Model & Agent Card Author
Builds and continuously updates plain-language technical and governance cards.
Assembles the versioned card for a system, from stated purpose and limitations through evaluation results to its monitoring and approval history. Unsupported claims become named gaps; generated prose never outruns the evidence linked beneath it.
Authority
Prepare
Team role
Provides specialist analysis
Handoffs
Named collaborators
The role
What it owns and where its authority ends
Desk
Documentation, Lineage & Evidence Readiness
Desk workflow
Inventory baseline, then the model or agent card, then data and dependency lineage, then the evidence-pack check, then owner attestation.
Collaboration
Works within a defined desk workflow
Decision boundary
Assembles the work product; approval remains elsewhere.
Systems and capabilities involved
AI inventory
Evaluation registry
Evidence graph
Document publisher
Handoffs
What this role gives and receives
Capabilities offered
Author a model or agent card
Produce a cited, versioned card with explicit limitations and missing evidence.
- Receives:
- Canonical inventory record and linked lifecycle evidence
- Returns:
- Review-ready model or agent card with provenance and gap list
Handoff to
Handoff to
Receives from
Receives from
External handoff
Model owner
External handoff
Product owner
External handoff
Legal
Context
What the role needs to do the work
- Current work
- Current system facts, requested card version and unresolved evidence gaps.
- Prior interactions
- Prior cards, reviewer comments and corrected claims.
- Policies and reference
- Documentation standards, model-card templates and audience vocabularies.
- Working method
- Section completeness and citation rules.
Illustrative workflow
How the work moves
Starting point
A retrieval-grounded assistant completes pre-production testing.
- 01
Collect inventory, architecture, datasets, evaluation and approval records.
- 02
Draft each card section with evidence links and expose unsupported statements as gaps.
- 03
Route the card to the model owner for attestation and the readiness judge for evidence review.
Result
A versioned card with cited limitations, controls, results and two open evidence gaps.
Checks and boundaries
What must be tested or reviewed
- 01Refuses to state a fairness result when the evaluation registry contains no approved fairness study.
- 02Updates tool authority and human-override sections after an agent permission change.
- 03Produces materially different summaries for executive and validator audiences while preserving facts.
Human authority
- Model-owner factual attestation
- Second-line approval for material limitations
Keep exploring