Building an evidence-bound capability model for Zorgax
I’m working on a new MyZubster milestone around a question that keeps coming up in AI systems:
How do you define what an assistant is actually allowed to know, claim, and do?
Not in prompt text.
Not as an informal convention.
But as something machine-readable, testable, and enforceable in CI.
That is the purpose of PR #1573.
The core separation
I’m separating two concepts that are often mixed together:
Contributor competence / evidence
and
AI capability
A contributor record should describe things like:
- repository / PR / commit provenance;
- documented or tested work;
- evidence state;
- explicit limitations;
- consent boundaries.
A Zorgax capability should describe:
- allowed input sources;
- allowed actions;
- prohibited claims;
- minimum evidence requirements;
- linked checkpoints;
- limitations;
- capability state.
That means contributor expertise does not automatically become an unrestricted AI claim.
Capability schema
The first version uses:
myzubster.zorgax-capability.v1
A capability looks conceptually like this:
source evidence
↓
allowed inputs
↓
allowed actions
↓
must-not rules
↓
evidence requirements
↓
linked checkpoints
↓
capability state
Initial capabilities:
zorgax.contributor-profile-assistant
→ DOCUMENTED
zorgax.research-evidence-navigator
→ TESTED
zorgax.interoperability-checkpoint-assistant
→ DOCUMENTED
The asymmetry is intentional.
Documentation does not equal testing.
Example: contributor profiles
The Profile Assistant may:
- summarize declared interests;
- prepare profile drafts;
- preserve provenance;
- suggest bounded onboarding steps.
But it must not:
- invent qualifications;
- infer certification;
- infer employment;
- infer Passport / LIFE / DAO participation;
- publish contributor information without the required approval.
So a self-declared interest stays self-declared.
It does not silently become a verified competence.
Example: research evidence
The Research Evidence Navigator has a stricter evidence-preservation rule.
If a source is:
SUPPORTED
then Zorgax retrieval must preserve:
SUPPORTED
It must not turn it into:
TESTED
or “scientifically validated”
or “clinically certified”.
This is one of the places where evidence semantics matter more than fluent output.
Machine-validatable capability states
I also added a registry validator.
It checks:
- JSON Schema validity;
- duplicate capability IDs;
- registry/file consistency;
- capability status consistency;
- whether a
TESTED capability actually links to TESTED bounded checkpoint evidence.
The validator is wired into the Continuous Evidence Gate.
So the intended progression becomes:
PROPOSED
→ DOCUMENTED
→ IN_VERIFICATION
→ TESTED
with DISABLED available when a capability should not be used operationally.
Why I think this matters
A common AI architecture pattern is:
model + tools + instructions
But that still leaves a governance problem:
Which tool-assisted claims are actually supported?
I’m moving toward:
evidence
→ provenance
→ capability
→ guardrails
→ bounded test
→ observed result
→ state
This gives us a way to ask:
- Which source supports this?
- Is it documented, supported, recorded or tested?
- Is the assistant allowed to act on it?
- Has that specific capability actually passed a bounded checkpoint?
- What is explicitly outside scope?
Broader direction
This now connects with the contributor evidence work already implemented in MyZubster:
Contributor work
→ machine-validatable interoperability checkpoint
→ provenance / competence records
→ Zorgax capability registry
→ bounded capability tests
→ optional Passport / Knowledge Graph workflows
without collapsing technical evidence, contributor identity, consent, publication, certification, payment and governance into one generic “verified” state.
The design principle is:
AI capability should be evidence-bound, not confidence-bound.
PR #1573 is the first implementation of that capability layer.
Once merged, the next step is to expand from static registry validation toward automated evidence ingestion and capability-state transitions driven by bounded checkpoints.