# Source spot check October 23 2026

Result: no discrepancies found in the sampled count rows, valid N, published means, source-page references, or the independently reconstructed custom percentages. This is a targeted check of two reports, not a complete audit of 58 judges or a verification of ballot eligibility.

## Sources and method

Used the original PDFs in the supplied transfer: [Brad Astrowsky](https://jpr.azcourts.gov/portals/2/148/jpr_reports/2026/2026-1549.pdf) and [Geoffrey Fish](https://jpr.azcourts.gov/portals/2/148/jpr_reports/2026/2026-1608.pdf). Physical PDF pages 5 and 6 were rendered and visually inspected. Commission conclusions were checked against extracted text on page 1. These are the local source snapshots used for the dataset, not a claim that the remote PDFs have remained unchanged.

Compared the individual “Unbiased and impartial” row for all five groups in each report with `data/judges.json`. The table below records counts in source order: Very Good, Good, Acceptable, Poor, Very Poor. Independently calculated higher-category shares from these counts; then compared them with the application’s rollup. Tests retain these manually transcribed expectations in `tests/source-spotcheck.test.mjs`.

## Ten individual source rows

| Judge | Group | Page | VG / GD / AC / PO / VP | Valid N | Published mean |
| --- | --- | --- | --- | --- | --- |
| Astrowsky | Attorneys | 5 | 28 / 5 / 4 / 0 / 1 | 38 | 4.6 |
| Astrowsky | Litigants/witnesses/others | 5 | 44 / 5 / 8 / 4 / 5 | 66 | 4.2 |
| Astrowsky | Jurors | 5 | 7 / 2 / 0 / 0 / 0 | 9 | 4.8 |
| Astrowsky | Judicial peers | 6 | 9 / 1 / 0 / 0 / 0 | 10 | 4.9 |
| Astrowsky | Appellate judges | 6 | 3 / 0 / 0 / 0 / 0 | 3 | 5.0 |
| Fish | Attorneys | 5 | 31 / 5 / 1 / 1 / 0 | 38 | 4.7 |
| Fish | Litigants/witnesses/others | 5 | 2 / 0 / 0 / 0 / 0 | 2 | 5.0 |
| Fish | Jurors | 5 | 4 / 0 / 0 / 0 / 0 | 4 | 5.0 |
| Fish | Judicial peers | 6 | 18 / 0 / 0 / 0 / 0 | 18 | 5.0 |
| Fish | Appellate judges | 6 | 2 / 0 / 0 / 0 / 0 | 2 | 5.0 |

N uses the individual question row, not the number of returned questionnaires in the column heading. For example, Astrowsky’s litigant column reports 70 returned questionnaires but 66 valid answers to this question; Fish’s appellate column reports 3 returned questionnaires but 2 valid answers here.

## Independently reconstructed custom percentages

Settings: all five respondent groups at 1×; current Superior Court instrument; selected section as named below. Integrity has one child question. The N filter keeps results at N = 10 and excludes those below 10.

| Judge and section | Filter off | Exclude N below 10 | Coverage off / on |
| --- | --- | --- | --- |
| Astrowsky Integrity | 92.216906% → 92.2% | 87.028177% → 87.0% | 100% / 60% |
| Fish Integrity | 98.947368% → 98.9% | 97.368421% → 97.4% | 100% / 40% |
| Astrowsky Legal Ability | 97.722723% → 97.7% | 95.570571% → 95.6% | 33.333333% / 20% |

Astrowsky Integrity without filtering is the equal average of 33/38, 49/66, 9/9, 10/10, and 3/3. Filtering removes jurors (N = 9) and appellate judges (N = 3), retaining attorneys, litigants, and judicial peers. The drop in percentage reflects removal of perspectives, not a correction to the answers.

For Astrowsky Legal Ability, six selected questions yield these question-level higher-category shares with filtering off: `1`, `(34/37 + 1)/2`, `1`, `(32/36 + 1)/2`, `(34/37 + 1)/2`, `1`. Their equal average is 97.722723%. These terms follow the actual populated cells on pages 5–6; blank groups are not zero. Filtering leaves four questions and six question/group results, producing 95.570571% and coverage of 6/30.

## Official findings and different definitions

Both page-1 conclusions read “Meets Judicial Performance Standards,” matching the dataset. The report’s page-1 category summaries count “Acceptable or Above.” Our custom result counts only Good and Very Good, uses user-selected group weights, and averages selected questions. These are different definitions; the custom percentage must not be described as reproducing the official category percentage.

The detailed tables explicitly warn that category summaries are averages and may not add because of rounding. The application uses individual question counts, not those summary counts. This spot check supports that choice but does not establish statistical representativeness or uncertainty bounds.