The classifier is distilroberta-finetuned-financial-news-sentiment-analysis in
Xenova's quantized ONNX export, pinned at revision e6953e30 and checked by size and
SHA-256.
- Model download: the model is not in the repository. The news job downloads it in
the background into the models folder of the data directory, the first time, and
retries after 6 hours if that fails. Until then the keywords judge alone, as
SPEC 10.3 requires.
- Tokenizer: RoBERTa's byte-level BPE travels inside the DLL, read with
BpeTokenizer and the RoBERTa pre-tokenizer. CodeGenTokenizer.Create adds ids for
runs of spaces that the model rejects. No special tokens are recognised in
untrusted text.
- Token check: the test compares the ids on 14 headlines with an independent copy
of the GPT-2 encoder run on the same files.
- Model test: it reads the tone of clear headlines at 23-24 ms per title on one
thread. It runs where the model is already downloaded and is skipped in CI.
Microsoft.ML.OnnxRuntime 1.31.1 and Microsoft.ML.Tokenizers 2.0.1 come from the
ADR 0004 list. Google.Protobuf joins the known indirect dependencies.
Anubis and Anubis.Server now declare win-x64 among their runtime identifiers, so
their lock files hold the Windows native packages. Publishing the Windows zip no
longer rewrites the lock files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>