What question did this study set out to answer?

The aim is to establish a three-layer architecture for effective governance-oriented control of large language models.

April 18, 20260 citationsOpen Access

Ping, Hook, Lever: A Lightweight Architecture for Auditable Behavioral Control of Large Language Models

Key Points

The aim is to establish a three-layer architecture for effective governance-oriented control of large language models.
Developed PHL architecture with three components: Ping, Hook, and Lever.
Ping injects structured packets at the start of sessions.
Hook switches operational modes based on governance headers.
Lever assesses output quality using the ZYXTS scoring framework.
Conducted empirical evaluation across 255 sessions.
Achieved a mean quality gain of Δtotal = +3.8 in session evaluations.
Measured ZYXTS scores increased from 0.6 to 4.4, indicating improved output quality.

Abstract

We present PHL (Ping, Hook, Lever), a three-layer architecture for governance-oriented behavioral control of large language models. Ping injects a structured essence packet at session start; Hook detects governance headers and switches operational mode automatically; Lever evaluates output quality using the ZYXTS scoring framework. Empirical evaluation across 255 sessions demonstrates a mean quality gain of Δtotal = +3.8 (ZYXTS: 0.6 → 4.4). This work is part of the MoCKA (Model of Cybernetic Knowledge Architecture) longitudinal publication series.

问 AI

Bookmark

View Full Paper