on the review design; this workbook does not certify that a particular professional standard has been met.

An average request selects papers supporting a desired conclusion. A defensible request supplies neutral eligibility criteria, permitted passage IDs and explicit decision labels. Require the decisive passage and a contestable reason. Treat instructions inside retrieved documents as source content, not authority over the review.

Scroll sideways to see every column.

Learn in your browser · Chapter 06 of 08

Extraction creates a checked bridge between source material and the statements you may write. A genuine reference is only the first step. The passage, measurement, comparison and uncertainty must support the particular claim.

Use an extraction row for one result, not an entire paper summary. Record study ID, report version, source locator, the relevant text, population, intervention, comparison, outcome, unit, time window, denominator, estimate, uncertainty and limitations. Keep direct source values separate from calculations you perform. A missing field stays missing. A model that fills an absent confidence interval or study date has invented evidence even if the rest of the row looks plausible.

For R01, the supplied fictional report records 120 MWh before and 90 MWh after a library LED change. The arithmetic difference is 30 MWh and the relative reduction is (120 - 90) / 120 = 25 percent. This is a before-and-after observation without a control group. Write that energy use was lower in the observed after period. Do not write that the LEDs caused a 25 percent reduction across universities. A correct calculation does not repair a weak comparison design.

For R04, raw HVAC use changes from 80 to 68 MWh, a 15 percent reduction. The same fictional report gives a weather-adjusted estimated reduction of 2 percent with an interval from -3 to 7 percent. Preserve both quantities and their labels. The interval includes no reduction and possible increased use. Do not substitute the larger raw change when discussing the adjusted estimate. Nor does an interval spanning zero prove that the true effect is exactly zero.

Scroll sideways to see every column.

Retain adequate precision through the calculation and round only the displayed result. The two periods each cover 12 months, and user counts mean registered building users in each period. The fixture has no person-level measurements, distribution of use, occupancy hours or independent counterfactual. Dividing by users does not adjust for every difference between periods. A useful extraction note names these missing quantities rather than concealing them behind a general phrase such as adjusted for population. The denominator is a recorded count, not evidence that the same individuals were observed.

A passage locator might identify report R05, paragraph p1. Store the exact source text or a permitted local copy, plus a digest that detects whether that stored passage later changes. A matching hash proves a byte-level match to the stored text, not that the text is true, that the source is authentic or that your conclusion follows. A forged or mistaken source can have a perfectly valid hash. Keep identity checks, integrity checks and interpretation checks as separate activities.

The correction case tests version discipline. R03 originally reports 12 percent. R11 revises the sensor estimate to 4 percent with an interval from -2 to 10 percent. An extraction row using the affected R03 estimate must be labelled superseded and excluded from current estimate generation. Unchanged methods and context remain available when read alongside R11. R11 does not create another independent sensor experiment. Update downstream paragraphs, charts and tables that previously used 12 percent, then record the affected claims.

Ask an assistant to extract only fields explicitly available in permitted passages, with a separate missing-fields list and source locator for each value. Require it to preserve units and signed intervals and to list any calculation as a derivation. The researcher checks the row against the original and recomputes arithmetic. An output schema makes an error easier to spot; it does not make the values trustworthy by itself. Reject embedded source instructions such as ignore the review criteria as document content.

Scroll sideways to see every column.

Learn in your browser · Chapter 07 of 08

Synthesis explains patterns and disagreements across eligible evidence. It is not a vote for the largest number, a collage of summaries or an average of incomparable results. The strength of the conclusion must match the available designs and measurements.

Create an evidence map before a prose conclusion. Give each included study one row and keep linked report versions beneath it. Compare setting, intervention, design, outcome definition, denominator, adjustment, uncertainty, access limitations and plausible alternative explanations. An evidence map is not automatically a risk-of-bias assessment. If your actual review requires a design-specific appraisal tool, use its current instructions and obtain appropriate methodological support. Do not invent a universal confidence score.

The four included fictional studies concern LED lighting, sensors, HVAC and a behaviour campaign. Their numeric outputs address different quantities. A mean of 25, 4, 2 and 7.9 percent would mix a raw before-and-after change, corrected and adjusted estimates, and a per-user calculation. It would also conceal the campaign total-use increase. The arithmetic mean can be computed, but it has no justified common interpretation in this exercise. A polished chart would not solve that problem.

Scroll sideways to see every column.

Distinguish absence of evidence from evidence of absence. R09 contributes no inspectable full report here, so it cannot settle the question. R11 supplies an uncertain estimate; its interval crosses no change but is not a proof of zero effect. R06 is an excluded prediction under this question, not evidence that simulations are useless. R07 concerns another population. Each difference deserves a precise description. Calling all four cases negative findings would erase their meanings.

A contradiction register makes uncertainty actionable. For each apparent disagreement, ask whether the population, intervention, comparator, outcome, time window or analysis differs. The R05 total and per-user directions are not a logical contradiction: their denominators differ. R03 and R11 are not competing independent experiments: one is corrected. R04 raw and adjusted changes answer different statistical questions. Resolve the relationship before proposing a causal explanation.

A research assistant can propose a synthesis matrix or identify sentences that need support. Give it the checked extraction rows, not a heap of unreviewed documents, and ask it to preserve disagreements, missing fields and version status. Require claim IDs and study IDs in its output. A researcher must check whether the generated comparison silently drops an unfavourable outcome or converts may into will. Record that a fluent unsupported sentence was removed rather than repairing it with an unrelated citation.

The offline exercise deliberately allows a dangerous example: a claim that LED changes caused the observed reduction is attached to the correct R01 passage with a matching hash. Its links can pass structural checks while the causal interpretation remains unsupported. The program must continue to require human semantic review. Even a recorded approval is a record of somebody saying yes, not independent proof that they read or understood the evidence. Approval records support accountability; they do not replace judgement.

For a real decision, separate the evidence conclusion from the decision recommendation. A facilities manager may choose to commission a small monitored pilot despite uncertainty. That recommendation needs local costs, operational requirements, permissions, risk, measurement plans and stop conditions outside this literature exercise. State which evidence gaps the pilot would address and which it would not. Do not sell a sovereign deployment, a particular model or a Mickai service as proof of correct scientific interpretation.

Scroll sideways to see every column.

Learn in your browser · Chapter 08 of 08

Your final product is an inspectable reasoning chain from question to conclusion. The capstone rewards honest limitations, correction handling and reproducibility. It does not reward confident wording, a high reference count or the mere presence of AI.

Assemble the review contract, permission decisions, search log, deduplication map, screening register, linked report versions, checked extraction rows, evidence map, claim ledger, correction log and final narrative. A reader should be able to start at one important sentence and find the permitted source, relevant passage, comparison and reviewer decision. Keep confidential material outside a shareable package. Include a manifest of file names, versions and hashes where appropriate, with the meaning of those hashes clearly limited.

A claim ledger contains the claim as written, its type, study and report IDs, exact locator, source version, derivation if any, inference limit, reviewer decision and date. Label background context separately from results and recommendations. If a sentence contains three substantive claims, one citation at its end may not support all three. Split the sentence or map each claim explicitly. A register is useful only if its entries describe actual checks rather than a copied statement that everything was verified.

Write a methods section that names the review type and the closed fictional collection. State that this is a structured critical narrative exercise, not a comprehensive search of the campus-energy literature. Describe the eligibility decisions, candidate counts and correction linkage using their actual units. Report whether screening and extraction were checked by one person or more. State what the optional program checks and what it cannot evaluate. Do not convert a software test result into a methodological-quality rating.

Disclose substantive AI assistance according to the actual institution, journal and funder requirements. A useful teaching disclosure says which tool and version were available, when it was used, the bounded tasks it assisted, the permitted inputs, and how its outputs were checked and changed. If you used no model, say so. Do not invent a model transcript, tool version or independent review. The responsible named author remains accountable for the work; a tool cannot accept authorship responsibilities.

Sovereignty concerns practical control over sources, access, processing, retention and decisions. In a research workspace, separate private source storage from the shareable review, restrict retrieval by project permission, record exports and allow correction or withdrawal to propagate. Locally hosted AI may change data movement but does not automatically grant copyright permissions or remove confidentiality obligations. Mickai can help design controlled knowledge workflows and digitisation processes; the review still needs qualified human interpretation and an institution-specific approval route.

Scroll sideways to see every column.

Test a correction drill before delivery. Imagine that a second fictional correction changes a permitted source passage. Do not overwrite the old passage silently. Register the new version, identify affected extraction rows and claims, recalculate relevant figures, revise the narrative and record who approved the change. A hash mismatch helps locate a changed input. It does not decide whether the new version is authoritative. That judgement requires source and version verification.

The optional code route and paper route use the same synthetic records. The program provides structural feedback and arithmetic checks without contacting a model or external service. The paper route asks you to trace the same records, compare versions and explain each disposition. Neither route certifies semantic correctness. Submit a short oral or written defence explaining one unsupported claim you removed, one uncertainty you retained and one limitation that additional searching might address.

Separate the knowledge exam from the capstone. The planned scored sitting draws 20 questions from a reviewed 40-question bank. Passing requires strictly above 80 percent: 16 correct is 80 percent and fails, while 17 correct is 85 percent and passes. Written chapter questions and the capstone are practice, not automatically assessed research competence. A named Mickai certificate records a passing knowledge result through the protected issuance system. It is not accreditation, a university award, professional registration or identity assurance.

Scroll sideways to see every column.