Fraud detection & response quality rules
Every response is scored in near real time at both the question and session level. Behavioral telemetry is captured silently while a respondent answers, run through deterministic rules, then reviewed by an AI auditor. Scores roll up to the admin dashboard for triage.
Behavioral signals captured
Milliseconds from question render to submit, compared against an expected reading time derived from prompt length.
Keystroke counts per open-end answer, used to catch text that appears without matching typing activity.
Clipboard pastes captured per question — a common signature of scripted or recycled answers.
Window blur / tab-switch counts while a question is open, indicating multitasking or automation.
User agent, platform, screen size, timezone, language and connection reputation (VPN / proxy / hosting).
Choice position history, response variance, and open-end text passed to gibberish, low-effort and duplicate checks.
Question-level rules
| Signal | What it means | Weight |
|---|---|---|
Open-end looks like a keyboard mash gibberish | Vowel ratio, character runs and keyboard-row sequences flag mashed text. | 40 |
Same option position picked repeatedly straightlining | Three or more consecutive choice questions answered in the identical position. | 32 |
Answered faster than the prompt could be read speeding | Submitted in under 40% of the expected reading time for the question. | 30 |
Text appeared without matching typing activity no_typing | Answer longer than 25 characters with keystrokes below 40% of its length. | 25 |
Repeats an earlier open-end answer verbatim duplicate_text | Exact match against any previous open-end response in the same session. | 25 |
Answer was pasted in pasted | One or more clipboard paste events during the open-end question. | 20 |
Throwaway open-end answer low_effort | Matches a dictionary of non-answers such as 'na', 'idk', 'good', 'nothing'. | 18 |
Always selects the first option first_option_bias | First position chosen across every recent choice question. | 15 |
Unusually fast answer fast | Submitted in under 75% of the expected reading time. | 12 |
Single-token open-end answer too_short | One word and fewer than four characters. | 12 |
Left the survey window repeatedly distracted | Three or more focus losses while the question was open. | 10 |
Very long pause before answering stalled | More than three minutes on a single question. | 8 |
Session-level rules
| Signal | What it means | Weight |
|---|---|---|
Majority of answers were rushed session_speeder | Over half of all answers carried a speeding or fast flag. | 35 |
Zero variance in choice selections no_variance | Every choice question answered in the same position. | 30 |
Completed faster than humanly plausible impossible_completion | Total completion time below the floor for the survey length. | 30 |
Median time per question under 1.5s low_median_time | Session median response time falls below a human reading threshold. | 25 |
Straightlining detected across choice questions session_straightlining | Repeated positional patterns across multiple questions. | 25 |
Connection looks like a VPN, proxy or hosting provider proxy | IP reputation indicates a datacenter or anonymizing network. | 25 |
Nonsense open-end answers session_gibberish | Scales with the number of gibberish open-ends in the session. | 20 + 8 each |
Repeated identical open-end answers session_duplicate_text | Same text reused across multiple open-end questions. | 20 |
Many rushed answers session_fast | Over a quarter of answers flagged as fast (when not already a speeder). | 18 |
Risk scoring
Triggered rule weights are summed and clamped to a 0–100 risk score. Question scores feed the session aggregate, which is bucketed into a risk level.
AI-assisted auditing
Open-end answers are reviewed by a language model for relevance, coherence and off-topic or machine-generated content. A second pass summarises the whole session, explaining in plain language why a respondent looks suspicious.
Privacy posture
Only behavioral metadata and answer content needed for scoring are stored. Respondents are told upfront that response behavior is tracked for fraud detection and usability analysis.