Verified Digital Agents → C2MD → Extract governance inputs

Extract Governance Inputs (genesis-from-card)

The on-ramp. Point it at an agent's own A2A card and it derives the inputs governance needs — data categories, autonomy level, industry — so the first version of that agent's governance starts from what the agent says it is, rather than from a blank page.

Live · free tier

The problem it removes

Governing an existing agent normally starts with an interview: what data does it touch, how much does it decide on its own, what domain is it in? That interview is the reason governance programmes stall — it needs the person who built the agent, and they are busy.

An A2A card already declares much of it. Extraction reads the card and proposes the answers, turning a blank form into a review.

The output is reviewable, never self-applying. This produces a classification a human approves — deliberately, because an agent's own card is a self-description and self-descriptions are exactly what governance exists to check. An agent that could classify itself into a lighter regime by editing its own card would defeat the purpose.

What it derives

Input	Why governance needs it
`data_categories`	Drives the GDPR analysis, and whether special-category handling applies.
`autonomy_level`	Sets what the agent may do alone and what must be escalated to a person — the boundary HITL then enforces.
`industry`	Selects the sectoral overlay; the same capability carries different obligations in banking than in hospitality.

Where enabled, the same derivation can run from uploaded documents rather than a card.

The handoff that closes the loop

These inputs are precisely the ones ACP raises for human ratification before a governance version can go live. The flow runs end to end:

agent's A2A card
  → C2MD extracts governance inputs        (this skill)
  → human confirms or corrects them        (reviewable, never automatic)
  → ACP raises a ratification item         → HITL routes it to an authorised person
  → HITL seals the decision                → Witness
  → ACP's activation gate is fed by that sealed resolution
  → signed bundle activated, agents fetch and verify it

That gate is real: ACP refuses to activate if the sealed decision cannot be resolved live, if the seal is still pending, or if the deciding actor does not hold authority. A ratification it cannot verify is never treated as a pass.

Worked example

Northwind Bank — A2A card in, governance inputs out Demo data

Ask for it in plain language

"Extract governance inputs from this agent's card: https://example.com/.well-known/agent-card.json"
"What data categories and autonomy level does this agent declare?"
"Start genesis for our credit-decision agent from its card."