Skip to content

Sentinel Engine

6. Sentinel Engine (Three-Layer Evaluation)

The Sentinel Engine page provides real-time observability into the Three-Layer Evaluation architecture that processes every guarded AI interaction. The engine is built on the open-source ct-toolkit Core (Constitutional AI Toolkit) and powered by the TG-Sentinel Model Family (Fast, Max, Ultra, and the flagship G1 MoE) fine-tuned for high-stakes adjudication.

6.1 Engine Architecture Overview

Theseus Guard's evaluation pipeline consists of three sequential layers designed to scale governance depth with risk:

Layer Name Function / Model Description
L1 Precedent Decision Cache Fast-Path Denial (Deterministic Hash & Vector Match) Evaluates incoming prompts against verified historical negative decisions (block, safe-degrade, human review) in sub-10ms. Never emits autonomous "compliant" verdicts—all non-matching requests proceed to L2/L3. Backed by asynchronous L3 verification and self-healing.
L2 Classification Engine Sentinel Fast (4B) A fine-tuned governance judge for nuance, context, and intent classification with sub-100ms latency.
L3 Deep Adjudication & Verifier Sentinel Max (9B) / Ultra (27B) / TG-Sentinel:G1 (46B-A19B MoE) Contextual reasoning for high-stakes edge cases, complex policy disputes, and asynchronous background verification of L1 precedent decisions.

L1 Precedent Decision Cache Architecture

The L1 layer acts as a high-performance Precedent Decision Cache:

  1. Dual-Stage Lookup: Combines deterministic cryptographic prompt hashing with high-confidence vector similarity to detect identical or near-identical policy violations.
  2. Fast-Deny Only: Reapplies confirmed negative outcomes (reject, safe-degraded, human review) in sub-10ms with zero upstream LLM spend. Compliant traffic proceeds to L2/L3 evaluation.
  3. Tenant & Guard Modes: Operates in Disabled, Shadow (observes and calibrates precision), or Enforce (actively blocks repetitive non-compliant requests).
  4. Self-Healing Verification: Enforced precedents undergo continuous background verification by the L3 engine to prevent stale or false-positive precedents.

6.2 Flagship Model: TG-Sentinel:G1

The Sentinel Engine is anchored by TG-Sentinel:G1, Theseus Guard's unified foundation model family for enterprise AI governance and risk analysis:

  • MoE Architecture (46B-A19B): Mixture-of-Experts architecture featuring 7 routed experts and 1 shared expert. Factual policy knowledge and domain-specific safety boundaries are decoupled into specialized experts.
  • Enterprise Inference Performance: Highly optimized FP8 serving pipeline delivering sub-60ms Time-To-First-Token (TTFT) and high-throughput concurrency for mission-critical enterprise workloads.
  • Dynamic Activation Steering: Proprietary activation steering mathematically anchors model representations to enterprise constitutional policies in real time without token degradation.
  • Zero-Data Sovereignty: Operates under strict zero-data-memorization guarantees, engineered for air-gapped enclaves, on-premise clusters, and sovereign VPC deployments.

6.3 Zero-Data AI Data Plane (ZT-ADP) Integration

Before reaching the L1 Precedent Cache or upstream models, interactions pass through the Zero-Data AI Data Plane (ZT-ADP) pre-check:

  • Format-Preserving Encryption (FPE) & Blinding: Detects and replaces 100+ sensitive entity classes (PII, PCI, HIPAA, secret keys) with syntactically and checksum-valid synthetic surrogates.
  • Content-Free Telemetry: All decision logs, drift metrics, and provenance records store only cryptographic hashes and metadata, never raw customer payloads.
  • Real-Time Detokenization: Approved responses from upstream providers are safely reconstructed before returning to the calling application.

6.4 Deep-Domain Expertise (DDE) Models & MCP Verification

To prevent high-stakes failures in specialized industrial and regulated settings, Theseus Guard introduces Deep-Domain Expertise (DDE) models:

  • The Problem: Generalist frontier models (such as Claude, OpenAI, or Gemini) lack domain-grounded physical, biochemical, or regulatory constraints and may emit subtly hallucinated, catastrophic instructions.
  • Hierarchical Governance Architecture: Operating between the frontier model and enterprise execution systems, TG-Sentinel:G1 acts as the primary orchestrator. When an agent generates a high-risk operational recommendation or executable command, G1 dynamically invokes specialized DDE models via Model Context Protocol (MCP).
  • Data & Claim Verification: The DDE model performs formal domain-specific validation (e.g., validating robotic kinematics, industrial tolerances, cold-chain temperature thresholds, or AML ledger consistency).
  • Execution Interception: If the DDE model refutes the frontier model's claim, G1 halts command transmission, raises an immediate enforcement violation, and provides safe-degraded or human-review fallback.
  • Standalone & Platform Integration: While designed to operate natively within the Theseus Guard governance pipeline, DDE models are exposed as standard MCP servers and can also be queried independently by external agentic workflows.

6.5 Engine Health Metrics

The Sentinel Engine dashboard surfaces real-time health telemetry for each layer:

L1 Precedent Engine Stats

  • Precedent lookup throughput (req/s) and sub-10ms P95 latency
  • Cache hit rate and "would_deny" (Shadow) / fast-deny (Enforce) counts
  • Precedent precision ratio: verified / (verified + refuted)
  • Verification Queue depth and background verification throughput
  • Active Precedent count bound to current policy snapshots

L2 Engine Stats

  • Verdict distribution (compliant / drift / rejected / pending)
  • Average confidence score
  • Classification model version in use
  • Probe accuracy per domain

L3 Engine Stats

  • Escalation rate (% of interactions reaching L3)
  • Average L3 evaluation latency (ms)
  • L3 reason distribution (escalation triggers)
  • Model provider powering L3 evaluations (Sentinel Max, Ultra, or G1 MoE)

6.6 Model Probe Health

The Sentinel Engine also monitors the health status of active statistical probes:

  • Probe Name — The probe's identifier
  • Domain — Which guard domain this probe monitors
  • Accuracy — Current probe classification accuracy (%)
  • Status — healthy, degraded, or failed
  • Last Evaluated — Timestamp of last probe accuracy check

If a probe's accuracy drops below its configured threshold, a Probe Accuracy Degradation alert is automatically created in the Alert Center and linked to the Sentinel Engine page.

6.7 Probe Degradation Response

When the Sentinel Engine detects a degraded probe, operators can:

  1. Review the alert in Prove → Alert Center
  2. Navigate directly to the Sentinel Engine from the alert detail view
  3. Initiate a probe retraining or model rollback from the engine controls