Semantic QC is live — 41% more invalid open-ends caught before they reach your dataRead the method
Data quality
The most expensive line in a study is the rework after a decision was made on bad data. Quality control decides at submit, not at the debrief.
Three gates
The three layers run independently and any failure stops the complete. Every verdict and its reason is retained, so you can audit it — or appeal it.
Cross-referenced identity, mobile and employment data, plus device, payment-binding and history linkage to surface multi-account users, farms and paid-answering rings before they enter a study.
Question-level timing, pointer and scroll traces, answer order and consistency modelling. Straightlining, speed-running and click-without-reading all surface here.
Open-ends are reviewed for AI cadence, topic drift, stem restatement and zero-information answers — and you get a readable reason, not an opaque score.
Reject at submit
The conventional workflow runs a cleaning script after collection — by which point bad completes have already consumed quota and budget. Quality control decides at submit: rejected responses free the quota instantly and recruitment refills it.
Rules are public
Quality control should not be a black box. These are the rules currently in force — no hidden clauses, and no "the system determined the response was invalid".
| Rule | Trigger | Action | Billing |
|---|---|---|---|
| Device and account linkage | Same fingerprint across 3+ accounts, or a match against a known farm registry | All linked accounts frozen for review | Not billed |
| Timing anomaly | Any question under 1.5s, or whole interview under 40% of median | Auto review plus human spot check | Billed once corrected |
| Straightlining | Same option across 5+ matrix rows with zero variance | Rejected | Not billed |
| Semantically empty open-end | Over 0.7 overlap with the stem, AI cadence, no information | Rejected with reason recorded | Not billed |
| Contradictory answers | Internal conflict, e.g. stated role incompatible with tenure | Re-screened, rejected if unresolved | Not billed |
| Supplier quality drift | 90-day rejection rate 1.5× the mean | Downgraded, throttled or removed | — |
Deliverable
Not a screenshot claiming the file was cleaned, but a document you can hand to a colleague: where the sample came from, how much was lost, why, and what is still trustworthy.
1,321 received · 1,284 valid
37 rejected (2.8%), against a 4.6% category benchmark
Most rejections came from semantics (21)
Then straightlining (9) and internal contradictions (7)
Median interview 6 min 12 s
Normal distribution at question level, no speed clusters
All 16 quota cells met
Highest cell oversampled 4%, weighted down
Verdict: fit for quantitative brand tracking
Recommend reading 10% of open-ends before quoting them
Start your first project
$500 in sample credit on your first order. No minimums, no contracts, no platform seat fees. If you would rather have someone run it with you, one of our researchers will field the whole study.
No credit card 14-day refund window Export to SPSS / Excel / CSV
QUALITY FAQ
Not covered here? Email support@pandaopinion.com — you will hear back within two hours during business hours.
Talk to usThe semantic layer looks for information density: answers naming a store, a price, a moment or a specific reason pass. Only boilerplate, stem restatement and suspiciously smooth but empty prose get rejected. We calibrated on 12,000 human-labelled responses and keep the false-positive rate on genuine respondents under 0.7%. Every rejection keeps the original text so you can appeal.
Yes. There are three preset thresholds plus custom rule combinations, for example rejecting anything under three minutes. Stricter settings raise CPI and slow down fielding; the dashboard shows the projected impact as you change them.
Rejection is the last step, not the first. The identity and behaviour layers stop most ineligible accounts before they ever enter a project, so they never consume quota or budget. A low rejection rate reflects earlier filtering, not a lax gate.
Yes. You can export a record-level file including question timing, device type, source supplier and rejection flags, so your own rules can be applied on top. We do not lock the data.
