@0nork/cr0n-engine

Federated AI Closed Learning Loop — Multi-model optimization engine with adaptive weight learning.

The core innovation: a federated AI approach where multiple models (Claude, OpenAI, Gemini, Grok) compete, and the system learns which model performs best per task type. More models connected = faster convergence.

Installation

npm install @0nork/cr0n-engine

Quick Start

import { createEngine } from '@0nork/cr0n-engine';

const engine = createEngine({
  models: {
    claude: { apiKey: process.env.ANTHROPIC_API_KEY },
    openai: { apiKey: process.env.OPENAI_API_KEY },
    gemini: { apiKey: process.env.GOOGLE_AI_KEY },
    grok:   { apiKey: process.env.XAI_API_KEY },
  },
  weights: savedContentWeights,      // From previous cycle (optional)
  modelWeights: savedModelWeights,   // From previous cycle (optional)
  maxTasksPerRun: 50,
});

const result = await engine.runCycle({
  pages: pageDataArray,
  completedActions: pastActions,
});

// result.plan          — Ranked tasks with AI-generated briefs
// result.weights       — Updated content weights (save for next cycle)
// result.modelWeights  — Updated model weights (save for next cycle)
// result.modelStats    — Per-model performance dashboard
// result.learningLog   — What changed and why

Two-Dimensional Learning

Dimension 1: Content Weights

Which metrics matter most for scoring opportunities:

{ impressions: 0.20, position: 0.20, ctrGap: 0.20, conversions: 0.20, freshness: 0.20 }

Adjusted per cycle based on action outcomes.

Dimension 2: Model Weights (per bucket)

Which AI model produces best results for each task type:

CTR_FIX:           { claude: 0.25, openai: 0.25, gemini: 0.25, grok: 0.25 }
STRIKING_DISTANCE: { claude: 0.25, openai: 0.25, gemini: 0.25, grok: 0.25 }

Starts equal. After evaluating outcomes, converges — e.g. "Claude is best for CTR_FIX, Gemini for LOCAL_BOOST."

Why More Models = Faster Learning

Models	Perspectives/Cycle	Learning Speed
1	1	Linear
2	2	2x
4	4	4x hypothesis testing per evaluation window

Architecture

src/
├── core/           — Types, constants, config
├── engine/         — Scoring, bucketing, analysis
├── federation/     — Multi-model orchestration
│   ├── adapters/   — Claude, OpenAI, Gemini, Grok
│   ├── router      — Task-to-model routing
│   ├── consensus   — Weighted aggregation
│   └── model-weights
├── briefs/         — Content brief generation
├── learning/       — Dual-dimension weight learning
└── data/           — GSC/GA4 data normalization

5 Action Buckets

Bucket	Trigger	Action
CTR_FIX	High impressions, low CTR	Meta/intro rewrite
STRIKING_DISTANCE	Positions 4-10	Content expansion
RELEVANCE_REBUILD	Positions 11-50, stale	Full refresh
LOCAL_BOOST	Local intent	Location optimization
MONITOR	Top 3 positions	Protect rankings

API

createEngine(config)

Creates an engine instance with available models.

engine.runCycle(input)

Full cycle: evaluate past actions → score & classify → federate across models → learn → return plan.

engine.analyze(pages)

Local-only analysis without federation (fast, no API calls).

engine.federateBrief(page, bucket)

Generate a single federated brief for a specific page.

Dependencies

• ai — Vercel AI SDK unified interface
• @ai-sdk/anthropic — Claude adapter
• @ai-sdk/openai — OpenAI + Grok adapter
• @ai-sdk/google — Gemini adapter

Zero external runtime dependencies beyond the AI SDKs.

License

MIT