RWE Tools
Protocol Checker
Upload a real-world-evidence protocol and check it, item by item, against HARPER (the ISPE/ISPOR protocol template) — or TARGET for target-trial emulations. You also get an auto-drawn study-design scheme and a review of the planned shell tables and figures. The flip side of the Protocol Generator: audit a protocol you already have.
Your protocol
or paste the text instead
ⓘ Checked by a free AI model (Google Gemini). Text is sent to Google for analysis and is not stored by this site, but please don't upload confidential or unpublished-sensitive material. Results are AI-generated guidance to help self-audit a protocol's completeness — not a substitute for the official guideline or peer review.
The check could not be completed
Nothing was scored. A partial or unreadable answer is never turned into a conformance percentage.
⚠ This report is out of date
Nothing below has been re-checked, and both download buttons still export this report — the file will say so. Press Check protocol again to score what is in the input now.
This looks like a target-trial emulation.
Conformance
How to read this. Met = the protocol says it, substantively. Partial = the topic is raised but a key element (usually an operational definition, code list, or window) is missing. Missing = no trace in the text. N/A = the item cannot apply to this kind of document. Not assessed = the AI returned no usable verdict, so nothing is claimed either way. Not scored = the AI did answer, but the row is a heading for sub-items counted separately, so counting it too would count the same content twice. All three of N/A, Not assessed and Not scored are excluded from the score, so the denominator can be smaller than the checklist — the sentence beside the score says when it is, and by how much.
Neither HARPER nor TARGET defines a conformance score. This ratio is this tool's own summary of the item-level verdicts below; the guidelines are checklists, not scales, and a percentage is not what either was written to produce. The verdicts are the output — the number is a way of sorting them.
This score is one sample, not a measurement. The same protocol, submitted four times unchanged, scored 19, 20, 21 and 22 of the 23 HARPER items in our own testing — the model is not deterministic. Treat a one- or two-item difference as noise, read the per-item evidence rather than the total, and re-run if a verdict surprises you.
Study design scheme
This diagram was drawn by this tool from your text. It is not evidence that your protocol contains a study-design diagram — that is item 7.2, judged separately below. A window the AI declared it had assumed is drawn dashed and marked *; a window it could not place on a day axis is not drawn and is named under the diagram. The AI does not always declare an assumption, so a solid bar is not a guarantee that the duration was stated in your protocol — check every window against it.
Timeline (days relative to index date)
Planned outputs & figures
Does the protocol pre-specify the shell tables and figures a rigorous RWE study mocks up in advance — Table 1, a Love plot, a participant flow diagram, results shells, and so on?
What the five labels mean. Present / Partial / Absent are claims about your document: whether it pre-specifies this output — or the equivalent your design owes, since these seven names are written for a person-level, propensity-score-adjusted, time-to-event cohort. A case-selection flow counts as the flow diagram; a case-series or period-composition description counts as the Table 1; an incidence-rate-ratio profile by risk window counts as the survival curve. N/A is a claim about your design: its estimator cannot produce this output or any equivalent — there is no between-person balance to diagnose in a within-person design, and no survival curve where nobody is followed from a shared time origin. Two rows can never be N/A and are recorded as Absent if the AI says otherwise — the primary results shell and the sensitivity-analysis outputs, because every design has a primary estimate, and assumptions worth probing: identifying assumptions where a contrast is estimated, and measurement assumptions — outcome and denominator misclassification — where none is. Two of the five labels can be this tool’s rather than the AI’s: Not assessed, which is always this tool recording that the AI gave no verdict or gave one it does not recognise, and an Absent on one of those two barred rows where the AI had said N/A, which is this tool overruling it. An ordinary Absent on those rows is the AI’s own verdict like any other. Every label this tool supplied rather than read says so in its own evidence line. Read the evidence line before acting on any of them.