From a question to a reviewable finding

Work through the evidence

Record deviations, failed searches and unresolved objections. Stop when the evidence cannot support the intended claim.

1. Define the decision. State the question, intended user, what decision it informs, and what observation would change that decision. Record the author and the date the scope was defined. Do not presuppose that pollution worsened or that a policy caused harm.
2. Bound the study. Specify population, place, dates, variables, units, inclusion criteria and exclusions before searching. For monitoring coverage, define pollutant, active dates, geographic boundary and the exact meaning of a gap. Do not equate distance to a monitor with validated exposure coverage.
3. Map relevant sources. Select authoritative primary records for each claim type. Consider the five discovery/data providers, recording “not applicable” with reasons where appropriate. Add necessary sources outside those five; do not force every question through every provider.
4. Collect and log. Record exact query, filters, retrieval time, pagination, retrieved/screened counts, failures, source identifier, version and reuse conditions. Keep permitted input snapshots. A researcher-reported search must be labeled as such.
5. Inspect originals. Open original papers or datasets. Capture exact page/table/section or short passage, study method, population, limitations and claim-to-evidence explanation. Mark records with missing locations or unread originals as “Needs source inspection.” A title alone is a lead.
6. Deduplicate and align. Match DOI/PMID/trial/dataset identifiers, versions and shared cohorts. Record independence as unresolved where needed. Align geography, dates, units, endpoints and populations before comparing. Preserve unmatched records as gaps.
7. Draft a comparison. Separate what sources report from researcher inference. Build a claim-by-evidence table identifying support, challenge, context and uncertainty. Produce only as many candidate claims as the evidence supports; zero is acceptable. Record calculations and assumptions so they can be rerun.
8. Challenge the inference. Identify alternative explanations, confounding, missing comparisons, uncertainty and selection bias. Optional AI suggestions are unverified drafts. A human accepts, rejects or revises each material suggestion with reasons. Retain unresolved objections.
9. Review and decide. A named researcher checks the original evidence and calculations and records review scope, date and conflicts. Obtain a separate qualified reviewer for material domain or methods claims where feasible; otherwise state that independent review is missing. Record the next test and decision rule. A checkbox does not confer scientific verification.
10. Release and maintain. Only an authorized human may release a dated version after reviewing source rights, privacy, evidence and unresolved limitations. Keep earlier versions, corrections, authorship and a stable link. Record the actual publication date only after release. Specify an achievable review schedule and owner; do not promise continuous updates. Invite a replication attempt and record disagreements and failures.

When to stop or change course

• No bounded question → narrow it before collection.
• Source unavailable or original unread → request the material and label the record “Needs source inspection”; do not infer findings from metadata.
• Same study across providers → one underlying evidence unit, multiple discovery paths.
• Incompatible geography/time/units → do not combine until a justified transformation exists.
• Contradictory evidence → retain both; propose a discriminating check rather than vote by provider count.
• Insufficient evidence → report “No supported conclusion yet.”
• No independent reviewer → disclose the missing review; never relabel self-review as independent validation.
• Publication not performed → label a draft, not a published finding.

Choose sources for the question

OpenAlex and PubMed help discover literature. Crossref helps resolve publication identifiers and metadata. They can describe the same study. World Bank indicators and OpenAQ measurements answer different questions at different geographic and time scales. Consider all five, but record a reason when a provider is not applicable.

For a proposed Atlanta or DeKalb monitoring study, start with EPA AQS and AirData. Check monitor operators, pollutants, active periods and completeness. An OpenAQ gap alone does not establish absence of a regulatory-network monitor.

For local poverty estimates, inspect Census ACS data at the appropriate geography and vintage, including uncertainty. Do not assume a national indicator describes an Atlanta neighborhood. Monitoring location is not itself a measure of personal exposure or health effects.

Provider documentation: PubMed · Crossref · OpenAQ.

AI assistance and publication

The workspace prepares an AI handoff prompt and a nine-section worksheet from supplied materials. A model's self-critique is not independent review. No claim becomes verified because several databases repeat it, a user checks a box, or a method is described transparently.

AI suggestions remain drafts. A researcher checks original evidence, calculations and reasoning before deciding what to release. Record an actual publication date only after release; keep earlier versions and unresolved objections visible.

Test whether the method helps

Compare matched tasks with researchers’ existing process. Record missed relevant evidence, citation errors, next-test quality, total researcher and reviewer time, and reviewer disagreement. Define the evaluation before collecting results and report failures as well as improvements.

No time savings, research effectiveness score or completed finding is claimed. Read the Funder & Reviewer Brief.