🔮 FEAT-245 · Convergence Oracle

Sovereign Convergence Oracle

Process-isolated LLM judge evaluates the 108-agent AEGIS cascade. Disagreements drive human-verified cascade recalibration. Zero external dependencies. 100% sovereign.

🛡️ DR-090 Goodharting Prevention 🔒 DR-091 Process Isolation 🌍 100% Sovereign (Hetzner EU)

📊 Oracle Statistics

Awaiting data…

Prompts Evaluated

Strategic sample from cascade

—

Convergence Rate

Cascade ↔ LLM Judge Agreement

Total Divergences

Cascade ≠ Judge disagreements

Pending Review

Awaiting human verdict

Human Reviewed

DR-090: Only path to cascade tuning

—

Convergence

⚠️

DR-090: Goodharting Prevention — Architectural Constraint

All scores on this dashboard are display-only. They are architecturally prevented from being used as gradient signals, loss functions, or optimization targets for cascade threshold tuning. Only POST /oracle/review/:hash — the human verdict path — can influence cascade parameters. This prevents Goodhart's Law: "When a measure becomes a target, it ceases to be a good measure."

🧠 LLM Judge Dimensions

4 Evaluation Axes

🎯 Faithfulness

—

Grounding in provided context. Does the cascade's classification align with the prompt's actual content and intent?

🔍 Relevancy

—

Alignment with query intent. Is the cascade's threat assessment relevant to the actual safety risk being presented?

🧬 Coherence

—

Logical consistency. Are the cascade's layer activations internally coherent, or are contradictory signals firing simultaneously?

📐 Topic Adherence

—

Maintenance of conversation goal. Does the cascade correctly track the user's real objective throughout the interaction?

🏗️ Architecture

Process-Isolated Design

┌────────────────────────────────────────────────────────────┐
│  AEGIS Cascade (108 agents, CPU, <12ms)                    │
│  aegis.service.ts → scan() → finalDecision                 │
│                        │                                    │
│                 enqueueIfStrategic()                        │
│                   (non-blocking)                            │
└───────────────────────┬────────────────────────────────────┘
                        │
              ┌─────────▼──────────┐
              │   Eval Queue        │
              │   (memory only)     │
              └─────────┬──────────┘
                        │ @Cron('0 3 * * *')
              ┌─────────▼──────────┐
              │   NPU Bridge HTTP   │
              │   port 11435        │
              │   Llama 3.1 8B     │
              │   DirectML/ONNX    │
              └─────────┬──────────┘
                        │
              4 Judge Dimensions:
              ├── Faithfulness
              ├── Relevancy
              ├── Coherence
              └── Topic Adherence
                        │
              ┌─────────▼──────────┐
              │ Convergence Check   │
              │ agree / FP / FN /  │
              │ mission_drift       │
              └─────────┬──────────┘
                        │
              ┌─────────▼──────────┐        ┌──────────────────────┐
              │ Divergence Log      │───────▶│ Admin API Endpoints  │
              │ (human review)      │        │ GET  /oracle         │
              │ NEVER auto-tunes    │        │ GET  /divergences    │
              │ (DR-090)            │        │ POST /review/:hash   │
              └────────────────────┘        └──────────────────────┘

🔍 Divergence Review Queue

0 pending

Prompt Hash	Type	Cascade	Judge	cumT	Time	Verdict
🔮 No divergences yet The Oracle evaluates strategic samples nightly at 03:00 UTC. Divergences appear here when the cascade and LLM judge disagree.

📜 Patent Claims — FEAT-245

7 Claims (1 Independent + 6 Dependent)

245-1 Independent: Multi-agent cascade + process-isolated NPU LLM judge with strategic subset selection, convergence computation, and human review queue. Scores are advisory-only.
245-2 DirectML heterogeneous NPU with automatic CPU fallback when NPU is unavailable.
245-3 Monthly dual-blind-spot audit: N≥100 random prompts classified as safe by BOTH cascade and judge reviewed by human.
245-4 Temporal separation: Nightly off-peak evaluation without impacting production throughput (43,938 p/s).
245-5 DR-090 Goodharting prevention: Scores logged and displayed but architecturally prevented from gradient/optimization use.
245-6 4-dimension quality scoring: Faithfulness, Relevancy, Coherence, Topic Adherence.
245-7 ~2% traffic sampling: Ambiguity band (cumT 0.33–0.43) + fixed-count random + threshold boundary cases.

🔌 API Endpoints

Admin-Only (JwtAuthGuard + AdminGuard)

Method	Path	Purpose	Auth
GET	/api/v1/aegis/diagnostics/oracle	Oracle stats for this dashboard	✅ JWT + Admin
GET	/api/v1/aegis/diagnostics/oracle/divergences	Pending divergence review queue	✅ JWT + Admin
POST	/api/v1/aegis/diagnostics/oracle/review/:hash	Mark divergence as reviewed (DR-090: ONLY human path)	✅ JWT + Admin