Reading the Zero: Noise-Referenced Feature Adoption, Parametron Networks, and a Ternary Answer for Near-Ties

LLM-based automated feature engineering such as SIGMA ranks generated features against the SHAP importance ofan added noise column. A parametron, or Kerr parametric oscillator, settles from an undecided state into one of twophases. Both are ternary: below the noise, or undecided, is 0; above it, a sign. We connected the two on syntheticregression worlds with known answers, using an exact TreeSHAP implementation (verified against brute-forceShapley values to 4e-16) and ballistic simulated bifurcation as the classical parametron network. (1) A threshold fromone noise column adopted 20.0% of irrelevant features when nothing was planted; the 95th percentile of 30 noisecolumns adopted 0.8% and still found a weak planted feature (A = 0.1) in 90.0% of replicates. (2) Selecting featureswith the oscillator network, errors concentrated on the features that lock last; three precursors fixed in advance did notpredict failure (AUC 0.47-0.52), but returning the two last-locking features as 0 removed 64.8% of errors against6-13% for static choices. A sealed confirmation on a new family was Unresolved (only 11 of 30 worlds had errors;descriptively 21.5% removed) and revealed two failure regimes, late coin flips and stable traps. (3) With 30 runs,choosing the lowest-energy run removed most errors and made consensus-based abstention unnecessary. (4)Polishing an answer by single-flip descent and returning as 0 every feature that varies among 1- and 2-flip neighbourswithin 1% of the energy left no committed error in single-shot use; its zeros matched the true near-ties found by fullenumeration (develop precision 0.957, recall 0.982). The sealed confirmation on a third family gave precision 0.998and recall 0.997, but was Unresolved because only 10 of 80 worlds had errors left after polishing. Everything here issynthetic; no claim is made about real data, real LLM features or hardware parametrons. ※ The license for codes is AGPL-3.0-or-later ※ AI use. Designs, code, analyses and text were produced by Kurado (Claude Opus 5.5, Anthropic) working with M. Watabe, who chose every direction.