How validation works
A report is only worth as much as the judgement behind the feedback in it. This is what happens to every submission.
1. Every submission is scored, not sampled
Each item is evaluated by the validation engine before it can be approved. Nothing reaches a report without passing through it.
2. Scored against its own product’s history
A submission is compared with the distribution of prior submissions for that product and channel, not against a global average. A product with polarised feedback and one with uniform feedback are judged on their own terms.
3. The method is chosen per product category
Z-score, interquartile range, isolation forest and embedding distance are all available, selected per category rather than fixed platform-wide.
4. Outliers are flagged, never auto-rejected
A statistically unusual submission goes to a human review queue. An unusual opinion is not a false one, and discarding it automatically would bias the corpus toward the consensus it is supposed to measure.
5. Near-duplicates are detected by similarity, not string matching
Submissions are embedded and compared by vector similarity, so templated or lightly-reworded repeats surface even when no two are textually identical.
6. Every decision stays reconstructable
Each result records the method used, the score, and a snapshot of the thresholds in force at that moment — so a decision can be explained years later, after the thresholds have moved.
7. Reports carry no personal data
Names, emails and phone numbers are removed where a report is served, not where it is generated. The masking sits at the boundary every consumer crosses.
Feedback is processed in India, in-region, and is never sent abroad for analysis.