# Declaration and Persistent State Change in Artificial Systems
## Source-verified secondary analysis — Version 1.0

Prepared for Joseph Anady, 17 September 2026. Author review pending.

### Completed work

One public training trajectory was analyzed through all 167 released rank-one adapter checkpoints and its full 792-step trainer record. All 330 previously transcribed pilot measurements match the original source exactly. The study computes effective weight-update geometry, source-style local path bends, filter/lag sensitivity, and PCA. It does not train models or generate new model responses.

The manuscript distinguishes measured parameter changes from an unmeasured declaration and from the hypothesis of subjective consciousness. It has not been submitted, peer-reviewed, or externally preregistered.

### Read first

- `Declaration_and_State_Change_Final.pdf`: formatted manuscript.
- `Declaration_and_State_Change_Final.docx`: editable manuscript with native equations.
- `manuscript/Declaration_and_State_Change_Final.md`: manuscript source.
- `Declaration_Study.executed.ipynb`: executed offline reproduction.
- `protocol/Declaration_Study_Protocol.md`: separate proposed follow-up, not completed research.

### Reproduce the delivered results offline

Use Python 3.11 or newer in a virtual environment. Install `requirements.txt`, then run from this directory:

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

The notebook performs those checks and shows observed results. Tensor calculations were originally executed in a network-enabled analysis session; the offline notebook does not claim to load raw tensors not bundled here.

### Recompute from the original public sources

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

This downloads about 13 MB of safetensors, JSON configurations, and trainer states at revision `1551d514e5977637d5a6fd5198edef4d1d6b80e4`. It does not download the full base model, use paid inference, or execute a misaligned model. Network access to Hugging Face is required. The program checks configuration and tensor identity assumptions rather than silently accepting a different model.

### Data contents and boundaries

`gradient.npy`: all 792 step/gradient pairs. `eval.npy`: all 79 step/loss pairs. `pca.npy`: all 158 regular-grid PCA score rows. `selected_checkpoint_metrics.csv`: 16 exact sampled checkpoints, not the entire geometry table. `corner_sensitivity.csv`: all 18 tested lag/filter combinations. Supporting JSON stores exact comparisons, source receipts, and validation results.

All 167 checkpoints were analyzed remotely. The complete 167-by-16 geometry table can be regenerated by the source-connected code. Raw upstream tensor files are not redistributed. Derived data transferred into this package were verified against source-session SHA-256 receipts. Selected CSVs also match their source-session byte digests exactly.

The figures include four manuscript figures and two clearly labeled supplementary plots based on selected checkpoint rows. The notebook does not count time points as independent models or assign consciousness labels.

### Main findings

- First saved nonzero effective adapter update: step 2.
- Maximum logged gradient: step 169.
- Retained descriptive split of the 300-step prefix: after step 181.
- At amplitude threshold 0.002, the largest path bend for each tested lag: step 190.
- With no amplitude filter, the two shorter lags instead select step 725.
- First two B-vector principal components explain 94.990% of variance.
- Evaluation loss falls 63.694% from its first to last record; one tiny late increase remains in the table.
- 22 local software and identity checks pass; 462 stable-form corner comparisons were checked in the source-connected calculation.

These are different operational measurements, not alternative measurements of an established consciousness-onset event. No declaration was observed in the acquired artifacts.

### Attribution and use

Original models, training, logs, and experimental findings are due to ModelOrganismsForEM and the authors cited in the manuscript. New analysis and manuscript are prepared for Joseph Anady with AI assistance. The author must review methods, conclusions, attribution, and the target venue's disclosure requirements before submission. No third-party model/code license is replaced by this package. No universal or biological claim is inferred from this single trajectory.

`SHA256SUMS` records integrity of distributed files. `provenance/source_verification.json` identifies original sources separately from derived artifacts. `provenance/environment.json` records local and source-connected environments. No API keys, private endpoints, or browser-session credentials are included.

### Manuscript files

`Declaration_and_State_Change_Final.pdf` is the 23-page reader copy. The DOCX retains editable equations and text. `manuscript/Declaration_and_State_Change_Final.md` is the plain-text source. Optional document rebuilding with `code/build_document.py` additionally requires Pandoc and compatible fonts; it is not required to reproduce the numerical analysis. No font files are distributed.