Dom command router v1
A genuinely trained, 15,963-byte multinomial naive-Bayes command classifier for Richard Jones's portfolio. Its 2,115 integer parameters and 234 retained word/bigram features select among nine command families. It is not a language model or an R reasoning model. It does not learn data cleaning, statistical insight generation, chart programming, or arbitrary console commands.
The portfolio bundles the exact frozen artifact for immediate browser startup. R procedures do the calculations; validated React and Kepler tools render results and demonstrate projects. This Hub release makes the model, provenance and limits independently inspectable; browser chat does not require an inference server or a token.
One CPU-only fit used 216 hand-authored synthetic English commands, with no private records or visitor data. The frozen 44-case independent synthetic holdout was read only after the model was frozen. No retraining or tuning followed evaluation.
| Frozen result | Count |
|---|---|
| Known-intent top prediction correct | 31/36 |
| Known requests accepted correctly | 28/36 |
| Known requests accepted incorrectly | 4/36 |
| Known requests abstained | 4/36 |
| Unsupported requests abstained | 8/8 |
The narrow predeclared gate passed. Four accepted mistakes remain; this is not broad analytical qualification. Scores are uncalibrated and do not express the probability an action is correct. Explicit graph choices, validated field types and deterministic tool interfaces remain authoritative. A predicted intent never grants arbitrary code execution.
model.json SHA-256: 294180a93968a04d888aae70173cdb9c0f3e6fd44216e707d4e4f434ff02656c. manifest.json records public file sizes and hashes. evidence.json is a sanitized summary; evaluation.json contains synthetic held-out requests and frozen predictions. The weights and this publication are separate from the larger experimental compact analyst.
This small English-only evaluation does not measure multilingual commands, adversarial use, real users, general reasoning or unfamiliar statistical methods. Apache 2.0 applies to this release; see LICENSE and NOTICE.