Rules, version 1.0.0
Every poll is graded by these rules, on the same day. A rule reads facts taken from the poll's own documents. If any fact it needs could not be read, the result is could not check, never clear.
Extraction accuracy: measured 2026-10-23 on 450 labelled facts (labelled by an independent AI model, blind to our answers). A rule's results are shown only when every fact it reads was extracted with precision of at least 0.80.
| Rule | Severity | Reads | Shown? |
|---|---|---|---|
| Did not disclose its basic method: sample size, field dates and how people were reached. | high | Sample size, Field start date, Field end date, How people were reached | yes |
| Sample is below 400 for a statewide race (300 for a district). | medium | Sample size | yes |
| Fielded in a single day. | low | Field start date, Field end date | yes |
| First listed more than 21 days after fieldwork ended. | low | Field end date, First listing date (from the poll listings) | yes |
| Sponsored by a campaign, party or political committee. | medium | Sponsor, Partisan mark (from the poll listings) | withheld: extraction accuracy below 0.80: sponsor recall 0.60 |
| Did not disclose crosstabs. | medium | Crosstabs | yes |
| Party mix of the sample differs from the benchmark by more than 6 points. | medium | Party mix of the sample | withheld: extraction accuracy below 0.80: party mix of the sample precision 0.67, party mix of the sample recall 0.67 |
| Age mix of the sample differs from the benchmark by more than 8 points. | low | Age mix of the sample | yes |
| Surveyed registered voters or adults, not likely voters, within 60 days of the election. | low | Field end date, Who was surveyed (from the poll listings) | yes |
| Did not disclose how the sample was weighted. | medium | Weighting | yes |
| The ballot question's wording may lead respondents. | low | Ballot question wording | withheld: decided by a wording judge that is not built or measured |
Version history
- 1.0.0 (2026-10-23): First version: the eleven DESIGN §8 rules.