# Evidence: the run of 7 October 2026

The files a reader needs to check the figures quoted for this example. They
come from one run on a clean rented machine (8 virtual CPUs, 32 GB, Ubuntu
24.04, Python 3.12.3, Shuddhi v1.2.0, PyTorch 2.14.1 CPU) that followed the
README in the parent folder command for command, starting from nothing.

| File | What it is |
|---|---|
| `shuddhi-out/build/BUILD-MANIFEST.json` | The build manifest: the three hashes, drops by reason, the filter configuration, one sha256 per cleaned file |
| `shuddhi-out/run/MANIFEST.json`, `COMPOSITION.md` | The corpus manifest and its readable summary |
| `shuddhi-out/REPORT.md` | The draft training-content summary Shuddhi wrote |
| `model/run.json` | The run card: architecture, corpus hashes, the full loss log |
| `model/samples.jsonl` | The six generated samples, as generated, with the copied-run measurement |
| `logs/` | Console output of every step, the training log, timings, package versions, sha256 lists, and the provider's billing reading |
| `registry.json` | The registry the build used |

`logs/SHA256SUMS.txt` lists the sha256 of every file the run produced,
including the prepared corpus, the cleaned text, the tokenizer and the model
checkpoint, none of which are stored here.

## Redactions

Two things were changed in these copies, and nothing else:

- the `host` field in both manifests, which Shuddhi fills with the build
  machine's hostname, now reads `[redacted]`;
- the home directory in `logs/console.log` and `logs/doctor.log` reads
  `/home/[user]`.

Because of the first, the sha256 of the two manifest files here differs from
the one in `logs/SHA256SUMS.txt`. The hashes *inside* the manifests are
untouched.

## The samples, and the drop log

`model/samples.jsonl` is raw model output. It was read and screened before
publication; it is clumsy and none of it is unsuitable, but it has not been
edited.

The drop log (`shuddhi-out/build/dropped.jsonl`, every dropped story with its
reason and a preview) is not in this published copy, because it quotes the
passages that got stories removed by the content screen. It is kept with the
example's repository copy. The counts are in the build manifest here, and
`logs/why-dropped.log` shows what each contamination drop matched.