@ai-operations/spark-engine

Self-Perpetuating Adaptive Reasoning Kernel — A closed feedback loop that makes CORD safety scoring learn from outcomes.

The Spark

Most AI safety systems use static rules. SPARK closes the loop: Predict → Act → Measure → Learn.

Step arrives → Predictor predicts outcome → CORD scores (with learned weights)
                                                ↓
                                          Step executes
                                                ↓
           LearningCore compares ← OutcomeTracker measures
                    ↓
          WeightManager updates (bounded by SENTINEL)

Core Modules

Predictor

Before each step, predicts the CORD score, expected outcome, and confidence.

import { Predictor } from '@ai-operations/spark-engine';

const predictor = new Predictor(sparkStore);
const prediction = predictor.predict(stepId, runId, 'gmail', 'send');
// { predictedScore: 35, predictedOutcome: 'success', confidence: 0.72 }

OutcomeTracker

After execution, measures what actually happened.

import { OutcomeTracker } from '@ai-operations/spark-engine';

const tracker = new OutcomeTracker(sparkStore);
const outcome = tracker.measure(step, runId, wasApproved);
// { actualOutcome: 'failure', signals: { succeeded: false, hasError: true } }

LearningCore

Compares prediction to reality and adjusts weights.

import { LearningCore } from '@ai-operations/spark-engine';

const core = new LearningCore(sparkStore);
const episode = core.learn(prediction, outcome);
// { adjustmentDirection: 'increase', reason: 'CORD scored 15 but action failed' }

AdaptiveSafetyGate

Wraps CordSafetyGate with learned weight multipliers.

import { AdaptiveSafetyGate } from '@ai-operations/spark-engine';

const gate = new AdaptiveSafetyGate(cordGate, weightManager);
const result = gate.evaluateAction('gmail', 'send', input);
// score adjusted by learned weight, decision may change

Safety Bounds (SENTINEL)

• All weights bounded ±30% of base (0.70–1.30)
• Destructive and financial categories can NEVER decrease below 1.0
• Minimum 3 episodes before any learning occurs
• EMA smoothing (α=0.1) prevents oscillation

License

MIT