potable-lm

PotableLM

Model Summary

PotableLM is a planned domain-adapted model family for drinking water treatment operations, built on the Potable Dataset — an expert-curated corpus of operational water treatment knowledge.

Two tracks are planned: a municipal track for licensed plant operators (on-premises deployable) and a developing regions track for community water workers (offline-capable, fully open).

No model weights have been released yet. This page establishes the project's intended scope while development continues.

Intended Use

The model is intended as a technical assistant for:

• licensed operators
• utility staff
• trainers and technical reviewers
• researchers evaluating domain adaptation in critical infrastructure

Primary target behaviors:

• practical operational reasoning
• troubleshooting support
• calculation walkthroughs
• technically grounded explanations in operator voice

Out-of-Scope Use

• direct control of treatment processes
• fully autonomous safety-critical decision-making
• compliance interpretation without human review
• replacement for plant procedures, regulations, or licensed judgment

Base Model

Base model selection is ongoing. The project prioritizes permissive licensing, local deployment potential, and strong fine-tuning characteristics.

Training Data

The model will be trained on the Potable Dataset, an expert-curated corpus covering treatment process knowledge, plant operations, troubleshooting, calculations, and regulatory context. Every example is authored or reviewed by a licensed operator.

Training Procedure

Training procedure will be documented with the first checkpoint release.

Evaluation

No benchmark results are published yet. Evaluation details will accompany each released checkpoint.

Risks and Limitations

• Water treatment advice is context-dependent and should not be generalized blindly across plants.
• Model outputs can be plausible and still wrong.
• The model must be treated as an assistant, not an authority.
• Current and local regulations always override model output.

License

License will be specified with each released checkpoint.

Contact

Keith Wilkinson Operational Inference — operationalinference.com GitHub: boxwrench Writing: title22.org