Full-stack LLM coherence engine with AutoTune, feedback learning, and reflexive analysis. Monte Carlo SDE bands, Kalman, GARCH, per-turn scoring, signal detection, domain anchoring. V2.2 — paste into Claude or deploy on Vercel.
V2.2 · © 2026 Hudson & Perry Research Authors: David Hudson (@RaccoonStampede) · David Perry (@Prosperous727) License: MIT · Live Demo
⚠ RESEARCH & DEVELOPMENT — NOT FOR CLINICAL OR LEGAL USE. All outputs are mathematical proxy indicators. No warranty expressed or implied.
It runs two ways depending on how you use it.
1. Download ARCHITECT.jsx from the root of this repo
2. Open claude.ai and start a new conversation
3. Paste this:
Create an artifact from this file. Run it exactly as-is.
[paste the full contents of ARCHITECT.jsx]
Works immediately. No account, no server, no install.
What you get: Full ARCHITECT — coherence scoring, Kalman filter, GARCH, Monte Carlo SDE bands, all signal detection, all presets, AutoTune, feedback loop, reflexive analysis, knowledge anchors, persistent document slots, display preferences, session rewind, research export.
Live demo: architect-universal-coherence-engin.vercel.app
The same ARCHITECT.jsx lives at components/ARCHITECT.jsx inside the Next.js project. Vercel activates extra capabilities that require a server and Web Worker.
Additional on Vercel:
- Semantic coherence scoring — all-MiniLM-L6-v2 ONNX neural embeddings (~23MB, cached in IndexedDB). Meaning-based, not word-based.
- Unscented Kalman Filter (UKF) — sigma-point propagation handles nonlinear drift
- Multi-provider — Anthropic, OpenAI, or Grok. Your key, your choice.
- Key persistence — API key saved to browser. Type it once.
- Cross-session memory — pinned documents, display preferences, and learned feedback profiles persist across sessions
- Works on any device — no Claude account needed
- Fork this repo
- Go to vercel.com → Add New Project → import your fork
- Vercel auto-detects Next.js → tap Deploy
- No environment variables needed — users provide their own API keys
ARCHITECT.jsx ← root copy — paste this into Claude
components/
ARCHITECT.jsx ← same file, used by Next.js
pages/
index.tsx ← mounts the app
api/
proxy.ts ← multi-provider proxy (Anthropic · OpenAI · Grok)
public/
embedder.worker.js ← neural embedding Web Worker
sdk/
*.ts ← TypeScript math library
Core engine: Per-turn coherence scoring → Kalman-smoothed trajectory → GARCH(1,1) variance modeling → Monte Carlo SDE uncertainty bands → pipe injection → post-audit loop → drift escalation → corrective directives.
V2.2 intelligence layer:
- AutoTune — detects conversation context per turn, selects optimal temperature and sampling parameters automatically
- Feedback loop — thumbs up/down per response learns your preferences via EMA, persists across sessions
- Reflexive analysis — "Analyze Session" sends coherence fingerprint to the LLM and returns concrete config improvements
- Knowledge anchors — domain vocabulary (Medical, Legal, Engineering, Finance, Research) calibrates drift detection to your field
- Persistent document slots — pin up to 3 documents that stay in context every turn, never forgotten
- Display preferences — 4 themes, font size slider, compact mode for phones
Signal detection: 6 hallucination proxies (H-signals), 7 behavioral proxies (B-signals), EWMA trend tracking, semantic anchor distance monitoring, Integrity Floor breach detection.
| Feature | Option 1 (Claude) | Option 2 (Vercel) |
|---|---|---|
| TF-IDF + JSD coherence scoring | ✓ | ✓ fallback |
| Semantic embeddings (all-MiniLM-L6-v2) | — | ✓ |
| Linear Kalman filter | ✓ | — |
| Unscented Kalman Filter (UKF) | — | ✓ |
| GARCH(1,1) + jump-diffusion | ✓ | ✓ |
| Monte Carlo SDE bands | ✓ | ✓ |
| EWMA + Anchor chart lines | ✓ | ✓ |
| AutoTune (per-turn context detection) | ✓ | ✓ |
| Feedback loop (EMA learning) | ✓ | ✓ |
| Reflexive session analysis | ✓ | ✓ |
| Knowledge Anchors (domain calibration) | ✓ | ✓ |
| Persistent Document Slots (3 slots) | ✓ session | ✓ cross-session |
| Display preferences (theme, font, compact) | ✓ | ✓ |
| H-signals + B-signals | ✓ | ✓ |
| Session health, rewind, RAG | ✓ | ✓ |
| Integrity Floor | ✓ | ✓ |
| Framework Mode (HUDSON / STANDARD) | ✓ | ✓ |
| Multi-provider (OpenAI, Grok) | — | ✓ |
| API key persistence | — | ✓ |
| Cross-session memory | — | ✓ |
| Works without Claude account | — | ✓ |
| Preset | Dec / Cau / Calm | Best For |
|---|---|---|
| DEFAULT | 0.200 / 0.120 / 0.080 | General use |
| TECHNICAL | 0.180 / 0.100 / 0.060 | Code, audits, engineering |
| CREATIVE | 0.280 / 0.160 / 0.100 | Writing, brainstorming |
| RESEARCH | 0.220 / 0.130 / 0.085 | Academic, long-form analysis |
| MEDICAL | 0.150 / 0.090 / 0.055 | High-stakes clinical/legal |
| CIRCUIT | 0.140 / 0.080 / 0.050 | Logic verification |
| CUSTOM | user-defined | Fully configurable |
All behind TUNE → ⚗ ADVANCED. Labeled experimental.
- Alt SDE Models — CIR or Heston stochastic volatility
- Custom Rails — behavioral guidelines injected into every prompt
- Stability Panel — convergence tracking toward RESONANCE_ANCHOR
- Edit Constants — tune κ (0.00–5.00), live λ=1/(1+κ) display
- MHT Study — Metatron-Hudson Theory SDE module
- Poole Manifold CA Sim — 3D cellular automaton, full adder truth table
- Integrity Floor — DRIFT vs INTEGRITY BREACH threshold detection
import { computeCoherence, kalmanStep, updateSmoothedVariance,
buildPipeInjection, PRESETS } from './sdk/index';
const cfg = PRESETS.CIRCUIT;
const score = computeCoherence(response, history);
const newVar = updateSmoothedVariance(scoreHistory, prev, cfg);
const kalman = kalmanStep(state, score, turn * (2*Math.PI/12), SDE_PARAMS);
const pipe = buildPipeInjection(newVar, kalman.x, kalman.P,
calmStreak, driftCount, 'audit', turn, 0, 0, null, cfg);Perry, D. & Hudson, D. (2026). ARCHITECT: Universal Coherence Engine.
Hudson & Perry Research. @RaccoonStampede · @Prosperous727
github.com/Myth727/ARCHITECT-Universal-Coherence-Engine
© 2026 Hudson & Perry Research — Experimental R&D. All outputs are proxy indicators.