FOR ARCHITECTS
Lakhowal is a deterministic enforcement layer, not a model wrapper. You deploy it once. Your predicates are logic. Your boundaries hold.
SYSTEM DESIGN
The AI model and the enforcement boundary are causally independent. The model cannot know about LUIPM's decision. The decision cannot be influenced by the model's request.
Given the same input action, LUIPM always produces the same decision. No randomness. No internal state drift. You can reason about the system.
On any predicate failure, the action is blocked. SAFE_STATE is the default. No compensations. No majority voting. Hard boundary.
The 100ms latency budget is architectural, not aspirational. Each layer has a fixed latency ceiling. If a predicate exceeds its budget, the action is blocked.
Every decision produces an ERTuple committed to the Hydra Ledger. The ledger is append-only. Chain integrity is cryptographically verified.
DEPLOYMENT TOPOLOGIES
Lakhowal runs in-path between the AI model and your action system. Every action is evaluated before execution.
Your application imports the Lakhowal SDK and wraps AI action calls. Evaluation happens locally.
Lakhowal runs as a non-blocking observer. Actions fire normally; permits are logged in parallel.
Lakhowal runs in your VPC with no egress. Full computational autonomy. Highest assurance.
PREDICATE DEVELOPMENT
Predicates are functions that evaluate an action and return a boolean: permitted or denied. They can be written in Python, Go, JavaScript, or any JVM language.
Deploy your predicate to the Laboratory sandbox. Run historical action logs through it. See counterfactual permits. Tune logic without affecting production.
Predicates can be deployed as:
Predicates are versioned. You can enable/disable versions without redeployment. If a predicate starts denying too many actions, roll back instantly. Full history is immutable.
OBSERVABILITY
Real-time histograms of evaluation latency (p50, p95, p99, p99.9). Identify slow predicates.
How many actions are approved vs denied? By action type? By time of day? Spot drift in approval patterns.
Query the Hydra Ledger in real time. Understand causal chains. Debug permit decisions. Build compliance reports.
I_φ trending up? Model reasoning is becoming unstable. Trending down? Reasoning is stabilizing. Use for anomaly detection.
NEXT STEPS
What decisions will AI make? (e.g., approve_transfer, deny_loan, adjust_threshold). Each needs its own predicate set.
Start simple: Is the action within policy bounds? Use the Laboratory to test and iterate.
See what permits would be issued in production without blocking real actions. Build confidence in your predicates.
Enable enforcement. Every action now passes through LUIPM. Boundaries hold. Evidence is immutable.
Ready to start building? The Laboratory is ready for your first predicate.
Access the Laboratory