Preserve evidence before choosing labels
The first design decision in review analysis is how to handle evidence, not what to call the categories. A smooth summary can erase conditions and uncertainty. I recommend separating the source wording from its assigned labels so a reviewer can inspect the basis for a classification. Preserving evidence does not mean retaining unnecessary personal information: exclude identifying details before importing notes.
Fitnote currently processes CSV notes using rule-based logic in the browser. Matching defined expressions and conditions should not be presented as AI understanding. Explain what the rules detect and what they cannot decide. Useful classification does not, by itself, demonstrate human-level comprehension.
Negation changes the evidence
Finding the word “small” is not enough to classify a complaint about small sizing. “It is not small” and “I thought it was small, but it fits after trying it on” express different judgments. Negation, contrast, and changes over time deserve review rather than a promise that rules resolve every sentence.
All quoted examples in this essay are invented illustrations, not actual reviews or analysis results. “It is not short, but it rides up when I move” cannot safely be reduced to a length complaint. The movement can be retained as an observation, while its cause remains unresolved. Review the wording and either defer a narrow label or use a broader review category.
Do not force ambiguity into sentiment
“Not what I expected” signals a difference from expectations without naming an attribute. “It is okay, but I am unsure” does not provide a precise product assessment. Filling these gaps with size, material, or satisfaction labels produces a tidier table but weaker evidence. Unknown and needs-review states describe the limits of the source.
Reviewers should separate what the sentence says from what they infer. Existing product knowledge must not be treated as if it appeared in the note. Keep unanswered questions distinct, and do not elevate an opinion with missing context into a representative customer requirement.
- Do not infer an attribute when none is named.
- Keep conflicting statements visible.
- Distinguish missing information from negative sentiment.
Overlapping labels change the meaning of totals
An invented note such as “The sleeves are long and the body is narrow” can support both sleeve-length and body-width labels. Adding label counts may therefore exceed the number of classified rows. That is a consequence of multi-label counting, not necessarily an error. It is different from importing the same row twice.
State the counting unit and whether multiple labels are allowed. Input rows, unique notes, and label occurrences are different quantities. Shares using the same denominator can sum above the whole when categories overlap. Mutually exclusive categories require an explicit priority rule, and reviewers should consider which secondary observations that rule suppresses.
Exclude personal information and retain human review
Before importing a CSV, exclude unnecessary names, telephone numbers, email addresses, addresses, and order identifiers. Free text can also contain identifying details, so checking column headings is insufficient. Browser processing alone does not resolve every privacy concern or guarantee complete security. Start with the minimum information needed for the analysis.
Human review should cover negation, vague subjects, multiple topics, and missing sources, as well as apparently straightforward classifications. When changing a judgment, distinguish the original label from the correction and its reason. Check whether a revised rule damages other interpretations.
In the shared output, separate verified observations, interpretations, and unanswered questions. Suggested wording should be a human-reviewed checklist or template grounded in source notes and verified product facts, not described as AI-generated copy. Evidence-first analysis does not make every statement certain. It makes the remaining uncertainty inspectable.