What question did this study set out to answer?

The aim is to understand when a cognitive system emits a say-event without relying on the content of the event.

April 16, 2026Open Access

Predictive Feature Architectures For Self-Supervised Say-Events in VDM

Key Points

The aim is to understand when a cognitive system emits a say-event without relying on the content of the event.
Utilized a 1k-node run segment with over 68,000 ticks for analysis.
Developed a logistic motor-gate model to predict say-events.
Performed time-split testing to assess model performance using metrics like AUC and average precision.
Achieved near-perfect prediction with an AUC of 0.9994 and an average precision of 0.8600.
Identified the boundary-pulse proxy as the dominant predictor with a univariate AUC of 0.9812.
Showed minimal improvement in prediction accuracy when including input text-topic features.

Abstract

This paper reports an empirical readout for when a zero-training cognitive runtime actuates (emits a say-event), independent of the emitted semantic content. Using an archived 1k-node run segment containing 68, 403 ticks (t = 316, 741. . . 385, 143) and 312 actuation ticks (rate 0. 004561), a compact logistic motor-gate model predicts didₛay from a small set of internal state proxies with near-perfect discrimination (time-split test AUC 0. 9994, average precision 0. 8600). The dominant predictor is a boundary-pulse proxy (b1ᵦ), which alone achieves univariate AUC 0. 9812 and yields a high-precision actuation aperture when thresholded (e. g. , duty cycle about 0. 4% at conditional say-rate about 0. 68). Ablations indicate that adding input text-topic features provides negligible improvement to actuation prediction, supporting a separation between motor gating and downstream content selection. Appendix A documents known proxy-physics components in the logged runtime and the canonical replacements required for physics-native interpretation.

AIに質問

Bookmark

View Full Paper