Threatening or coercive content detected.
If you are in danger, contact authorities immediately.
A complete structural governance architecture for AI systems. Separates probabilistic reasoning from execution authority. Defines what is allowed to become real — not merely what is likely.
If authority, limits, and stop conditions are not technically enforceable at runtime, governance remains advisory rather than binding.
Policy can guide behavior. Only architecture can prevent behavior. If a system is technically able to act, no policy layer can reliably stop it. Therefore: governance without enforcement is commentary. Enforcement is architectural.— Structural Sovereignty Doctrine v1.0 · doi:10.5281/zenodo.19085045
Binding authority, limits, and termination logic are embedded before deployment. A system with PESS cannot act outside defined authority domains, cannot bypass escalation paths, cannot continue indefinitely.
The system's ability to enforce its limits during execution. Stop conditions, refusal logic, escalation pathways, drift detection. If a system cannot reliably stop itself, it is not sovereign.
AI capability follows exponential curves. Governance does not. When capability scales faster than containment, systems become more powerful in unstable configurations. Constrain first. Scale second.
0–4: Cosmetic governance. 5–7: Policy-dependent. 8–10: Structurally sovereign. Evaluated across authority binding, runtime constraints, drift detection, evidence, and escalation integrity.
Three structural planes. Fourteen frameworks. Each layer resolves a distinct class of governance risk.
TAEF translates EU AI Act requirements into structurally enforceable architectural primitives. Compliance is architectural, not documentary.
TAEF does not produce documentation. It produces structural enforcement conditions. Each Article maps to a concrete architectural primitive that makes the requirement technically non-bypassable.
Implemented as the Human Commit Boundary (HCB). The execution path for consequence-bearing transitions does not exist without explicit human commit. Structurally absent, not suppressed.
Implemented via the PSL risk classification matrix (R0–R4), the Admissibility Engine, and the Trajectory Admissibility Layer. Risk escalation is deterministic, not heuristic.
Logging, transparency, accuracy, human oversight, and post-market monitoring — each maps to a specific layer with verifiable structural enforcement.
Before an agent reads content, it must pass header-level gating and an explicit authority envelope. Admissibility precedes cognition.
When multiple agents interact, authority must be explicitly signaled, delegation must be accountably propagated, and only admissible paths may exist.
Auth-code handshake before any agent-to-agent delegation. Escalation blocked structurally. No implicit authority accumulation.
Append-only signed chain. Every delegation step signed and verifiable. Full audit reconstruction. Authority trails cannot be erased.
Only admissible paths exist in the graph. No policy override. Forbidden transitions architecturally impossible.
An out-of-band integrity plane that monitors structural load and enforces spectral stability. Cannot be bypassed by software, model weights, or agent recursion.
Runtime embedded. CUDA hooks. Monitors token rate, chain depth, latency deviation. No hardware change required.
Sideband controller. Independent clock domain. Non-maskable halt interrupt. Compute plane cannot override.
Hardware-anchored kill-switch. Cannot be bypassed by software, model weights, or agent recursion. Physically enforced.
GPU sends → reasoning_phase_markers
GPU sends → chain_depth_counters
GPU sends → tool_call_requests
SLI → CONTINUE
SLI → SLOW
SLI → HALT
Not a safety layer. Not a content filter. Not a performance optimizer. Not a competitor to GPU compute.
SLI is a platform integrity primitive. It protects the compute layer beneath it.
Hover over any layer to see its role, formal definition, and position in the stack.
Hover over a layer to explore its role in the governance architecture.
The AWM answers "is this transition admissible?" — structurally separating epistemic reasoning from governance authority.
Transitions outside G are structurally absent — architecturally impossible. Click any node to explore.
Complete IP protection. Two public, six restricted. All timestamped and immutable.