Developer danielioni has introduced a significant architectural shift in the MyZubster ecosystem with PR #1573, targeting the Zorgax AI assistant. The core problem addressed is not how to feed an AI more data, but how to strictly define what it is permitted to do with that data. By implementing a machine-readable capability model, the project aims to prevent AI assistants from making unfounded inferences, a common failure mode in current generative AI integrations.

Separating Contributor Competence from AI Capability

The new schema, myzubster.zorgax-capability.v1, explicitly decouples contributor records from AI capabilities. A contributor record tracks provenance, evidence state, and limitations of human input, whereas a Zorgax capability defines allowed actions, prohibited claims, and required source states. This distinction ensures that if a contributor has expertise in a domain, Zorgax does not automatically inherit that expertise as an unrestricted claim, preventing the AI from acting as an authority it hasn't earned.

Strict Guardrails and Initial Capability States

The implementation adopts a deliberately conservative approach to initial states. The Contributor Profile Assistant and Interoperability Checkpoint Assistant are marked as DOCUMENTED, while only the Research Evidence Navigator is currently TESTED. For instance, the Contributor Profile Assistant can summarize declared interests but is explicitly prohibited from inferring employment, certifications, or qualifications. If a user states they are interested in usability testing, Zorgax must preserve that exact phrasing rather than converting it into a claim of being a 'Certified software tester.'

CI Enforcement via Continuous Evidence Gate

To ensure these rules are not just theoretical, a validator has been integrated into the Continuous Evidence Gate. This CI check enforces schema validity, detects duplicate capability IDs, and verifies that any capability marked as TESTED is linked to bounded, reproducible checkpoint evidence. This moves capability state from prose documentation to a hard technical constraint that CI pipelines can reject if violated, ensuring that 'verified' status is backed by actual test results rather than vague descriptions.

Key Takeaways

  • The myzubster.zorgax-capability.v1 schema enforces strict boundaries on AI actions and claims based on evidence states.
  • Only one initial capability, the Research Evidence Navigator, is granted TESTED status; others remain DOCUMENTED.
  • A new CI validator rejects pull requests where capability definitions lack linked bounded test evidence.
  • The model prohibits AI from inferring qualifications or employment status from self-declared contributor interests.

The Bottom Line

This is a pragmatic approach to AI governance that prioritizes evidence over confidence. By making capability states machine-validatable and CI-enforced, MyZubster is building a trust layer that most AI projects are currently too lazy to implement.