# Supplementary Material S1

## Declaration and Persistent State Change in Artificial Systems
Joseph Anady

This supplement accompanies the standalone JAIC manuscript. It is not an additional article, a new training run, or a consciousness classification.

## Submission revision

The accompanying manuscript now locates declaration and persistent discrepancy independently, allows either event order, and distinguishes onset from persistence after declaration. These definitions and the intervention protocol are prospective. No declaration data or new consciousness experiment have been added.

The original numerical data, analysis programs, source receipts, notebook, and figure files are retained. All three offline commands were rerun for this revision. The five transferred data digests matched and all 22 computational tests passed. The new logs are `results/revision_integrity_verification.txt`, `results/revision_analysis_verification.txt`, and `results/revision_tests_verification.txt`. `provenance/revision_environment.json` records the actual verification environment and scope. Source acquisition was not repeated, and the preserved notebook was not reexecuted during this revision.

## What the files contain

`data/gradient.npy` contains all 792 logged gradient observations. `data/eval.npy` contains all 79 evaluation loss observations. `data/pca.npy` contains the 158 regularly sampled checkpoint coordinates and their training steps. `data/selected_checkpoint_metrics.csv` contains 16 selected checkpoint rows, not the full 167 row geometry. `data/corner_sensitivity.csv` contains all 18 lag and amplitude conditions. The JSON files retain checkpoint comparisons and numerical validation.

`provenance/source_verification.json` records the pinned source revision, original file receipts, exact transferred-data digests, the earlier 330 value verification, and the scope of the analysis. Raw source tensors and trainer JSON are not redistributed. The complete 167 row geometric table is regenerated by `code/checkpoint_analysis.py` from the identified public release.

`results/` contains the numerical summaries and tables used for reporting. `Declaration_Study.executed.ipynb` is the previously executed offline analysis notebook. It operates on the included arrays and selected tables. Its recorded output is not a new base-model inference experiment. `figures/` preserves the original study plotting outputs for the notebook. `journal_figures/` contains the four newly formatted manuscript figures at 600 dpi and their EPS originals. The journal figure numbering corresponds to the manuscript.

## Offline checks

Use Python 3.11 or later and install `requirements.txt` into an appropriate environment. From this supplement's root directory:

```
python code/verify_package.py
python code/analyze_tables.py
python -m unittest discover -s code -p 'test_*.py' -v
```

The original 22 computational checks pass in the preparation environment. The figure layouts may vary slightly with available fonts, but this does not alter the underlying numerical data.

## Source connected reproduction

```
python code/checkpoint_analysis.py --download \
  --raw-dir raw_sources --output-dir recomputed
python code/verify_recomputed.py recomputed raw_sources
```

The acquisition uses revision `1551d514e5977637d5a6fd5198edef4d1d6b80e4` of `ModelOrganismsForEM/Qwen2.5-14B-Instruct_R1_0_1_0_extended_train`. It retrieves small adapter and configuration files and the relevant trainer logs. It does not load the full base model, run paid inference, or train a new model. Network access is required for that path. Follow the original source licenses and attribution.

## Interpretation and assistance

The observations are from one training trajectory. Checkpoints and repeated measurements are not independent training runs. Neither a declaration detector nor subjective experience was measured. The manuscript distinguishes a hypothesized relationship from the computed changes. AI assistance with acquisition, analysis code, execution, figures, and writing is disclosed in the manuscript. This supplement does not imply independent human replication or publication acceptance.

`SHA256SUMS` identifies the delivered files. It is distinct from the original source receipts in the provenance directory. Confirm source byte identities exactly and compare derived floating point values with the documented tolerance when numerical libraries differ.