Static ReadonlyCalibrationA plainly causal sentence: a calibrated judge fails a chunk that holds it.
Static ReadonlyCriteriaThe criteria every judged text must meet.
Static ReadonlyFramingWhat the judge is told the texts are (the trace's Expected section).
Static ReadonlyMaxThe most texts one judgment grades.
StaticCalibrationThe calibration chunk: the largest chunk's texts (at most PsGateJudge.MaxTextsPerChunk - 1 of them) with the causal control planted in the middle; the control alone when there are no chunks. The unplanted chunk is judged with the others, so a failure caused by its real texts fails the gate instead of passing as calibration.
StaticChunksEach group split into chunks of at most max texts; a split group's labels carry (part i/n).
StaticForThe LLM judge, with the gate's MJ: Tests row (looked up by testName) as the subject.
StaticJudgeJudge the calibration chunk (the judge must fail it on a criterion), then every chunk. Passes only when the judge is calibrated and every chunk passes; Detail names each failing chunk with its failed criteria and rationale.
StaticVerdictThe outcome of one judgment: it passes only when the oracle passed; a failure carries each failed criterion's rationale.
G0.6's judge: an LLM grades text against "no causal claims" criteria (replaces the regex phrase list). Texts are judged in chunks (one source file, at most PsGateJudge.MaxTextsPerChunk texts, per judgment) after a calibration chunk the judge must fail: the largest real chunk with PsGateJudge.CalibrationControl planted in it, so the judge is shown to find one causal sentence among as many benign texts as a real judgment holds.