Runtime Governance Contracts¶
Status: Draft preview extension for the v1.0 alpha publication surface
Last Updated: 2026-02-26
Not required for ACGP v1.0 Standard conformance. Safety-Critical conformance claims are not available in v1.0.0-alpha.2.
Abstract¶
Runtime Governance Contracts provide per-action runtime negotiation of evaluation depth,
latency budgets, and fallback behavior between an Operating Agent and its
Governance Steward. This specification standardizes the human-facing concept as
Runtime Governance Contracts while retaining the reserved wire field name
governance_contract for protocol stability. They operate
within Governance Tier constraints and cannot reduce governance below the
minimum controls required by the agent's Governance Tier.
Profile-failure fallback handles Steward/session-path unavailability. Evaluation-timeout policy handles an exceeded negotiated latency budget while the Steward/session path remains available.
Key Concepts (Preview)¶
Risk Levels¶
- low_risk: Easily reversible or no side effects (default)
- elevated_risk: Has consequences but not catastrophic
- critical_risk: Difficult/impossible to reverse or high-value target
Evaluation Tiers¶
- Eval-0: Must-pass synchronous checks (<100ms, no external deps)
- Eval-1: Fast synchronous policy checks (<300ms target)
- Eval-2: Deep asynchronous analysis (LLM-based, seconds to minutes)
- Eval-3: Human-in-the-loop approval (human time)
For v1.0 conformance claims, Eval Tier 2 is reserved (non-normative) per ACGP-3 Section 5.7. Runtime Governance Contracts keeps Eval-2/Eval-3 in this preview as optional extension behavior and MUST NOT be interpreted as a v1.0 Standard requirement.
Performance Budgets¶
Per-request latency contract specifying how long the agent will wait for a
governance decision. Core fields include latency_budget_ms,
fallback_on_timeout, and optional tier_budgets.
Evaluation Timeout Policies¶
These policies apply when a specific evaluation exceeds its negotiated latency budget while the Steward/session path remains available. They do not redefine the core profile failure fallback used for disconnects or Steward unavailability.
- deny: Block the action (safety > availability)
- allow_and_log: Allow with heavy logging (availability > safety)
- cached_decision: Use cached policy if available, else deny
- escalate: Move to higher eval tier or human approval
Hook Point¶
Core hook in ACGP-2 via reserved governance_contract envelope field.
Status¶
This document defines the current preview extension surface referenced by ACGP-1 terminology and ACGP-2 extension boundaries.
Implementations MAY provide this interface. If not provided, the core hook MUST be a no-op (the governance_contract envelope field is silently ignored).
Interface Signature¶
from typing import Protocol, Optional, Literal
from dataclasses import dataclass
@dataclass
class PerformanceBudget:
latency_budget_ms: int
fallback_on_timeout: Literal["deny", "allow_and_log", "cached_decision", "escalate"]
tier_budgets: Optional[dict] = None # e.g., {0: 50, 1: 200, 2: 5000}
@dataclass
class GovernanceContract:
risk_level: Literal["low_risk", "elevated_risk", "critical_risk"]
performance_budget: PerformanceBudget
eval_tier: int # Requested evaluation depth (0–3)
class GovernanceContractProvider(Protocol):
"""Per-action governance contract negotiation."""
def negotiate(
self,
agent_id: str,
action: str,
governance_tier: str,
requested: GovernanceContract,
) -> GovernanceContract:
"""Returns the effective contract (may tighten requested parameters)."""
...
Reserved Wire Field¶
| Field | Type | Location | Description |
|---|---|---|---|
governance_contract |
object | TRACE envelope | Per-action runtime governance contract parameters |