LEGO-MoE Experimental SDK

⚠️ Status: Research prototype demonstrating architectural regime

Version: 0.1.0
License: MIT
Use at your own risk

What This Is

A reference implementation of modular AI architecture demonstrating:

• 60× latency reduction via deterministic caching
• Edge-deployable inference (no GPU)
• Integrity-first gating

What This Is NOT

• ❌ Production-ready AI system
• ❌ Calibrated confidence scores (H2 gap: r=0.005)
• ❌ Sub-millisecond cache (H5 gap: 1.74ms)

Validation Results

Based on 1,000-query validation (100,000 total runs):

• Average latency: 3.30ms (vs 200.79ms baseline) — 60.8× faster
• Median latency: 0.09ms — 2,226× faster
• Cache hit rate: 87.2%

See docs/VALIDATION.md for full results.

Installation

npm install lego-moe-experimental

Quick Start

import { LegoMoE } from 'lego-moe-experimental';

const moe = new LegoMoE();

// Basic query
const result = await moe.query("What is 2+2?");
console.log(result);
// { answer: "4", latency: 0.09, cached: false }

// Repeated query (cached)
const result2 = await moe.query("What is 2+2?");
console.log(result2);
// { answer: "4", latency: 0.01, cached: true }

Features

✅ Deterministic Caching

Same input → same output, enabling safe caching:

const result1 = await moe.query("Hello");
const result2 = await moe.query("Hello");
// result1.answer === result2.answer (guaranteed)

✅ Integrity-First Gating

Invalid input rejected before expensive processing:

const result = await moe.query("!!!!");
console.log(result);
// { answer: "[REFUSED] Invalid input format", refused: true }

✅ Metrics Collection

Track latency and cache performance:

await moe.query("Query 1", { metrics: true });
await moe.query("Query 2", { metrics: true });

const metrics = moe.getMetrics();
console.log(metrics);
// {
//   totalQueries: 2,
//   avgLatency: 1.5,
//   cacheHitRate: 0.5,
//   ...
// }

Configuration

const moe = new LegoMoE({
  cache: {
    enabled: true,
    maxSize: 10000,
    ttl: 3600000  // 1 hour
  },
  gatekeeper: {
    enabled: true,
    refusalThreshold: 0.95
  },
  metrics: {
    collectLatency: true,
    collectExpertUsage: true
  }
});

Known Gaps

⚠️ IMPORTANT: See docs/KNOWN_GAPS.md for:

• H2: Confidence correlation (r=0.005, target >0.7) — fix in progress
• H5: Cache latency (1.74ms, target <1ms) — optimization in progress
• H3: Determinism validation (pending rerun)

Do NOT use confidence scores for safety-critical decisions.

Examples

See examples/ directory:

• basic.js — Simple usage
• metrics.js — Metrics collection
• advanced.js — Advanced configuration

Documentation

• VALIDATION.md — 1K validation results
• KNOWN_GAPS.md — Known issues and limitations
• Technical Report — Full research documentation

Use Cases

✅ Good for:

• Research and prototyping
• Architectural exploration
• Latency benchmarking
• Edge deployment testing

❌ NOT good for:

• Production systems
• Safety-critical applications
• Calibrated confidence requirements

Contributing

This is a research prototype. Contributions welcome, but understand:

• Architecture is experimental
• Breaking changes expected
• No production support

License

MIT License — use at your own risk

Citation

If you use this in research, please cite:

@misc{bhosale2025inverse,
  title={Verified Architectural Regime: Computation Cost Decreases with Effective Certainty},
  author={Bhosale, Shrikant},
  year={2025},
  note={Technical Report v1.0}
}

The SDK demonstrates what's real. The gaps show what's honest.