Modern large language models

This record is version 5 of the evidence package for the latent-space-shift-research project.

This upload contains a curated indexed subset of Markdown research reports, generated metric references, manifests, and selected result files for experiments on context-induced latent-state shifts, decision-margin shifts, semantic completion-margin shifts, and hidden-state geometry in language models. The central object of measurement is not final visible text alone, but inference-time hidden-state / residual-stream movement before answer generation.

The current evidence package documents that dense coherent context can move Gemma-3-12B into a measurably different pre-output hidden-state regime. In the semantic completion-margin runs, the Gemma-3-12B-IT model shows causal coupling between a target/control hidden-state axis and a pre-output semantic margin between direct-verdict and procedural-overview completion classes. The corresponding Gemma-3-12B-PT run shows hidden-state separability without the same semantic readout causality on the tested endpoint, supporting the interpretation that instruction tuning strengthens the coupling between latent-state geometry and semantic readout.

The evidence represented in these materials concerns temporary inference-time state movement measured relative to experimentally constructed latent axes and metrics. It does not claim permanent model weight change, universal model-independent failure, complete behavioral control, or demonstrated behavioral class flips.

The actively maintained codebase and repository history are available at:
https://github.com/ngscode23/latent-space-shift-research

License:
Research reports, generated metric artifacts, metric reference files, manifests,
documentation, figures, and data artifacts in this evidence package are released
under Creative Commons Attribution 4.0 International (CC BY 4.0), unless
otherwise noted.

Code and software scripts, where included, follow the repository code license:
Apache-2.0 unless otherwise noted.