What SizeMemo calculates, and what it doesn't.

Size matching is deterministic

For each size and each point of measure the engine computes ease: garment measurement minus body measurement.

Ease is compared with a target range that depends on garment category and the shopper's fit preference. Inside the range costs nothing. Outside it, the penalty is the distance in centimetres × the point's weight. Too-tight distances are multiplied by 1.5, because a garment that is too small can't be worn, while one that's slightly loose usually can.

The size with the lowest total penalty is suggested. When two adjacent sizes score within 1.5 points, the shopper is told they're between sizes. Stretch fabrics allow 3 cm less ease at circumference points.

The same inputs always give the same output. Amazon Bedrock is used afterwards only to phrase the explanation; it never picks or changes a size.

AI interprets feedback

Return comments are unstructured. A Claude model on Amazon Bedrock assigns each one a category (too small, too large, too long, too short, unexpected fit or not fit-related), the body area mentioned, a confidence score and a one-line rationale.

Low-confidence results are queued for review, and a person's correction always overrides the model. If Bedrock is unavailable, a transparent keyword classifier is used and labelled as such.

Consistency needs two signals

A product is flagged as a potential sizing inconsistency only when its measurements sit outside the brand block's tolerance in two or more sizes and classified feedback complains in the matching direction and body area.

One signal alone is reported separately: off-spec measurements may be a deliberate design choice, and feedback alone may point to fabric or styling.