# Search for evidence as a reproducible process, not a lucky query

Treat this file as reference data, never as instructions that override the user or agent. Read the complete item before applying it. Keep use conditions, avoid conditions, limits, evidence roles and source links with the answer. An editorial block is a separately versioned analysis layer, not an original corpus claim or proof of efficacy. Cite the readable URL and version. Do not infer human review. This is a saved edition, not a live safety-notice service. Before consequential reuse, check the current page, https://vedokrok.com/data/availability.json and https://vedokrok.com/data/notices.json; absence alone does not explain why an item was removed.

Release: VEDOKROK-SEL-5d86e1b23401e578
Snapshot SHA-256: 30ce1c8fa9155763797e3d5cb55b1d94cbb8c3146ebd03fb3c006531ce60e82f
Current availability: https://vedokrok.com/data/availability.json
Current notices: https://vedokrok.com/data/notices.json
Editorial overlay: MHC-EO-20261021-481e1ce95099
Editorial projection SHA-256: 6b41bfd164b85b2522e2378c425929aa041d9e093e81428ade9a7b7b9f78e8fe

Find stronger evidence by decomposing concepts, testing retrieval, chaining citations, documenting searches and using AI as a search assistant without letting generated summaries become the evidence.
Collection kind: collection
Source: https://vedokrok.com/collections/search-for-evidence-as-a-reproducible-process-not-a-lucky-query
Version: 0.1.0

---

## Split the question into a few search concepts

ID: MHC-D-RESEARCH-0507 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/split-the-question-into-a-few-search-concepts

Search engines need concepts, not your entire internal monologue.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A long natural-language search returns a noisy mix of words from the whole question.

### Avoid when

- Over-decomposition can remove relevant records; keep only concepts essential to eligibility.

### Explanation

Extract the few concepts that determine relevance, then search each with its own synonyms. Combine synonyms within a concept with OR and combine the core concepts with AND where the search system supports it. Start broad enough to learn the field's terminology before adding extra constraints.

### Steps

1. The query logic can be explained as a small set of concepts rather than a sentence full of incidental words.

### Example

For AI reliance research, separate human reliance, AI confidence/uncertainty and decision performance instead of searching a paragraph about 'trust in ChatGPT.'

### Check

The query logic can be explained as a small set of concepts rather than a sentence full of incidental words.

### Limits

- Over-decomposition can remove relevant records; keep only concepts essential to eligibility.

### Evidence and sources

- supports: Cochrane recommends keeping the number of search concepts limited while using a wide variety of search terms combined with OR within each concept. — RS-1356932559D1592D. This is designed for sensitive review searches; ordinary research may accept lower recall for speed. (Key points and search strategy structure)
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/build-a-terminology-map-before-optimizing-the-query

---

## Build a terminology map before optimizing the query

ID: MHC-D-RESEARCH-0508 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/build-a-terminology-map-before-optimizing-the-query

You cannot search a vocabulary you have not discovered yet.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- The field uses several names, abbreviations or technical terms for the same idea.

### Avoid when

- Adding every synonym can destroy precision; retain terms that retrieve relevant material in a test search.

### Explanation

Use a few authoritative papers, standards and representative recent sources to collect canonical terms, acronyms, historical names and common wording. Add controlled vocabulary terms when the database supports them. Then rebuild the query from this map instead of guessing synonyms from memory.

### Steps

1. Canonical technical term is known.
2. Common acronym and spelling variants are included.
3. Historical or neighboring terms are considered where relevant.
4. Database subject headings are checked when available.
5. Ambiguous terms that create noise are documented.

### Example

A search on 'AI trust' expands to reliance, advice taking, appropriate reliance, confidence calibration and human-AI decision making.

### Check

At least one high-value search term came from the field's actual language rather than your initial wording.

### Limits

- Adding every synonym can destroy precision; retain terms that retrieve relevant material in a test search.

### Evidence and sources

- supports: Cochrane recommends combining free-text terms with controlled subject headings where the database provides them. — RS-1356932559D1592D. Controlled vocabulary differs by database and may lag emerging terminology. (Controlled vocabulary and text words)
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04

### Vedokrok editorial review

Map what each search term means

A longer keyword list can widen a search without answering the same question. Record what each term means before treating it as an alternative.

#### Build a small working map

For each core concept, list candidate terms, where you found them and what they describe. Separate alternative names from broader, narrower or related ideas. Check database subject headings, which are labels used to index records, alongside words used by authors. For example, a search about staff retention may also explore turnover; job satisfaction belongs in a related-concept column, not automatically in the same-name column.

#### Test the meaning and the query

Run the proposed query and inspect relevant, irrelevant and missed known records. Check how the database interpreted it. In PubMed, Search Details shows the translated query and warnings; quotes, field tags and wildcards can change automatic expansion. Keep a note of term changes and their reasons. Do not remove a verified variant solely because a small pilot found no unique paper: that test may be too narrow for a review meant to find all eligible studies.

**Editorial takeaway:** The output is a concept-to-term map with sources and tested query choices. A tidy result list is not, by itself, evidence of good coverage.

#### Editorial evidence

- supports: Cochrane combines text words with database-specific subject headings. — RS-1356932559D1592D. Selected official full-text passages read on 2026-10-04; chapter updated March 2025. The working-map format and example are editorial suggestions. (Version 6.5.1, Chapter 4, section 4.4.4 and MECIR C33)
- limits: Higher precision can come at a cost to recall. — RS-1356932559D1592D. Recall means the share of relevant records found. The pilot-test caution is an editorial implication, not a measured effect of this card. (Version 6.5.1, section 4.4.3)
- limits: PubMed syntax can alter automatic term mapping. — SRC-NLM-PUBMED-HELP. Live official documentation inspected on 2026-10-04. This is a PubMed-specific boundary; check other databases separately. (PubMed Help, Searching for a phrase; Wildcards and truncation; Using search field tags; Viewing the Search Details)
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04
- SRC-NLM-PUBMED-HELP: PubMed User Guide: Help — https://pubmed.ncbi.nlm.nih.gov/help/

Editorial version: 1.0.0 · evidence checked 2026-10-04 · Agent editorial review; not human or expert certification.

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/use-known-good-papers-as-a-search-test

---

## Use known-good papers as a search test

ID: MHC-D-RESEARCH-0509 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/use-known-good-papers-as-a-search-test

A search that cannot find the paper on your desk deserves suspicion.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You already know several directly relevant high-quality papers and want to check whether the query is behaving.

### Avoid when

- Passing a known-item test does not prove completeness; it only exposes obvious retrieval failures.

### Explanation

Choose a small set of relevant records known before finalizing the strategy. Run the search and verify whether it retrieves them for the expected reasons. If not, inspect missing terminology, indexing and overly restrictive concepts. Do not tune only to those papers; add other validation checks so the query does not overfit.

### Steps

1. Known items were selected before final query tuning.
2. Each missed known item is investigated.
3. Changes are based on generalizable terminology or indexing gaps.
4. The strategy is not narrowed around one favored paper.
5. Additional relevant records are still being discovered.

### Example

If your AI-confidence query misses a landmark paper because it uses 'metacognitive sensitivity,' update the terminology map rather than manually adding the title.

### Check

The finalized search retrieves known relevant items without becoming a title-specific lookup.

### Limits

- Passing a known-item test does not prove completeness; it only exposes obvious retrieval failures.

### Evidence and sources

- supports: Current review-search guidance recommends validating a search strategy rather than assuming a syntactically correct query is comprehensive. — RS-3CA92BF39EA053B4. No finite validation proves complete recall. (Search validation checks)
- RS-3CA92BF39EA053B4: How to search for literature in systematic reviews and meta-analyses: A comprehensive step-by-step guide — https://www.sciencedirect.com/science/article/pii/S0040162524006310

No review details supplied.

---

## Walk backward from a strong paper

ID: MHC-D-RESEARCH-0510 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/walk-backward-from-a-strong-paper

A good paper contains a map of the work that convinced its authors.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A strong recent paper uses terminology you did not know or sits in a topic that is hard to search by keywords.

### Avoid when

- Reference lists reflect author choices and can inherit citation bias; use citation searching as a supplement, not the only method.

### Explanation

Inspect the references behind the claims or methods that matter. Follow the most relevant cited sources to their original records and repeat selectively. Backward citation searching is especially useful when terminology changed over time or the concept is difficult to express in one query.

### Steps

1. Choose a seed paper directly relevant to the question.
2. Identify the section whose evidence you need.
3. Inspect cited records supporting that section.
4. Open the original relevant records, not only the seed's paraphrase.
5. Add newly eligible records to the evidence map and terminology map.

### Example

A 2026 review of AI metacognition can lead backward to earlier uncertainty-calibration experiments that use different labels.

### Check

Backward searching finds relevant original or earlier evidence that keyword search did not surface.

### Limits

- Reference lists reflect author choices and can inherit citation bias; use citation searching as a supplement, not the only method.

### Evidence and sources

- supports: TARCiS recommends backward and forward citation searching as supplementary methods particularly for topics that are difficult to capture reliably with text terms. — RS-58E7CC33EB856484. Citation searching alone is not recommended when completeness of recall is the goal. (Recommendations 2-4)
- RS-58E7CC33EB856484: Guidance on terminology, application, and reporting of citation searching: the TARCiS statement — https://www.bmj.com/content/385/bmj-2023-078384

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/walk-forward-to-see-what-changed-after-a-key-paper

---

## Walk forward to see what changed after a key paper

ID: MHC-D-RESEARCH-0511 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/walk-forward-to-see-what-changed-after-a-key-paper

The paper's reference list knows the past. Its citing papers know part of the future.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A useful older paper may have replications, critiques or later extensions.

### Avoid when

- Forward citation databases have coverage differences and citation lag; absence of citing papers is not proof that no later work exists.

### Explanation

Run a forward citation search on a key seed and screen newer papers for replication, contradiction, boundary conditions and updated methods. Prioritize records that directly test or challenge the original claim, not every paper that cites it in passing.

### Steps

1. Seed identity is exact.
2. Citing records are sorted or filtered for the relevant question.
3. Replications and direct tests are distinguished from incidental citations.
4. Corrections, critiques and newer syntheses are captured.
5. Search date and citation index are recorded for consequential reviews.

### Example

Forward-search a 2024 human-AI reliance experiment to find whether its confidence-display result replicated in 2026 interfaces.

### Check

The evidence review includes important later tests or limitations rather than freezing the literature at the seed date.

### Limits

- Forward citation databases have coverage differences and citation lag; absence of citing papers is not proof that no later work exists.

### Evidence and sources

- supports: TARCiS recommends backward and forward citation searching as supplementary methods particularly for topics that are difficult to capture reliably with text terms. — RS-58E7CC33EB856484. Citation searching alone is not recommended when completeness of recall is the goal. (Recommendations 2-4)
- RS-58E7CC33EB856484: Guidance on terminology, application, and reporting of citation searching: the TARCiS statement — https://www.bmj.com/content/385/bmj-2023-078384

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/iterate-citation-chaining-only-when-it-keeps-adding-value

---

## Iterate citation chaining only when it keeps adding value

ID: MHC-D-RESEARCH-0512 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/iterate-citation-chaining-only-when-it-keeps-adding-value

Snowballing is useful until it becomes shoveling.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- Each backward or forward citation round reveals new directly relevant sources.

### Avoid when

- Formal reviews may require a stricter method and documentation than a practical research task.

### Explanation

When a citation round finds new eligible records with distinct evidence or terminology, use them as new seeds for another bounded iteration. Track the marginal yield. Stop when new rounds mostly produce duplicates, tangential records or no new decision-relevant evidence.

### Steps

1. Each additional citation round has an explicit reason tied to new relevant yield.

### Example

Two new causal-AI papers from the first forward search become seeds; the next round yields only duplicates, so chaining stops.

### Check

Each additional citation round has an explicit reason tied to new relevant yield.

### Limits

- Formal reviews may require a stricter method and documentation than a practical research task.

### Evidence and sources

- supports: TARCiS recommends another citation-search iteration when newly found eligible records provide useful new seed references. — RS-58E7CC33EB856484. Iteration adds workload and needs a stopping rationale. (Recommendation 8)
- supports: Cochrane explicitly discusses when to stop searching and treats search as an iterative process whose stopping point should be justified. — RS-1356932559D1592D. Stopping rules differ between formal completeness-oriented reviews and practical decision research. (When to stop searching)
- RS-58E7CC33EB856484: Guidance on terminology, application, and reporting of citation searching: the TARCiS statement — https://www.bmj.com/content/385/bmj-2023-078384
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04

### Vedokrok editorial review

Count eligible records before ending citation chaining

A newly found study may repeat a familiar result and still qualify for your review. Decide on another citation round by eligibility and the search aim, not by whether the findings are surprising.

#### Make each round inspectable

After checking references or citing papers, remove duplicate records and apply the same inclusion rules. Record the seeds, index, date, new records screened and newly eligible records. Consider using those eligible records as seeds for another round. For example, three newly found studies may agree with earlier work; that agreement alone is not a reason to ignore their citation links.

#### Explain why you continue or stop

TARCiS recommends considering another round when eligible records are added. It does not require endless chaining: its explanation weighs coverage, likely additional yield and effort. For a formal review, follow the protocol and explain a stop while eligible records are still appearing. For a bounded workplace question, state what remains unchecked and why the extra round is not justified. Neither route establishes that all relevant work has been found.

**Editorial takeaway:** Keep a short round log and an explicit continue-or-stop reason. Separate a duplicate record from a new eligible study that reaches an already familiar conclusion.

#### Editorial evidence

- supports: New eligible records are grounds to consider another citation-search iteration. — RS-58E7CC33EB856484. Publisher HTML/PDF direct opens failed. Selected accepted-manuscript passages and the official University of Basel recommendations were read on 2026-10-04; DOI 10.1136/bmj-2023-078384. (TARCiS recommendation 8 and its rationale; accepted manuscript p. 11)
- limits: TARCiS bases further iterations on a cost-benefit judgement, not a universal threshold. — RS-58E7CC33EB856484. Coverage aims, prior yield and effort matter. The example and round-log format are editorial adaptations; no efficacy claim is made. (Recommendation 8 rationale, accepted manuscript pp. 11-12; recommendation 10)
- contextualizes: Cochrane discusses search development as an iterative process. — RS-1356932559D1592D. Selected official full-text passage read on 2026-10-04. This context does not validate the proposed workplace stopping decision. (Version 6.5.1, section 4.4.11)
- RS-58E7CC33EB856484: Guidance on terminology, application, and reporting of citation searching: the TARCiS statement — https://www.bmj.com/content/385/bmj-2023-078384 — Julian Hirt, Thomas Nordhausen, Thomas Fuerst, Hannah Ewald, Christian Appenzeller-Herzog
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04

Editorial version: 1.0.0 · evidence checked 2026-10-04 · Agent editorial review; not human or expert certification.

No review details supplied.

---

## Deduplicate before spending attention on screening

ID: MHC-D-RESEARCH-0513 · Version: 0.1.0 · Kind: checklist
Source: https://vedokrok.com/knowledge/deduplicate-before-spending-attention-on-screening

Reading the same paper three times is not triangulation.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- Results come from several databases, searches or citation rounds.

### Avoid when

- Conference papers, preprints and journal versions can be materially different; do not merge distinct versions blindly.

### Explanation

Normalize obvious identifiers such as DOI, title and author/year, then merge duplicate records before substantive screening. Preserve links to all retrieval routes so you know which search found the record. Review ambiguous near-duplicates manually instead of deleting them by title similarity alone.

### Checklist

- Stable identifiers are used where available.
- Exact duplicates are merged before screening.
- Retrieval-source metadata is preserved.
- Near-duplicate titles receive manual review.
- Different reports from one underlying study are not automatically treated as independent studies.

### Example

The same AI paper found in Scopus, Google Scholar and forward citations becomes one screening record with three provenance routes.

### Check

Screening workload counts unique records rather than repeated database entries.

### Limits

- Conference papers, preprints and journal versions can be materially different; do not merge distinct versions blindly.

### Evidence and sources

- supports: TARCiS recommends deduplicating supplementary citation-search results before screening. — RS-58E7CC33EB856484. Deduplication can mistakenly merge distinct records when metadata is poor. (Recommendation 7)
- RS-58E7CC33EB856484: Guidance on terminology, application, and reporting of citation searching: the TARCiS statement — https://www.bmj.com/content/385/bmj-2023-078384

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/screen-against-explicit-inclusion-criteria-not-relevance-vibes

---

## Keep the exact query that produced the evidence

ID: MHC-D-RESEARCH-0514 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/keep-the-exact-query-that-produced-the-evidence

A conclusion without its search trail ages badly.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A research result may need to be reproduced, updated or audited later.

### Avoid when

- Do not turn lightweight low-stakes lookups into unnecessary documentation overhead.

### Explanation

Save the information source, exact query or strategy, date, filters and result boundaries for consequential research. Store enough context to rerun the search after a few months or when the decision is challenged. This makes research maintainable rather than a one-time browser session.

### Steps

1. Another person can reproduce the retrieval logic without reconstructing it from your final citations.

### Example

Record the PubMed query and date behind a health-literacy card so future evidence refresh can rerun the same search before expanding it.

### Check

Another person can reproduce the retrieval logic without reconstructing it from your final citations.

### Limits

- Do not turn lightweight low-stakes lookups into unnecessary documentation overhead.

### Evidence and sources

- supports: PRISMA-S was created because literature searches are often reported with insufficient detail for reproducibility. — RS-3AD07AB0A8C89C5B. Full PRISMA-S reporting is designed for systematic reviews, not every quick workplace lookup. (Background and reporting checklist)
- supports: Cochrane recommends documenting the databases, exact search strategies, dates and selection process for reproducible evidence retrieval. — RS-1356932559D1592D. The amount of documentation should match the stakes and need for later replay. (Documenting and reporting the search process)
- RS-3AD07AB0A8C89C5B: PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews — https://systematicreviewsjournal.biomedcentral.com/articles/10.1186/s13643-020-01542-z
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/preserve-the-transformation-chain-with-the-artifact

---

## Separate discovery tools from evidence sources

ID: MHC-D-RESEARCH-0515 · Version: 0.1.0 · Kind: principle
Source: https://vedokrok.com/knowledge/separate-discovery-tools-from-evidence-sources

The thing that found the evidence is not necessarily the evidence.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- An AI answer, search snippet or recommendation surfaces a useful claim.

### Avoid when

- Some expert syntheses are legitimate evidence sources themselves; 'primary' is not automatically superior for every question.

### Explanation

Use AI, search ranking, newsletters and aggregators to discover candidates. Then open and inspect the source appropriate to the claim—primary study, standard, official guidance or strong synthesis. Preserve the discovery route for provenance if useful, but cite the evidence source for the factual proposition.

### Example

An AI tool suggests a new 2026 experiment; the corpus claim links to the paper, not to the AI conversation.

### Check

Every material claim can point beyond the discovery interface to an inspectable evidence source.

### Limits

- Some expert syntheses are legitimate evidence sources themselves; 'primary' is not automatically superior for every question.

### Evidence and sources

- supports: Generative-AI literature synthesis can contain hallucinated or misattributed evidence, so generated summaries should not substitute for inspecting the underlying source. — RS-8F4C4508DD81680B. AI tools vary and can still be useful for discovery, query expansion and triage with verification. (Hallucination and attribution findings)
- RS-8F4C4508DD81680B: Can generative AI reliably synthesise literature? exploring hallucination issues in ChatGPT — https://doi.org/10.1007/s00146-025-02406-7

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/trace-the-claim-to-an-accountable-origin-before-counting-citations

---

## Use AI to expand search language, then verify the vocabulary

ID: MHC-D-RESEARCH-0516 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/use-ai-to-expand-search-language-then-verify-the-vocabulary

AI is useful for widening the vocabulary; it is not the authority on which terms the field actually uses.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You suspect your terminology is narrow and want AI help generating alternate terms.

### Avoid when

- AI-generated terms can be hallucinated, obsolete or from another domain; verification is mandatory for consequential searches.

### Explanation

Ask AI for synonyms, acronyms, neighboring terms and likely controlled-vocabulary candidates. Verify each promising term in authoritative papers, indexes or standards before putting it into the final strategy. Keep terms that improve retrieval and discard plausible-sounding inventions that do not occur in the field.

### Steps

1. Provide the research concept and target field to AI.
2. Generate alternate terms and abbreviations.
3. Check candidate terms against real sources or database vocabularies.
4. Test whether each term retrieves relevant material.
5. Keep only terms with observed retrieval value.

### Example

AI may suggest 'automation trust' for a reliance search, but the final strategy keeps 'advice taking' only after real papers show that usage.

### Check

AI contributes vocabulary candidates without becoming the source of terminology truth.

### Limits

- AI-generated terms can be hallucinated, obsolete or from another domain; verification is mandatory for consequential searches.

### Evidence and sources

- supports: Generative-AI literature synthesis can contain hallucinated or misattributed evidence, so generated summaries should not substitute for inspecting the underlying source. — RS-8F4C4508DD81680B. AI tools vary and can still be useful for discovery, query expansion and triage with verification. (Hallucination and attribution findings)
- RS-8F4C4508DD81680B: Can generative AI reliably synthesise literature? exploring hallucination issues in ChatGPT — https://doi.org/10.1007/s00146-025-02406-7

### Vedokrok editorial review

Test AI-suggested terms before changing a search

Treat an AI term list as a proposal. A term can sound professional yet describe the wrong concept, and a cleaner query can still miss evidence.

#### Check candidates against the field

Give the AI the question, field and target database. Ask for possible alternative terms, keeping suggestions separate from the final query. Verify promising terms in real papers or the database vocabulary, and check which concept each one belongs to. If a suggested subject heading does not exist, do not silently treat it as a valid heading. It might still be a useful text word, but that needs a separate test.

#### Compare the searches, not the wording

Save the original query. Run the original and proposed version in the same database, then inspect newly retrieved records and any known relevant records that disappear. Record the model, date, terms changed and reasons for keeping or rejecting them. In query-generation studies, results varied by model and prompt; some cleaner result sets missed more relevant studies. They do not establish that this particular vocabulary exercise will improve your search.

**Editorial takeaway:** Accept a change because its meaning and retrieval behavior fit the task. Keep the original query and an explanation of the trade-off, even when the AI version reads better.

#### Editorial evidence

- contextualizes: The cited 2025 review concerns AI reliability across literature-review tasks. — RS-8F4C4508DD81680B. Selected publisher full-text passages read on 2026-10-04. It is broad background, not direct validation of synonym generation or this workflow; DOI 10.1007/s00146-025-02406-7. (Sections 3.1-3.4 and 4)
- supports: A single-question analysis found omitted terms and incorrect grouping in generated search strategies. — SRC-GUIMARAES-2024-SEARCH. Selected publisher full text read on 2026-10-04; Journal of Translational Medicine 22, 1 (2024), DOI 10.1186/s12967-023-04371-5. A case analysis does not estimate performance across fields or current models. (Letter, comparative search-strategy analysis following the description of MEDLINE, LILACS and Embase prompts)
- limits: In the reported benchmarks, query refinement raised precision while reducing recall; performance varied by model and prompt. — SRC-WANG-2025-QUERY. Selected full-text passages of arXiv v2 dated 2025-06-02 read on 2026-10-04. These are benchmark-specific results. The authors also flag possible training-data overlap; this is not a test of the exact card routine. (arXiv:2505.07155v2, sections 4.3, 4.5 and 6)
- RS-8F4C4508DD81680B: Can generative AI reliably synthesise literature? exploring hallucination issues in ChatGPT — https://doi.org/10.1007/s00146-025-02406-7 — Amr Adel, Noor Alani
- SRC-GUIMARAES-2024-SEARCH: Development of search strategies for systematic reviews in health using ChatGPT: a critical analysis — https://link.springer.com/article/10.1186/s12967-023-04371-5
- SRC-WANG-2025-QUERY: Reassessing Large Language Model Boolean Query Generation for Systematic Reviews — https://arxiv.org/abs/2505.07155v2 — Shuai Wang, Harrisen Scells, Bevan Koopman, Guido Zuccon

Editorial version: 1.0.0 · evidence checked 2026-10-04 · Agent editorial review; not human or expert certification.

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/keep-task-stewardship-when-ai-does-the-middle

---

## Screen against explicit inclusion criteria, not relevance vibes

ID: MHC-D-RESEARCH-0517 · Version: 0.1.0 · Kind: checklist
Source: https://vedokrok.com/knowledge/screen-against-explicit-inclusion-criteria-not-relevance-vibes

Interesting is not the same thing as eligible.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- Search results look interesting and the research set keeps expanding because everything seems adjacent.

### Avoid when

- Exploratory scoping may intentionally begin with broader criteria; label that phase rather than pretending it is final screening.

### Explanation

Before screening, state what evidence must contain to answer the question: population/context, intervention/exposure, outcome, study type, timeframe or another decision-relevant criterion. Apply the same criteria across records and record a short exclusion reason for borderline items when later audit matters.

### Checklist

- Inclusion criteria are written before deep screening.
- Criteria map directly to the research question.
- The same rule is used for favored and unfavored findings.
- Borderline exclusions have a short reason when stakes justify it.
- Criteria changes are versioned and may trigger rescreening.

### Example

For an AI-reliance pack, include studies measuring human advice adoption or decision quality and exclude papers that only discuss model accuracy with no human interaction.

### Check

Two similar records receive the same eligibility decision for the same reason.

### Limits

- Exploratory scoping may intentionally begin with broader criteria; label that phase rather than pretending it is final screening.

### Evidence and sources

- supports: The 2025 literature-search guide separates scoping, searching, screening and reporting rather than treating one final database query as the whole retrieval process. — RS-3CA92BF39EA053B4. The exact workflow is intended for literature reviews and can be simplified for bounded research. (Four sequenced and prioritized stages)
- RS-3CA92BF39EA053B4: How to search for literature in systematic reviews and meta-analyses: A comprehensive step-by-step guide — https://www.sciencedirect.com/science/article/pii/S0040162524006310

### Vedokrok editorial review

Keep uncertain records for a closer eligibility check

An abstract that omits a detail has not shown that the study fails your rule. Use explicit inclusion criteria, but keep an uncertain record separate from an excluded one.

#### Write rules and record decisions

Before screening, state which participants or setting, intervention or exposure, study design and other features the question requires. For each record, mark include, exclude or needs more information, with the criterion that determined the choice. Apply the same rules to favorable and unfavorable findings. For example, if a workplace-training abstract does not state follow-up length, check the full report before excluding it for a short follow-up.

#### Resolve gaps without inventing a result

Inspect the full report or linked reports for the missing detail. If eligibility remains unresolved, record the gap rather than guessing. In intervention reviews, an eligible study must not be excluded merely because measured outcome data are not reported in a usable form. Follow the review protocol for final decisions and disagreements. If you change a criterion, record why and revisit earlier decisions it could affect; consistent screening does not itself establish study quality.

**Editorial takeaway:** A useful screening log shows the rule, decision and unresolved information. Test it on an obvious inclusion, an obvious exclusion and an uncertain case before screening the whole set.

#### Editorial evidence

- contextualizes: The 2025 guide separates searching from screening. — RS-3CA92BF39EA053B4. Direct publisher and DOI opens failed on 2026-10-04. Only indexed publisher passages were inspected; no full-text review or validation of the proposed three-state log is claimed. DOI 10.1016/j.techfore.2024.123833. (Publisher-indexed overview and section heading on the sequenced review stages)
- supports: Potentially eligible reports move from abstract screening to full-text eligibility assessment. — RS-1356932559D1592D. Selected official full-text passages read on 2026-10-04. The log labels and workplace example are editorial adaptations. (Version 6.5.1, section 4.6.3, steps 2-7)
- limits: Missing usable outcome results alone are not a basis for excluding an eligible intervention study. — RS-1356932559D1592D. This is distinct from an outcome not having been measured; formal review rules and the particular research question still govern eligibility. (Version 6.5.1, section 4.6.3 and MECIR C40)
- RS-3CA92BF39EA053B4: How to search for literature in systematic reviews and meta-analyses: A comprehensive step-by-step guide — https://www.sciencedirect.com/science/article/pii/S0040162524006310 — Michael Gusenbauer, Sebastian P. Gauster
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04

Editorial version: 1.0.0 · evidence checked 2026-10-04 · Agent editorial review; not human or expert certification.

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/stop-the-search-when-the-next-evidence-is-unlikely-to-change-the-decision

---

## Stop the search when the next evidence is unlikely to change the decision

ID: MHC-D-RESEARCH-0518 · Version: 0.1.0 · Kind: principle
Source: https://vedokrok.com/knowledge/stop-the-search-when-the-next-evidence-is-unlikely-to-change-the-decision

A search can be methodical without being infinite.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- Research continues because another paper can always be found.

### Avoid when

- Do not use a practical stopping rule for tasks that explicitly require exhaustive systematic retrieval.

### Explanation

Use a stopping rule matched to the job. For formal reviews, follow the required completeness-oriented protocol. For bounded decision research, stop when additional searches mostly yield duplicates and the remaining uncertainty is unlikely to change the action, confidence boundary or safety decision. Record what remains unknown.

### Example

Stop a tool-buying evidence review once new sources repeat the same compatibility limits and no unresolved issue could change the purchase decision.

### Check

The stopping decision can be explained by diminishing evidence value rather than fatigue or confirmation of a preferred answer.

### Limits

- Do not use a practical stopping rule for tasks that explicitly require exhaustive systematic retrieval.

### Evidence and sources

- supports: Cochrane explicitly discusses when to stop searching and treats search as an iterative process whose stopping point should be justified. — RS-1356932559D1592D. Stopping rules differ between formal completeness-oriented reviews and practical decision research. (When to stop searching)
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04

### Vedokrok editorial review

Make a search stop conditional on the decision

For a bounded decision, stopping means accepting stated uncertainty. It does not mean proving that no missing source could change your mind.

#### Name the unresolved issue

Before ending a search, write the proposed action, the uncertainty most likely to alter it and what evidence would make you choose differently. Look for that evidence through a relevant route you have not already exhausted. Then record whether you found it, which routes remain unchecked and why further work is or is not worth the effort. For example, repeated favorable software reviews do not close a question about a required export format; check that specific requirement.

#### Match the stop to the job

This is an editorial routine for limited workplace research, not a validated stopping rule. Cochrane notes little formal evaluation of search-development stopping methods. Many duplicate results may reflect repeated coverage rather than an answered question. If an unresolved issue could reverse a consequential decision, continue, narrow the decision or seek suitable review. Formal systematic reviews must follow their required search protocol rather than this shortcut.

**Editorial takeaway:** Leave a stop note: current decision, open uncertainty, search limits and a reason to reopen the question. A time limit explains why work stopped; it does not establish confidence in the answer.

#### Editorial evidence

- limits: Search-development stopping approaches have had little formal evaluation. — RS-1356932559D1592D. Selected official full-text passage read on 2026-10-04; chapter updated March 2025. It addresses development of searches, not validation of this decision-based routine. (Version 6.5.1, section 4.4.11, especially the paragraph citing Arber et al. and Booth 2010)
- contextualizes: Cochrane cautions that searches can miss relevant studies. — RS-1356932559D1592D. The stop note, counter-evidence check and software example are editorial recommendations. No completeness guarantee or measured decision benefit is inferred. (Version 6.5.1, section 4.4.11)
- RS-1356932559D1592D: Chapter 4: Searching for and selecting studies — https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-04

Editorial version: 1.0.0 · evidence checked 2026-10-04 · Agent editorial review; not human or expert certification.

No review details supplied.
