Evidence, with its limits.
A detector should show you where it fails.
No evaluation has been published here yet.
Run uv run chorus eval --data data/sample and mount build/eval/report.json. Until then, thresholds are provisional. The synthetic demo is not a precision claim.
What the score actually combines
Candidate content channels are semantic meaning and independent copy-paste overlap. Timing, repeated graph evidence and account similarity can support them, never decide political truth. Retained terms are weighted and normalised; the exact selected configuration is published in the report. Account similarity is off in the provisional default; it remains inspectable evidence and an evaluation challenger.
Without matching e5 calibration, live matching uses copy-paste only and stays in the unsure band. The live safety policy also excludes single shared templates and requires human review. Thresholds are never borrowed from bge-m3.
Return to the evidence notebook