A Library of Validated Public-Sector Models
Standardized output contracts so improvements propagate — not isolated legacy forks.
Open source software promised shared infrastructure but largely produced shared code that each jurisdiction customized into isolated legacy systems. A model commons governed by standardized output contracts can break that pattern. When a county solves a hard problem well, the improvement propagates. The ecosystem learns. Every region benefits from real-world performance data that no benchmark can replicate.
How Contribution Works
Contributing to the Registry
A county fine-tunes a zoning review model on three years of local permit decisions. The fine-tune improves accuracy on edge cases involving mixed-use parcels. They submit to the registry with: (a) the model weights, (b) a model card describing training data, known limitations, and equity audit results, (c) a regression test suite against the standardized output schema, (d) performance delta vs. the baseline model.
The registry runs automated schema compliance checks. A multi-stakeholder review board (elected officials, civic tech orgs, academics, public input) evaluates the model card. If accepted, the improvement is available to every jurisdiction on the next registry update.
What "Validated" Means
Validation Criteria
A model is registry-eligible if it meets all of the following:
- Produces output compliant with the relevant state-defined schema
- Ships with a model card covering training data provenance, known failure modes, and bias/equity audit results
- Passes a regression test suite against historical decisions from the submitting jurisdiction
- Includes an explainability summary: what changed from the prior version and what the measurable downstream effects were
- Meets a minimum energy efficiency threshold (to be defined by the registry board)