How does the Lab interpret a question?
The builder looks for authored topic terms and phrases in your question. Specific phrases generally count more than single words. It selects up to 5 topics and reports recognized topics left out. Choose your own topics if it misses a synonym or part of a compound question. Search the site if the topic is missing from the library.
The answer formats change what the brief emphasizes. A study design opens an evaluation plan, a comparison highlights the three workflows, and an evidence review leads with source notes. Depth controls how much analysis is displayed. Full research memos include exercise conditions, stop criteria and questions to resolve before a pilot.
The topic notes and connections are authored. A current officeholder, legal clearance, live commodity availability or country forecast requires other evidence.
How do unresolved cases affect the answer?
Use one fixed binary outcome and time window in all approaches. Within each approach, let N be eligible cases, S confirmed successes and U unresolved. Confirmed negatives equal N − S − U. Don't reclassify an unresolved case as a failure just because it was absent from the records.
- Rate among classified cases: S ÷ (N − U), if any are classified.
- Lowest possible rate: S ÷ N, if none of the unresolved cases succeeded.
- Highest possible rate: (S + U) ÷ N, if they all succeeded.
- Difference between workflows: later lower − earlier upper to later upper − earlier lower, in percentage points.
Worked example: verification by a deadline
In a fictional exercise, current practice verifies 58 of 100 cases and leaves 8 unresolved, for a possible range of 58% to 66%. Structured human-led cooperation verifies 72 of 100 with 5 unresolved, so its range is 72% to 77%. The same process with AI verifies 78 of 100 with 10 unresolved, for 78% to 88%.
Organized cooperation versus current practice ranges from 6 to 19 percentage points. The AI step versus organized cooperation ranges from 1 to 16 points. If the test required a gain of at least 5 points, the first contrast clears it throughout the range. The AI contrast does not settle the question. A situation like this calls for completing the audit or defining a replicable next test.
The ranges place all unresolved cases at either extreme. They describe how missing outcomes could affect the result; they are not confidence intervals or causal estimates. Compare the same outcome, eligibility rules and period in all arms. Missing-outcome bounds do not correct eligibility differences, inconsistent measurement or unobserved confounding. Record whether any theoretical extremes are ruled out by verified facts.
What should a reviewer be able to reproduce?
- The exact original question or decision, selected topics, software version and source notes.
- The success rule, denominator, eligibility and cutoff, counts and unresolved reasons in each approach.
- The comparable workflow, data, review rules and resource assumptions held fixed, plus any deviations.
- The criteria for a follow-up, review session or rejection even when a numerical target appears met.
JSON includes the versioned scope, full authored topic records, selected source notes, and the count inputs and calculations if they were run. Markdown includes the readable brief. Review exports before sharing. Nothing you type is submitted to a server by the builder.
Read the evaluation protocol Read the measurement guide Read the interaction guide
Brief version 1.0.0. Last reviewed 15 November 2026.