A release-ready research brief has a traceable input set and a restrained output contract. Store the question, search or supplied-corpus snapshot, eligibility rules, included and excluded study IDs with reasons, extraction records, calculation script and output, unresolved conflicts, review notes, and final text revision. Before release, verify every number and design label against its report anchor, and check that the brief includes contrary findings and missing evidence. Use a label matching the actual process: a limited evidence packet is a limited evidence packet. Do not call it an exhaustive or systematic review if discovery and screening were not performed to that standard. If the brief could drive a consequential operational decision, route it to the accountable reviewer rather than turning the generated recommendation into an automatic effect.
Research synthesis release: expose the method and its limits
Operational case
The parcel operator's final packet contains ST-47 and ST-88 as controlled line trials and ST-61 as context. It reports the observed -7.5 and +2.5 point differences separately. The first draft says the scan panel 'has been proven to reduce misroutes' and omits ST-88. The release gate blocks it. The corrected brief says the two controlled results disagree, the changed-shift report cannot settle that disagreement, and additional line trials with specified allocation and missing-outcome reporting would narrow uncertainty. An operations owner reviews any proposed rollout. The published brief's dataset revision, script output, and text hash match the reviewed packet.
Input: frozen packet RS-214; include ST-47, ST-88; context ST-61.
Checked arithmetic: -7.5 and +2.5 percentage points.
Final claim: controlled results disagree; no pooled effect claimed.
Gate: contrary result present, exclusions logged, owner review recorded.
Release: reviewed text and input revision match distributed artifact.Performance and operating cost
For K included studies and C public claims, a direct evidence audit is O(K+C) when records and anchors are indexed, plus the time needed to inspect disputed passages. Re-running a deterministic calculation is cheap compared with locating a changed report or adjudicating an exclusion. A frozen input set makes revisions reproducible; a new search date requires a new packet and review. If the model rewrites the summary after approval, compare the exact new text against the approved version rather than assuming the earlier check carries forward.
Common Mistakes
- Do not describe a limited supplied packet as an exhaustive review.
- Do not release a summary that hides a contrary trial.
- Do not automate a rollout from an unreviewed generated synthesis.
Connected lessons
- Production prompt engineering
- Prompt Engineering
- Prompt release artifacts: version the whole decision path
- Research prompts: bind claims to a date and evidence record
- High-impact advice: route uncertain decisions to qualified review
- Research prompts: fix the question and eligibility first
- Study extraction prompts: preserve the result behind each claim
- Study effects: calculate the measure outside the model
- Conflicting studies: inspect design before averaging outcomes
- Project: synthesize conflicting parcel-scan trials
- Research-synthesis prompt decisions
Continue with: Interview prompts: turn bounded findings into testable decisions.