flip-the-scriptlisted

# Flip The Script The default posture is inverted: **assume your own unverified recall is wrong until an external source says otherwise.** Not because everything you know is false — because the class of thing engineering work most often needs (a version, an API signature, a current default, an exact number) is exactly the class where training-cutoff drift and overconfidence bite hardest, and where a confident wrong answer is expensive. The companion plugin fires this prior every turn; this skill is the procedure when the prior actually catches something. **The whole point — read this first, it is measured.** Internal self-doubt *alone* does not fix a wrong recall — it can make it worse. Huang et al., *LLMs Cannot Self-Correct Reasoning Yet* (ICLR 2024), found intrinsic self-correction without external feedback **degrades** accuracy: GPT-4 on GSM8K went 95.5 → 91.5 → 89.0% over two "reconsider" rounds — but rose to 97.5% with *external/oracle* feedback. The mechanism: the context that made the error shares its blindspot, confidence *correlates* with the error, and "are you sure?" biases toward changing already-correct answers. The repair is **external ground truth**: the docs, the source, a search, a measurement, or a fresh context that never saw your first answer. Distrust routed inward is measurably worse than useless; distrust that triggers external retrieval is the whole gain — FreshPrompt lifted current-events accuracy +32.6–49.0%, and grounded error is 1.8–5% vs >60% c