Map each atomic claim to evidence, contradiction, source quality, and a calibrated status
8 min setupTested with:ChatGPTReviewed: 2026-08-28
1
Add your context
Your text stays in this browser. AILesson Prompts does not send it to a model or server.
2
Your prompt
Unfilled fields remain visible as placeholders, so you can still copy and edit the prompt
Audit the evidence status of every claim against the supplied corpus.
Claims or document:
[claims]
Sources with IDs, dates, authors, methods, and passages:
[evidence]
Intended use, status labels, quality rules, currentness, and burden of proof:
[standard]
Split compound sentences into atomic claims and preserve their exact scope, modality, comparison, time, and attribution. For each claim return type, supporting sources, contradicting sources, directness, source quality and independence, recency, population/measure fit, limitations, and calibrated status using only the supplied label rules. Distinguish no evidence supplied from evidence of absence. Do not count repeated reports of one underlying source as independent corroboration, infer causality from association, or upgrade an abstract, headline, opinion, or vendor claim beyond its evidence. Flag circular citations and source conflicts. Finish with corrected wording for overstated claims, missing evidence priorities, and claims that should be removed for the intended use.
Private by defaultPrompt assembly happens locally in your browser. Avoid placing confidential information into any AI service unless your organization allows it.
From input to outcome
A worked example
See how concrete context turns this recipe into a usable result
Actual input
Claims to audit
C1: The four-day-week trial increased productivity by 20% and improved wellbeing for all employees. C2: It caused no customer-service decline. C3: Most organizations can adopt it without additional staffing.
Evidence corpus
S1 (2025 company report): one 80-person software company compared self-reported completed tasks during a 12-week voluntary pilot with the prior quarter; reported +20%; seasonal workload changed; no control group. Wellbeing survey: 61/80 responded, 49 reported improvement, 8 no change, 4 worse. S2 (2025 company support dashboard): median response 3.9h before and 4.1h during; target ≤4h; complaint count not supplied. S3 and S4 are news articles that both summarize S1 without new data. S5 (2024 cross-sector survey abstract): 38 organizations reported varied staffing changes; full methods unavailable; abstract does not report adoption feasibility.
Evidence standard and use
For a public guidance article. Labels: Supported, Partly supported, Unsupported, Contradicted, Unclear. Causal language needs a credible comparison; 'all/most' needs population evidence. Company reports may support observations about that pilot but not generalization. Evidence current through 31 December 2025.
Example output
C1a “trial increased productivity” | causal | S1 reports +20% self-reported completed tasks | no independent source; prior-quarter comparison, seasonal change, voluntary pilot, no control | Partly supported only as an observation, not causal. Corrected: “The company reported 20% more self-reported completed tasks during its 12-week pilot than in the prior quarter; the design cannot establish causation.” S3/S4 are not independent.
C1b “improved wellbeing for all employees” | universal | S1: 49 of 61 respondents improved, 8 unchanged, 4 worse; 19 non-respondents | Contradicted. Corrected: “49 of 61 survey respondents reported improvement.”
C2a “no customer-service decline” | universal/absence | S2 shows median response changed 3.9→4.1h and exceeded ≤4h target; other service measures absent | Contradicted for the target claim and unsupported as a universal statement. Corrected: “Median response time rose from 3.9 to 4.1 hours; other service effects were not reported.”
C3 “most organizations can adopt without staffing” | generalization | S5 mentions varied staffing changes but no feasibility result | Unsupported. Remove from public guidance until full multi-organization staffing evidence is available.
Priorities: obtain S1 definitions/raw task distribution, comparable control/seasonality analysis, complete service metrics, and S5 full methods/results.
Why this works
1
Atomic claims prevent mixed-status sentences from receiving one misleading evidence label.
2
Source independence and fit checks prevent citation count from substituting for evidential strength.
Check the result
Was every compound statement split without changing scope or certainty?
Are support, contradiction, absence, and missing evidence distinguished?
Does each status account for source quality, independence, recency, and population fit?
ChatGPT