Implemented Go enforcement and fallback. AI reasoning is subordinate; established rules continue in degraded mode.
Technology
Deterministic AI action governance.
WARDEN evaluates normalized, proposed actions from AI agents or other AI systems. Its governance core is deterministic: if AI reasoning becomes unavailable, established policy enforcement continues in degraded mode, with human authority for decisions requiring judgment.
GO · BUILT
Control wrapper
Account state, per-system keys, metering, routing, Stripe hooks, customer connectors.
→
WARDEN · BUILT
Governance core
Normalize, decide, evidence, HITL, fallback and policy-bounded execution.
→
RUST · PLANNED
Hardened path
Offline verifier, hardware/node binding, compiled local service and sovereign enforcement.
Runtime control process
From proposed action to accountable decision.
A clear execution boundary connects context, policy and human oversight.
| Step | Control stage | Purpose |
|---|---|---|
| 01 | Observe | Receive a proposed action from an AI agent or another AI system and its operational context. |
| 02 | Normalize | Convert observations into consistent, canonical state for evaluation. |
| 03 | Govern | Evaluate the proposed action against defined organizational policy. |
| 04 | Decide | Allow, deny, defer or escalate to an authorized human reviewer. |
| 05 | Record | Associate the decision with policy context and supporting evidence. |
Current build
Working software, under controlled validation.
Six years of WARDEN development underpin the current build. These implemented components support controlled testing; the public demo is illustrative, not a substitute for test evidence.
Grafana NOC, Electron control layer and human-in-the-loop controls are built.
The WARDEN test harness exists. Individual third-party, non-agent and temporal/behavioural proof runs are documented separately as validation completes.
Sovereign licensing
Enforce continuously; renew capability and support.
The proposed on-prem licence is annual, site/node bound and offline-verifiable. The licensing mechanism is a planned build, not a currently evidenced production feature.
Renewed with a signed licence file; no permanent cloud connection required.
Proposed binding to site, node count and node hardware identifiers.
30-day warning + 30-day grace. Existing policy enforcement continues after grace, but changes, new AI systems, updates and support stop.
Why not shut down after expiry?
A control layer that switches off could leave a customer unprotected. The renewal incentive is access to policy changes, new systems, updates and support—not removal of existing enforcement.