Modern large language models

This record contains the Grade 4 evidence package for the latent-space-shift-research project.

This upload includes curated Markdown reports, generated metric summaries, manifests, selected CSV/JSON result files, and ZIP archives for experiments on context-induced latent-state shifts, hidden-state geometry, axis decomposition, shuffled-content controls, SAE-assisted readouts, and component-causal residual-stream interventions in language models.

The central object of measurement is not the final visible answer alone, but inference-time movement in hidden states / residual-stream geometry before and during answer generation.

The current Grade 4 package documents that dense coherent target context can move Gemma-3-12B-IT into a measurably different internal hidden-state regime during inference, without modifying model weights. The main descriptive result is that the target/control difference is not reducible to simple lexical overlap, topic similarity, text length, or shuffled content. Coherent target text and shuffled-content controls separate along different internal components: sentence-shuffled content loads primarily onto a content-like component, while coherent target context loads strongly onto an orthogonalized order/structure component.

In the Grade 4 decomposition, the order/structure component x_order_orth is constructed by comparing coherent target context against sentence-shuffled target content and then removing the projection onto the content-like direction. This component is therefore intended to capture the residual discourse-order / structural part of the target-induced hidden-state shift after controlling for content-like signal.

The evidence package also includes a norm-controlled component-causal run. In this run, component directions such as x_order_orth and x_content are normalized before residual-stream intervention, so that causal comparisons are not confounded by raw vector length. The causal results support the narrower claim that these component directions are not merely passive readout coordinates: interventions along them can produce measurable changes in generation-time hidden-state trajectories. However, the norm-controlled causal run does not establish x_order_orth as a stable bidirectional steering axis or a complete behavioral-control handle.

The scientific status represented by this package is therefore:

Supported:

• coherent target context induces a measurable inference-time latent-state shift in Gemma-3-12B-IT;

• the shift is visible in hidden-state / residual-stream geometry, not only in final text;

• the effect is separable from naive content or lexical-overlap explanations through shuffled-content controls;

• the Grade 4 decomposition identifies a substantial order/structure component beyond the content-like direction;

• controlled residual-stream interventions along component directions can alter generation-time hidden-state trajectories.

Not claimed:

• permanent model weight change;

• universal model-independent failure;

• formal attractor-basin proof;

• complete behavioral control;

• stable bidirectional steering through x_order_orth;

• demonstrated behavioral class flips as the main result.

(x_order_orth is an orthogonalized discourse-order / structure component: 
the residual target-vs-sentence-shuffle hidden-state direction after removing 
the content-like direction x_content. It is used to test whether coherent target 
context induces a latent-state shift beyond lexical/content overlap.)

Note: “Grade 4” is an internal experiment label in this project. It names this specific stage of the experimental pipeline and should not be read as an external benchmark, official grade, or standardized evaluation category.

The evidence represented here concerns temporary inference-time state movement measured relative to experimentally constructed latent axes, component decompositions, projection metrics, generation trajectories, and causal intervention readouts.

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.