On the Impossibility of Internal Attractor Basin Detection: External Epistemic Gating as a Necessary Condition for Reliable Inference Across Cognitive Scales
Formal argument demonstrates the necessity of external verification for accurate inference in cognitive systems, suggesting implications for belief systems and AI.
Key Points
This research aims to prove that systems cannot reliably determine their attractor basin type solely from internal signals.
Formal argument derived from a dynamical-systems framework.
Experimental results from machine learning showcasing the impact of supervised fine-tuning and external verification.
Transfer of results to formal domains like Lean 4 theorem proving.
Supervised fine-tuning on internally generated signals decreased base capability.
Use of an external verification channel led to an 11.3% increase in task accuracy.
Findings apply to various phenomena, including post-traumatic stress and the alignment problem in AI.