Define the batch
Create a batch identifier tied to the exam, subject, section and date. Record the expected student and page count before scanning.
Keep absent students and extra answer booklets explicit; do not infer them from missing uploads.
Scan consistently
Use the same orientation, resolution and file format across the batch. Check the first few booklets before scanning the rest.
Preserve the source PDF and avoid sending personal student data to unapproved tools.
Validate student and page mapping
Confirm that every page belongs to the correct student and is in order. Detect duplicate scans, blank pages and unexpected page counts.
Mapping errors should stop evaluation for the affected booklet and create a visible exception.
Freeze the evaluation criteria
Finalize question maxima, model answers, accepted alternatives and partial-credit rules before processing. Version the rubric if changes are unavoidable.
Reprocess all affected answers when a scoring rule changes; do not mix outputs from different criteria silently.
Process and monitor
Track uploaded, extracted, evaluated, flagged, reviewed and failed counts. A progress bar without these categories can hide incomplete work.
Set a timeout or escalation rule for pages that cannot be processed reliably.
Review and reconcile
Review exceptions, perform the required sample or full check, and compare totals with maximum marks. Confirm that approved student count matches the roster.
Export only final, reviewed records. Keep failed or pending students out of the final marksheet until resolved.
Document the run
Record scanner setup, software version, rubric version, reviewers, correction rate and known issues. This makes later moderation and improvement possible.
A successful pilot is not only fast; it is complete, traceable and acceptable to faculty.
Go-live checklist
Before a live batch, verify user access, storage location, scanner settings, question maxima, rubric version, expected roster, exception handling and escalation contacts. Run one complete booklet from upload to export and compare it with the original paper.
During processing, reconcile counts at every stage. After review, lock or version the approved output so later changes are visible. Keep the original scans and audit records only for the retention period permitted by institutional policy.
Common failure patterns
Watch for two students sharing one identifier, a continued answer being assigned to the next question, blank pages counted as failures, reordered scan pages and rubric changes that affect only part of a batch. These problems are operational rather than purely “AI” errors, so workflow controls are as important as model quality.