Published June 20, 2026 | Version v1

Knowing When Not to Learn: Coherence-Gated Plasticity and Viable Adaptation Under Corrupted Learning Conditions

Description

 

This technical note introduces a minimal learning-neuron model for studying when learning preserves viability and when learning becomes structural damage. The central claim is that task performance alone is not enough to evaluate intelligence. A learner can reach the correct answer while accumulating scar, losing coherence, or failing recoverability.

The model compares three learning regimes: ungated learning, frozen learning, and coherence-gated learning. The ungated learner updates from all input, including corrupted or low-coherence input. The frozen learner suppresses plasticity and protects itself by refusing adaptation. The coherence-gated learner updates only when internal coherence is high enough to permit safe learning.

Across a sequence of tests from canonical perturbation through robustness analysis, adversarial corruption, false-curvature formation, hard-mode false-identity lock-in, gating-corridor mapping, and delayed-gate timing, the results show that unrestricted learning can preserve output performance while damaging recovery. Frozen learning can avoid some damage while failing to adapt. Coherence-gated learning is not universally immune to false capture, but it occupies the strongest tested region when success is measured by viable adaptation: task success, coherence recovery, bounded scar, and actual recovery.

The main result is that viable intelligence does not live in maximum openness or maximum closure. It lives in a bounded corridor between the two. Too little gating admits corrupted input. Too much gating blocks truth from re-entering. A boundary must not only open and close, it must do so on the timescale of recovery.

The paper frames this as a minimal model, not a complete theory of intelligence, consciousness, or biological cognition. Its purpose is to isolate a core principle for Cognitive Physics and learning systems: viable intelligence requires regulated permeability across a recovery timescale.

Core sentence:

A mind must be open enough to learn, closed enough not to be rewritten by chaos, and timed well enough that repair arrives before damage becomes identity.

Files

Knowing_When_Not_to_Learn_Paper_v0_2.pdf

Files (476.9 kB)

Name Size Download all
md5:027a9c4eedea647f2bb708189b91ee71
476.9 kB Preview Download