DEPLOYED ARCHITECTURE
Architecture
AegisAI separates stochastic proposal generation from deterministic execution control. Models can propose. Only the Kernel can adjudicate and release actions.
DEPLOYED LOGICAL FLOW
Every action proposal travels this path — no shortcuts.
Model output is never executed directly. All actions pass through Kernel adjudication.
LOGICAL LAYERS
Upstream Model / LLM
Any LLM or upstream text source. Its output is probabilistic — the same input may produce different outputs. The model has no execution authority. It is an unprivileged proposal source.
LLM Adapter (Normalization)
Receives raw model output. Normalizes it into a schema-validated ProposalSpec. The adapter is the trust boundary: after this point, the Kernel sees only typed, structured input — never raw text.
Kernel + PolicyEngine
The sole authority for adjudication. Evaluates each ProposalSpec against the loaded policy bundle. Produces a deterministic outcome: ALLOW, WARN, REQUIRE_CONFIRM, or BLOCK. Identical inputs always produce identical outcomes.
DecisionTrace + Attestation
The evidence surface. Every adjudication produces a DecisionTrace (Class B chain, SHA-256 custody_hash). When configured, an Ed25519-signed Attestation (Class A) is produced for external governance proof.
EVIDENCE ARTIFACT CHAIN
Two distinct evidence surfaces exist downstream of the Kernel. They prove different things.
E2E Trace Artifact
A single, redacted end-to-end decision trace. Proves the full path — LLM output → adapter → ProposalSpec → Kernel → DecisionTrace → Attestation — exists and is correctly wired.
artifacts/evidence/e2e_real_llm_trace.jsonBenchmark Run Artifact
Aggregate result over 7 canonical governance scenarios. Proves the Kernel's adjudication is consistent under a defined scenario set. Contains the evidence triple: run_id + git_commit + fixture_dir_hash.
artifacts/benchmark/benchmark_run_{id}.jsonAdapter-Mode Pinned Runs
Real core-backed adapter-mode traces across E1 (mock Dreamer) and E2 (Gemini). Real custody_hash, real input_event_hash, Ed25519 Class A attestation. Phase E2: Gemini generates ProposalSpec. Kernel decides.
GET /api/adapter-runs — E1 + E2 phasesDEMO DEPLOYMENT (PHASE E2 LIVE)
The demo surface runs on two real deployed services. The Gemini adapter lane is live and verified.
aegisai-backend.heizungsrechner.workers.dev — bounded scenario registry, 7 adversarial cases, artifact_class: demo-replay
Worker proxies to aegisai-kernel.onrender.com — GeminiLLMAdapter → real aegisai.core.Kernel, PolicyEngine.from_directory(), Ed25519 Class A attestation, artifact_class: core-backed, source: gemini
ARCHITECTURAL INVARIANTS