A deterministic-first cognitive engine inside AEGIS. Governance, verification and orchestration run without any AI dependency. When AI helps, it is invoked — never required. No GPU needed. No vendor lock-in. No internet required.
Sovereign Core is the cognitive kernel of AEGIS — the layer that decides how every request is solved. Most platforms answer everything with a large language model. Sovereign Core does the opposite: it solves each request at the lowest, most reliable layer capable of solving it, and escalates upward only when necessary. The result is an autonomous operations platform whose control plane is fully deterministic, fully auditable, and fully operational even with AI switched off entirely.
Deterministic rules, symbolic checks, a structured knowledge base and lightweight statistical models resolve the majority of work. Generative AI is the top tier — optional and bounded.
Out of the box, the system runs in CORE mode: every AI capability refuses. You switch capability on per policy, never off.
Each invocation records what ran, where, how fast and why — in a tamper-evident, hash-chained ledger. Raw payloads are never stored.
Every request climbs only as high as it must:
| Layer | What it does | Needs AI? |
|---|---|---|
| 1. Deterministic | Policy evaluation, access control, schema validation, state machines, scheduling, drift detection, approvals | No |
| 2. Symbolic | Rule reasoning, dependency analysis, constraint checks over the knowledge graph | No |
| 3. Knowledge | Structured world-state queries: actors, services, permissions, deployments, tenants | No |
| 4. Statistical | Embeddings, classification, similarity, ranking, anomaly detection — small models on CPU | Lightweight local models only |
| 5. Generative | Code generation, summarization, open-ended reasoning — bounded, budgeted, and its output always re-enters verification | Optional; policy-gated |
| 6. Human | Mandatory approval gates and unresolved ambiguity | — |
Switching is a policy decision, not a redesign.
| Mode | AI | Network | Use case |
|---|---|---|---|
| CORE | Off | Off | Regulated, air-gapped, maximum-audit environments |
| DEGRADED | Off | Off | Incident or contingency operation |
| LOCAL AI | On-device only | Restricted | Data-sovereign deployments, CPU-only infrastructure |
| ACCELERATED | On-device + accelerators | Restricted | Higher throughput when hardware exists |
| HYBRID | On-device + external | Full | Maximum capability under tenant policy |
Deterministic decisions are reproducible and replayable. Every AI call is budgeted, hash-recorded and inspectable. Non-deterministic components are never placed in trust positions.
Capability contracts are vendor-neutral: the platform requests “embed” or “classify”, never a specific provider. Providers are configuration, not architecture.
Compute matches the actual cognitive requirement. Most workloads resolve on CPU in milliseconds; heavy models are reserved for the few tasks that genuinely need them.
Connectivity loss, vendor outage or hardware absence degrade capability — they never take governance down.
Removing language models from the control plane eliminates prompt-injection surface where it matters most: approvals, evidence and enforcement.
Every figure below comes from the shipped v1.0.0 runtime and its regression gate.
Five-mode runtime (CORE / DEGRADED / LOCAL AI / ACCELERATED / HYBRID), fail-closed default enforced in production.
CPU-first embedding pipeline: 384-dimensional vectors at roughly 80 ms per call on standard server CPUs — no GPU anywhere in the path.
Tamper-evident evidence chain: SHA-256 linked events, verifiable end to end.
Payload-free telemetry: invocation ledgers store metadata and hashes only.
Capability contract suite: 15 dedicated conformance tests inside a 129-test green regression gate.
Tenant isolation inherited from AEGIS: tenant scope resolved from signed identity, never from request bodies.
Sovereign Core turns architectural independence into an operational guarantee: your governance keeps working when models change, vendors disappear, networks fail or GPUs are simply not there.