@techreport{hamilton2026ethical, author = {Hamilton, Michael P., Ph.D.}, title = {Ethical AI Lab: Evaluating Local Ecological Summarization with Comma v0.1 on Apple MLX}, institution = {Canemah Nature Laboratory}, year = {2026}, number = {CNL-TN-2026-039}, month = {april}, url = {https://canemah.org/archive/document.php?id=CNL-TN-2026-039}, abstract = {This technical note documents the design, implementation, and evaluation of the Ethical AI Lab, an experimental dashboard that connects the Macroscope sensor network to a locally-running, ethically-sourced language model for ecological data summarization. The system uses Comma v0.1-2T, a 7-billion parameter model trained exclusively on public domain and openly licensed text (Common Pile v0.1), running via Apple MLX on an M4 Max processor. The dashboard fetches live monitoring data from three sensor types (WeatherFlow Tempest, BirdWeather acoustic network, and Ecowitt garden station) through the macroscope database, assembles structured prompts with ecological context from macroscope\_nexus curated place records, and sends them to the local model for summarization. Evaluation reveals that Comma produces serviceable prose synthesis of structured data but fails reliably on quantitative operations including ranking, comparison, and arithmetic. These findings define the operational envelope for small, ethically-sourced models in ecological monitoring: useful as a data-to-prose narration layer when paired with programmatic analysis, but unsuitable for independent analytical reasoning. The complete system runs without cloud dependency, using no data from proprietary or ethically questionable training sources.} }