Correct a rejected AI reply
The service identifies which finding or recommendation failed validation. Saved advice is unchanged.
How to use it
- Download or copy the corrections below the error. Send them to the same chat with the original assignment.
- For an image error, the service identifies the field and thumbnail token: the image may exist in the assignment but not be selected for this advice. AI must link every mentioned image, including those kept unchanged, to the original facts. Do not guess mappings from similar names.
- Obtain a complete corrected marpla-answer.json and upload it again. Do not refresh data merely to fix an error: that replaces the original assignment. Shortened metric names are recognized only by an exact match to daily data in the assignment. A specific product, set or device without a matching metric requires manual verification.
- For fresh data, use the sync icon beside Whole-game advice, then download the new assignment in step one and obtain a new AI reply. Downloading again without syncing preserves the existing assignment.
Advice and practical context
A correction preserves the original context
Validation ties findings to one assignment. Correct the complete reply for that assignment instead of creating a new data package.
Example of one failed recommendation
Example: when an error names a recommendation and field, send the correction file to the AI and obtain a complete JSON reply; a fragment cannot replace the response.
The correction file contains the rejected reply and field-specific instructions, so send it in full to the original chat. Request a complete new marpla-answer.json even when one recommendation failed because the importer validates the whole response. If the assignment expired or fresher evidence is needed, create a new assignment and do not combine it with corrections for the previous version.
Do not refresh without a reason
A new sync is for fresh evidence and makes the old reply obsolete. Decide whether you are repairing format or starting another analysis.
Technical detailsCalculations, permissions and behavior
When an AI reply file is rejected
MARPLA checks JSON, references to the original data and whether result verification is supported. A rejection shows specific reasons with finding or recommendation numbers; none of an invalid reply's advice is saved. Select Download corrections for AI or Copy corrections and send them to the same chat, together with the original assignment if it is no longer available there.
Corrections contain the rejected reply and exact fields to fix. Obtain a complete new marpla-answer.json and upload it in step three. You do not need to rebuild data to correct a reply.
If a newer assignment has already been prepared, the previous reply can still be imported within 7 days of its creation: MARPLA validates it against the original assignment. An older reply cannot replace a newer saved analysis. The import confirmation shows the date of the assignment used for the advice.
The first step contains only the assignment download button. Each download checks sources and retrieves available fresh data. Sources are queued in batches and checked concurrently with bounded Roblox traffic.
A temporary failure of one metric does not delay checks of other sources; completed checks are retained. The report window is fixed at startup and does not shift across an hour or day boundary. Temporary connection failures retry the same request automatically; if the connection remains unavailable, another click resumes the saved check.
Refreshing the advice panel does not delay the completed file. If Roblox rejects a specific metric query, the service continues checking other sources and explicitly lists unavailable metrics in the file. They are not zero values: AI must not use their old values for new conclusions.
If a breakdown by custom event name is rejected, MARPLA attempts to discover event names and query each separately for the same period. Collection continues in the background and saves successful parts. The whole source is checked only after every event completes.
An average cannot be reconstructed without the original sum and event participant count. Access errors and temporary failures are not counted as successful checks. The sync icon beside Whole-game advice also refreshes the assignment, without downloading a file.
Updating the service itself still allows correction of an old reply within the assignment's validity period. For fresh data and new features, including inline thumbnails, prepare a new assignment and obtain a new reply. If AI omits the metric: prefix in a monitor, the service restores it only for an exact daily metric in the original assignment.
Evidence references and all other checks remain required. Product, set or device checks without matching data require manual verification; whole-game metrics cannot replace it. Thumbnails must appear beside advice or its steps, not only in technical details.
MARPLA never silently substitutes metrics or changes the AI's confidence. The assignment contains the response schema, relationship checks and ready-to-copy campaign, group and image references. A campaign without a thumbnail still has an identity reference that AI must include.
Before delivering a file, AI must check references, required fields and limits as well as JSON syntax. During import MARPLA automatically adds missing evidence references to metrics, products and visuals already selected in the reply, when those facts exist in the original assignment and fit within the limit. Thumbnails placed only in technical details also appear in the main advice when there is room.
Missing presentation references for unambiguously cited imageless campaigns are restored in non-comparison advice. Advice without a verifiable setting change retains manual result verification. Prices, conclusions, confidence, data and user decisions remain unchanged.
All other checks still apply; substantive errors reject the file with an explanation. No instructions can eliminate every external AI error. Updated instructions are included when downloading an assignment, including a previously prepared assignment downloaded again.
Advice initially appears as collapsed headings. Open one to collapse the previous card while retaining its draft notes. After saving Done or Not doing, advice moves to the archive with your response.
Archived headings show the saved decision: Done, Not doing, or No decision if none was made. Recommendations superseded by a new analysis also appear in the archive. New advice on the same topic receives its own decision; an earlier rejection does not hide it.
AI receives saved rejection reasons and results. The absence of repeated advice does not prove that an issue is resolved.