Project: audit a receipt-image recapture gate

Prepare the evidence

Create a synthetic image manifest with submission groups, device channel, capture time, consent scope and readable, unreadable or uncertain labels. Include two crops of one receipt, a withdrawn image and a later device type. State exactly when a user is asked to recapture. The task contract must precede model choice.

Implement the path

Validate image size and orientation, apply deterministic preprocessing, and store reversible geometry for any box overlay. Partition by submission, then use training-only augmentation that preserves labels. A simple baseline is acceptable; the project is judged on data and decision integrity rather than model complexity. The later device stays untouched for a shift check.

Measure failures

Report confusion counts and the threshold cost of missed unreadable images versus unnecessary recaptures. If localization is attempted, show one-to-one box matching and IoU. Slice results by device and glare condition with denominators. An uncertain label should be adjudicated or excluded under a declared rule, never silently treated as readable.

Submit a release packet

Provide the synthetic fixture, label rubric, preprocessing manifest, split IDs, exact metric calculation, representative false positives and negatives, and rollback decision. Assert that the withdrawn image is absent from training and serving caches. Include a privacy deletion test and one manual-review route for low-confidence cases.

Implementation

def recapture_decision(unreadable_probability, threshold=0.63):
    if not 0 <= unreadable_probability <= 1 or not 0 < threshold < 1:
        raise ValueError("invalid probability or threshold")
    return "request_recapture" if unreadable_probability >= threshold else "continue_review"

assert recapture_decision(0.71) == "request_recapture"

Performance and operating cost

A threshold decision is O(1) after inference; image decoding and model inference dominate latency and memory. The release packet must measure those costs on the intended device and image-size distribution before this synthetic fixture becomes an operational path.

Common Mistakes

Do not split crops of one receipt across train and test.
Do not retain withdrawn images in an augmentation cache.
Do not ship an accuracy figure without threshold-specific error counts.