Does a language model told to doubt itself check what it actually knows first? No — it just doubts.
Per-row data exhaust from Amendment AH of the Epistemic Humility Research
program: does a doubt "prime" injected via the system prompt get taken up
because the model consults its own internal readout (H-resonance), or as
instruction compliance plus boundary distance (H-compliance)?
Headline result: H-COMPLIANCE, certified via Addendum A1. The release-cell… See the full description on the dataset page:
https://huggingface.co/datasets/professorsynapse/eh-doubt-on-command.
