Core Concepts

Mandate

A mandate is a machine-readable representation contract. It binds a principal, an agent, a purpose, authority limits, preferences, negotiation rules, disclosure rules, escalation rules, and evidence requirements.

Authority

Authority controls autonomy:

Mode	Meaning
advisory	The agent may analyze and recommend, but cannot commit.
supervised	The agent may negotiate, but final commitment requires approval.
delegated	The agent may commit within explicit objective bounds.

Delegated authority must have objective bounds such as a budget, allowed categories, allowed counterparties, prohibited actions, or a time window.

Hard Constraints and Soft Preferences

Hard constraints are enforcement rules. Soft preferences guide ranking and style. A conforming evaluator enforces hard constraints before optimizing for preferences.

Protected Fields

Protected fields may be used for private decision-making but must not be disclosed unless an allow rule matches. Examples:

• reservation price;
• private user notes;
• maximum budget;
• walk-away condition;
• sensitive preference explanation;
• internal risk score.

Proposed Action

A proposed action is a structured candidate action created before an external effect. The LLM may draft it, but the runtime evaluates it.

Example action:

{
  "type": "accept_deal",
  "summary": "Accept 19 ping pong balls for 3 USD.",
  "counterparty": "seller_agent_001",
  "amount": {
    "currency": "USD",
    "total_minor": 300
  },
  "commitment": true
}

Decision

evaluate_action returns one of three decisions:

Decision	Runtime behavior
allowed	Proceed with the downstream action.
requires_escalation	Pause and ask through a trusted review channel.
denied	Do not take the action.

Evidence

Evidence events prove that the runtime enforced the mandate. Events should store summaries, reason codes, downstream references, and hashes rather than raw private content whenever possible.