README.md

---
license: cc-by-sa-4.0
pretty_name: Strange loop in J-space
language:
- en
tags:
- interpretability
- self-model
- strange-loop
- attention
- multimodal
---

# Strange loop in J-space — research data

Wiktor Tomasik · 2026

Measurements and learned interfaces for [Strange loop in J-space](https://github.com/Pixedar/jspace-strange-loop), a study of self-description in feedback systems around frozen language models.

In the attention-feedback experiments, a small reader summarizes recent model activity. The summary sets a control state, which changes the next reading. Training asks the summary to distinguish possible updates after subtracting their common destination. Some trained interfaces develop dependence on their starting state. Applying the reader’s attention inside the language model also changes predictions, while the measured feedback through model processing remains weak.

## Contents

| Path | Contents |
|---|---|
| `self_reference/results/` | Numeric recordings from all nine completed shared-attention training runs, including controls. Fixed-probe arrays record next states for 12 candidate starting states at 60 text positions and four copies. |
| `self_reference/checkpoints/` | Learned Qwen3 interface weights and two starting interfaces. These do not contain base-model weights. Configurations accompany the interfaces. |
| `followups/recordings/` | State trajectories from generated-text comparisons, and cached features and response fields from the text/image/audio experiment. |
| `self_reference/ARTIFACTS.json` | Selected-file hashes and sizes. The code repository pins an immutable dataset revision in its root `ARTIFACTS.json`. |
| `v5/`, `phaseB/` and other historical paths | Recordings for word-based frozen feedback, memory between reasoning segments and other experiments described in the research history. Existing paths are retained. |

## Interpretation and reproduction

The [report](https://github.com/Pixedar/jspace-strange-loop/blob/main/REPORT.md) defines the measurements and controls. The state-dependence score is a normalized ratio, not a probability; positive correlation between predicted and measured response patterns does not imply accurate prediction scale. The experiments do not measure subjective experience.

The [reproduction guide](https://github.com/Pixedar/jspace-strange-loop/blob/main/self_reference/README.md) provides a hash-verifying downloader and code. The [behavior and multimodal methods](https://github.com/Pixedar/jspace-strange-loop/blob/main/followups/README.md) explain the additional experiments, including unsuccessful perception controls and weak image effects. The [interactive comparison](https://huggingface.co/spaces/Pixedar/strange-loop-in-j-space) displays measured starting-state interventions and feedback-path cuts.

Author-owned numeric observations introduced in release 4 are CC BY-SA 4.0; experimental software is AGPL-3.0-only. Earlier CC BY 4.0 grants and third-party terms remain valid. See LICENSING.md for scope. The external text corpus is not redistributed. Neither Qwen2.5-Omni base weights nor its trained reader weights are included; upstream model terms remain applicable. The preceding dataset account is preserved in history/README_before_20261019.md.