DCS-HF v0.1: Grey-Box Adapter for Directional Character Measurement in Open-Weight Language Models

DCS-HF v0.1 introduces a grey-box adapter for transferring the Directional Character Sandbox (DCS) apparatus from synthetic calibration trajectories to open-weight language-model probability fields. It builds on the Directionality Before Intention research program and the DCS v0.2/v0.3 methods/software notes, which established a proof-of-apparatus pipeline and baseline/sensitivity validation for directional-stabilization metrics.

This work does not claim to demonstrate intention, consciousness, agency, autonomy, or endogenous goals in language models. Its purpose is narrower: to define and package an adapter that maps candidate-string likelihoods, token-level probabilities, transition scores, or sampled outputs into DCS decision classes while preserving strict claim boundaries.

The central principle is that conditions are not intentions. Conditions shape the probability field in which directionality may or may not stabilize; they do not by themselves constitute a system’s intention.

The accompanying software bundle supports a dependency-free mock mode and an optional Hugging Face smoke-test mode. It exports raw likelihoods, class probabilities, adapter diagnostics, metrics, threshold sensitivity, template sensitivity, candidate-order sensitivity, failure cases, and a machine-readable decision file. The framework explicitly reports risks such as tokenization artifacts, verbalization bias, classifier Goodhart, prompt-template dependence, sham vulnerability, sign-flip insensitivity, ablation insensitivity, low signal, and seed instability.

DCS-HF v0.1 is therefore an adapter-transfer scaffold, not a model-level result. Its contribution is to provide a bounded, inspectable, and falsifiable bridge from synthetic DCS apparatus validation toward future preregistered grey-box studies on open-weight language models.