# T2-CHANNEL-001

Self-contained channel audit of T2-NETI-NETI. Read scientific_record.md for
results and preregistration.md for the prospective rules. No physical or
metaphysical claim is established by this computational study.

## Environment

Executed with Python 3.14.3, NumPy 2.4.4 and threadpoolctl 3.6.0. Figures use
Matplotlib 3.10.8. Exact environment and timestamps: results/execution.json.

```
python -m pip install -r requirements.txt
```

## Inspect the delivered result

results/summary.json holds the verdict, gates, bootstrap intervals and paired
architecture contrasts. all_metrics.json and each r*_metrics.json preserve
every fit. Each r*_model.npz stores trained weights. Each r*_test.npz holds
the 2,000 untouched test inputs, posterior means/SDs, labels and predictions.
Dataset generation is deterministic from the recorded seed. Code uses float64.

Check SHA-256 checksums from the package directory:

```
python -c "import json,pathlib,hashlib; c=json.loads(pathlib.Path('checksums.json').read_text()); assert all(hashlib.sha256(pathlib.Path(p).read_bytes()).hexdigest()==h for p,h in c.items()); print('All checksums passed')"
```

The lock hashes the preregistration, requirements and execution sources. It is
a local prospective record, not a third-party timestamp. report.py is post-run
reporting and is explicitly excluded from the execution lock.

## Repeat the full experiment without overwriting these results

Create a new empty folder. Copy src/, preregistration.md, requirements.txt and
execution_lock.json into it, preserving directory structure. Do not copy
results/. From that new folder:

```
python src/run_study.py calibrate
python src/run_study.py run
python src/run_study.py verify
```

The first command checks an analytic conventional channel and inherited
gradients. The second verifies the source lock and executes 72 fixed fits.
The third independently retrains the prespecified first A3 fit and compares
all model/posterior/prediction arrays bit for bit. Typical main-run time on
the original CPU is recorded in results/summary.json; hardware speed varies.

The delivered result can also be verified in place with the third command;
it writes only to results/verification/ and verification_metrics.json.
Regenerate figures, narrative and archive with `python src/report.py` after
copying this README into the new package. Reporting updates checksums.

## Provenance

common.py, vae.py and linear_gaussian.py are unchanged copies of the parent's
local training code, made public here for reproducibility. Their original
docstrings describe the parent design; the new protocol governs this study.
In particular, this study explicitly supplies d_z=4 to the linear control.
No third-party participant data or external model service is used.
