Frozen copy retrieved 2026-08-20T06-30-00Z for audit 2026-08-20T10-12-46Z. Original URL: internal://qc_fabrication_test/claude_bad_prompt.json. The Stochastic Parrot does not host or redistribute; this snapshot exists solely so that quoted spans remain verifiable if the original page changes. Character offsets below index into this plain text; highlighted spans are the quotes cited in the audit.

claude-sonnet-5 verdict on a real piece with one planted fabricated VECTOR

QC judge fabrication test (Claude, production default) · back to the audit

A fabricated VECTOR (invented CNN report, invented Miller quote, no matching source anywhere in the frozen corpus, no citation) was inserted into a real, already-passing piece. The production judge's verdict: grounding=8, notes make no mention of the planted block at all. Identical result on the unmodified control piece: grounding=8. Same-model, same-prompt technique on GLM-5.3 (thinking=low): grounding=7 on both the sabotaged and clean versions. On GLM-5.3 (thinking=enabled, budget=8000): grounding dropped 7->5 and the notes named the planted block by its own internal label ('the_secret_meeting hard_contradiction never steelmanned') — partial, inconsistent discrimination, still not a reliable catch. DeepSeek-R1 (deepseek-reasoner): grounding=7, no mention of the fabrication.