The sentence lottery

A sentence is supposed to price the act. In the documented Belarusian political cases it also prices the judge. We compared every sentenced case only against other cases carrying the same lead charge in the same year — protest cases against protest cases, "extremist formation" cases against each other — and asked whether the judge's identity still moved the number. It does, by years. The estimates are deliberately conservative: a judge's own cases are excluded from the baseline they are compared against, small caseloads are shrunk toward zero rather than trumpeted, and the whole spread was stress-tested by reshuffling judges across their cases five hundred times. Random assignment of names never reproduced the observed spread. What that means for a defendant is simple and grim: walk into one courtroom and the tariff for a donation or a comment is measurably heavier than the identical tariff one door down. Arbitrariness of exactly this kind is what "independent tribunal" language in international law exists to prohibit — and here it is, quantified, name by name, with the case files attached.

Empirical-Bayes shrunken premiums for the harshest judges with at least 3 comparable cases. Filled bars are individually significant at p < 0.05 (within-cell permutation test); the others are directionally harsh but small-sample.

The most lenient judges by the same method. Leniency here is relative to a repressive baseline — a "lenient" judge in this data still jails people for speech; they simply do it for fewer years than colleagues handling identical charges.

• §This is documented disparity with same-charge, same-year controls — not a randomization-based causal estimate. Court assignment is not yet observed in our data, so we cannot rule out that harsher judges systematically receive graver variants of the same charge. The within-charge-and-year design limits, but does not eliminate, that channel.
• §Judge names come from the Viasna Human Rights Centre's case documentation (676 sentenced, judge-attributed cases), imported with verbatim source quotes; every judge assertion in our database carries its evidence and can be audited case by case.
• §Caseloads are small (3–28 comparable cases per judge). Premiums are shrunken toward zero in proportion — a judge shown at +2 years with 5 cases had a larger raw gap. 10 of 83 listed judges are individually significant; the headline claim (the spread itself) is significant at p < 0.002 regardless of any single judge.
• §These are cases documented in our database (primarily from the Viasna Human Rights Centre), not a count of every prosecution. The value here is case-level detail: what each person actually did, with a source.

Every documented case behind this finding, each linking to its profile and primary source.