The automated checks.
Twenty-one automated checks run on every response. All of them, with their weights, are open for inspection under Signal weights on the Respondent Risk tab; the groups below mirror that panel exactly. Each weight is the check's ceiling: graded checks award a share of the weight by severity, so a mild speeder scores far less than an extreme one.
| Check | Weight | What it catches |
|---|---|---|
| Traps & rules | ||
| Failed a trap | 90 | Claimed a decoy brand, failed an attention check, or tripped a logic rule set on the Traps tab. Effectively conclusive. |
| Quality-fail tag | 90 | A hand-applied quality tag from the survey team. A human verdict, so effectively conclusive. |
| Logical impossibility | 45 | Age contradicts birth year, an allocation grid that does not sum to 100, a number outside anything else in the sample. |
| Speeders | ||
| Speeder | 70 | Finished far under the median duration. Graded: the further below the threshold, the more of the weight they score. |
| Skimmer | 70 | Too little time per question actually answered. Catches racers on long paths whose total time looks normal. |
| OE checks | ||
| Bad open-ends | 70 | Scored on the share of a respondent's text answers carrying red flags, verified on the OE Checks tab. |
| OE duplicate (corroborated) | 60 | Shares a written answer with someone AND agrees with them beyond chance elsewhere. A lone shared answer is only an amber OE flag. |
| Straightlining | ||
| Straightlining | 40 | Flat, patterned or diagonal answering on grids. Low-differentiation grids are soft yellows, since using two scale points is often honest. |
| Duplicates | ||
| Device duplicate | 70 | Same device fingerprint twice in one wave. The first entry keeps a soft 10 as the original; later entries take the full weight. |
| Panel-id duplicate | 70 | Same panel id on a different device: the panelist coming back a second time where the fingerprint cannot see it. |
| Patterns | ||
| Near-identical answer profile | 70 | A respondent's whole closed profile matches another respondent far beyond chance. Scored higher when the two also arrived within minutes of each other. |
| Careless responding | 45 | Psychometric evidence of not reading: uncorrelated related items, inconsistent halves, outlier answer profiles. |
| Erratic answering | 35 | Within-person variability far above the sample: random clicking, which never straightlines and never speeds. |
| Response style | 15 | Extreme, midpoint or acquiescent responding well beyond the sample. A style, not necessarily fraud, so kept low. |
| Improbable profile | 35 | Low-probability options chosen right through the survey, standardised per question. |
| Over-claiming | 35 | Ticks far more multi-select options than peers. Fifteen of seventeen brands when the average is three or four. |
| Timing | ||
| Duration cluster | 30 | Total duration near-identical to an unusual number of other respondents: a script signature. |
| Entry burst | 25 | Started within two minutes of many others on one channel. Weak on purpose: panels legitimately launch in bursts. |
| Impossible duration | 55 | Reported duration longer than the wall-clock time between start and finish. Reads 0 on Glow-collected data by construction; it exists to catch imported files with forged durations. |
| History & returning panelists | ||
| On the blocklist | 75 | Device was removed from an earlier study and banked on the cross-project blocklist. |
| Attribute drift | 25 | A returning panelist who changed an answer returners almost never change (birth year, postcode). Deliberately low: a lead to check, not a verdict. |
Tuning weights and thresholds
Every weight and threshold is editable, per survey. Two common adjustments:
- Speeders and skimmers default to flagging anyone who finished under 40% of the median duration. On a long or repetitive survey honest respondents can legitimately race, so dropping the line to 33% narrows the net to the extreme cases; on a short, rich survey the default is usually right.
- Straightliners require a minimum of 8 statements per grid by default, which suits same-valence batteries (all positively worded). For mixed-valence grids (positive and negative statements mixed) a flat line is much stronger evidence, so dropping the minimum to 5 is reasonable.
Rule of thumb: tune thresholds before you start deleting, not after. Changing a threshold rescores the sample, and you want your verdicts recorded against the scoring you actually used.