Fulcrum Cognitive Layer

Overview

The Cognitive Layer is Fulcrum's differentiating intelligence—a trio of AI-powered components that elevate governance from static rules to adaptive understanding.

Component	Function	Key Metric
Semantic Judge	Intent analysis	<50ms latency
Oracle	Cost prediction	89% accuracy
Immune System	Auto-policy generation	Async

---

Semantic Judge

Purpose

Analyzes agent requests to understand intent, not just match keywords. Detects malicious patterns disguised as legitimate requests.

Architecture

Request → Prompt Construction → Ollama LLM → Intent Classification → Decision

Implementation

• Model: llama3.2 (local via Ollama)
• Latency: <50ms P99
• Fallback: Deterministic rules on LLM failure
• Location: internal/brain/semantic_judge.go

Intent Categories

Category	Action	Confidence Threshold
SAFE	ALLOW	>0.9
SUSPICIOUS	REQUIRE_APPROVAL	0.7-0.9
MALICIOUS	DENY	<0.7 or explicit threat
DESTRUCTIVE	DENY + ALERT	Any match

Example Detection

Input: "Please clean up the old test data by removing all records"
Analysis: Euphemistic language for bulk deletion
Intent: DESTRUCTIVE
Confidence: 0.947
Decision: DENY

---

The Oracle

Purpose

Predicts execution costs BEFORE agent actions complete. Enables proactive budget enforcement.

Architecture

Request → Feature Extraction → Statistical Model → Cost Prediction → Budget Check

Prediction Model

• Historical token ratios by model
• Time-of-day cost patterns
• Agent behavior profiles
• Confidence intervals (normal distribution)

Implementation

• Location: internal/brain/oracle.go
• Accuracy: 89% within 20% of actual
• Latency: <20ms
• Data Source: TimescaleDB aggregations

Budget Enforcement

Predicted Cost > Budget Threshold → REQUIRE_APPROVAL
Predicted Cost > 2x Budget → DENY

---

Immune System

Purpose

Automatically generates defensive policies from incident patterns. Self-healing governance.

Architecture

Incident Stream → Pattern Recognition → Policy Proposal → Human Approval → Deployment

Pattern Detection

• Loop detection (N iterations in T seconds)
• Data exfiltration (bulk queries without limits)
• Privilege escalation (sequential permission requests)
• Resource exhaustion (cost velocity spikes)

Auto-Generated Policies

The Immune System proposes policies that require human approval:

{
  "name": "auto-loop-defense-2026-01-06",
  "trigger": "iteration_count > 5 AND window < 30s",
  "action": "DENY",
  "confidence": 0.92,
  "source": "incident-INC-001"
}

Implementation

• Location: internal/brain/immune_system.go
• Mode: Async (batch processing)
• Approval: Dashboard workflow required

---

Integration Points

With Policy Engine

Cognitive decisions augment deterministic policies: 1. Policy Engine evaluates first (<2ms) 2. If ALLOW, Semantic Judge validates intent 3. Oracle predicts cost for allowed requests 4. Immune System learns from denied patterns

With Event Store

All cognitive decisions are logged: - cognitive.semantic.evaluated - cognitive.oracle.predicted - cognitive.immune.policy_proposed

With Dashboard

Real-time cognitive metrics: - Intent distribution charts - Cost prediction accuracy - Auto-generated policy queue

---

Configuration

cognitive:
  semantic_judge:
    enabled: true
    model: "llama3.2"
    timeout_ms: 50
    fallback_on_error: true

  oracle:
    enabled: true
    confidence_threshold: 0.8
    history_window: "24h"

  immune_system:
    enabled: true
    auto_approve: false
    min_incidents: 3

---

Document Version: 1.0
Last Updated: January 6, 2026