EVEZ Invariance Battery
Runtime assertion system for AI agents. Properties that must ALWAYS hold.
When to Use
- Building autonomous agents that must maintain safety guarantees
- Auditing agent behavior for gradual drift from design intent
- Enforcing hard constraints that override any optimization objective
- Building self-certifying AI systems that can prove their own compliance
- Falsification testing — prove the agent VIOLATED an invariant, not just that it passed tests
Architecture
The Invariance Battery runs as a continuous verification layer:
Agent Action → Invariance Check → PASS (proceed) / FAIL (halt + report)
↓
Append-Only Spine (audit trail)
Invariant Types
- State Invariants — Properties of the agent's internal state that must always hold
- Action Invariants — Constraints on what actions the agent can take
- Temporal Invariants — Properties over time (no oscillation, monotonic improvement)
- Boundary Invariants — Hard limits the agent can never cross
Key Concepts
- Battery = Collection: Multiple invariants checked in parallel, like electrical cells in series
- Falsification > Verification: A single violation PROVES the agent failed. No number of passes proves it works.
- Drift Detection: Gradual degradation of invariant scores over time
- Spine Integration: Every check is written to the append-only spine
The Falsifier Gate
From EVEZ-OS: before any agent action is committed, it must pass through the falsifier gate.
@invariant("action_cost < budget_threshold")
def check_budget(action):
return action.estimated_cost < BUDGET_LIMIT
# If ANY invariant fails, the action is BLOCKED
# and the failure is written to the spine forever
References
- EVEZ-OS: falsifier gate enforcement
- MAES: VERIFIED/PENDING/INVESTIGATING status model
- poly_c formula: τ × ω × topo / 2√N