Quality Control

As responses come in, a QC pipeline scores each submission against a battery of heuristics. Straight-lining detection inspects multi-item Likert and Matrix blocks and flags responses where a respondent picked the same value for every item, optionally with a tolerance for one-or-two deviations to avoid false positives on grids where the same answer is genuinely valid. Speed outlier detection captures the time spent on each question and the survey as a whole; submissions that complete in less than 10% of the median total duration — calibrated per survey rather than globally — are flagged as likely speeders.

Drop-off analysis records the last interacted question and the action that preceded abandonment (closed tab, navigated away, idle timeout) so authors can see exactly where attention is lost. The dashboard renders a funnel from invitation to first answer to each subsequent question to completion, with conversion rates per step and side-by-side comparison across segments to surface design problems such as a confusing matrix or an over-long open-ended block. Every submission carries a quality_flags array combining these signals plus optional custom flags from validators (impossible answers, contradictory branches, suspected bot patterns when Turnstile risk score is elevated).

Analytics queries and exports accept a quality filter — strict (no flags), permissive (timing flags allowed), or raw — so the same dataset can serve a board-ready report and a sensitivity analysis without forking the data. Flags are advisory rather than destructive: no response is auto-deleted, and authors can review flagged submissions individually with the full timeline (timestamps, branches taken, edits) before deciding to keep, weight, or exclude them.