Applicable to: Language models · Software agents · Inference pipelines · Multimodal systems · Any sequential generative process
© 2026 Hudson & Perry Research
Authors: David Hudson (@RaccoonStampede) · David Perry (@Prosperous727)
Version: V2.0.0 · License: MIT · Live Demo
⚠ RESEARCH & DEVELOPMENT — NOT FOR CLINICAL OR LEGAL USE.
All outputs are mathematical proxy indicators. No warranty expressed or implied.
Most generative systems produce outputs and hope for the best.
VECTOR treats sequential generation as what it actually is: a stochastic process that can destabilize. It measures volatility in real time, detects when a system is drifting toward degraded output, and injects corrective signals before the output degrades further.
Accurate classification: VECTOR is a high-dimensional observability engine with a soft heuristic controller. It is not a provably safe control system — all outputs are proxy indicators. The correction loop is advisory, not enforced. See Validation Status below for what has and has not been proven.
The core loop:
Score output → Estimate state (Kalman) → Track volatility (GARCH) →
Detect instability → Inject correction (u_drift) → Measure effect (ΔC) → Repeat
What this means in practice: When a generative system starts drifting — getting sycophantic, inflating claims, losing context, contradicting itself — VECTOR catches it mathematically and injects corrective directives dynamically. Since V1.8.0, it also measures whether those corrections actually improved coherence (causal delta with k=1..5 lag binning), building the evidence base for future calibration.
Seven problems that affect every serious generative AI workflow:
- Drift & Incoherence in Long Contexts — LLMs wander, loop, confabulate. VECTOR quantifies it live and injects corrective directives automatically.
- Unreliable Self-Reflection — No reliance on the model to "check itself." External statistical harness enforces coherence.
- Context Bloat & Token Waste — Pruning + RAG + mute/gate keep prompts lean and relevant throughout long sessions.
- Domain Mismatch — Presets and tunables adapt tolerance to your context: tight for code/audit, looser for creative work.
- Observability Black Box — Dashboard, signals, exports, and session rewind provide audit trails and reproducibility.
- Prompt Engineering Fatigue — Pipes and harness automate the steering that normally requires manual system prompts every turn.
- False Confidence in Outputs — Proxy signals flag hype, sycophancy, and low-source replies. Post-audit catches quiet failures.
- Download
VECTOR.jsxfrom the root of this repo - Open claude.ai and start a new conversation
- Paste:
Create an artifact from this file. Run it exactly as-is.followed by the full file contents
Works immediately. No account, no server, no install.
Live demo:vector2026.vercel.app
- 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.
Vercel adds: Semantic embeddings (all-MiniLM-L6-v2 ONNX ~23MB) · Multi-provider (Anthropic, OpenAI, Grok) · Cross-session persistence via proxy
Volatility measurement:
Every output is scored using TF-IDF + Jensen-Shannon Divergence — measuring how much vocabulary and semantics shifted from prior context. That score feeds a GARCH(1,1) model that tracks volatility clustering over time and a Kalman filter that smooths the trajectory into a reliable state estimate.
Why this matters: A single unusual output might be fine. The system asks: is volatility rising across multiple turns? Is the system in a high-variance regime? That's what GARCH catches that individual turn scoring misses.
Detection:
When volatility crosses thresholds, VECTOR fires signal detectors:
- H-Signals (hallucination proxies): High-confidence language under elevated variance. Low source consistency vs attached documents. Self-contradiction with prior turns. Low response entropy. High vocabulary novelty under instability.
- B-Signals (behavioral proxies): Sycophancy, hype inflation, roleplay drift, question flooding, topic hijack, unsolicited elaboration, phrase repetition.
All signals are proxy indicators. Honest framing enforced throughout.
Correction:
When instability is detected, VECTOR injects a corrective directive into the next system prompt — u_drift(t) in the SDE framework. AUDIT mode detects only. MODERATE adds light correction. DEEP CLEAN and EXTREME apply progressively stronger constraints.
The compressed pipe format:
[V|t7|v=0.142|st=CAU|kx=0.887|kp=0.0004|cl=2|dr=1|md=AUD|h=0|b=0]->CONSOLIDATE.[/V]
60–70% fewer tokens than verbose injection formats. Identical information content.
| Component | Origin | Function in VECTOR |
|---|---|---|
| SDE (OU process) | Physics / stochastic control | Models output trajectory evolution over time |
| GARCH(1,1) | Quantitative finance | Tracks volatility clustering across turns |
| Kalman filter | Aerospace / signal processing | Smooths noisy per-turn scores into reliable state estimate |
| TF-IDF + JSD | Information theory / NLP | Lexical drift scoring with exponential Bayesian blending α(t)=1−exp(−t/τ) |
| Pipe injection | Control theory | u_drift(t) — corrective force applied to next output |
| Langevin noise | Spintronics / MTJ physics | Hardware-realistic stochastic uncertainty bands |
| Lévy flight noise | α-stable distributions | Heavy-tail noise for rare large behavioral jumps |
| PID controller | Classical control theory | Proportional-Integral-Derivative on variance |
| EKF / Particle Filter | Aerospace / robotics | Nonlinear and non-Gaussian state estimation |
| Berry phase | Quantum geometry | Geometric phase of coherence trajectory |
| Spin Hall Effect | Spintronics | Torque coupling between variance and Kalman state |
| Mutual information | Information theory | Statistical dependence between turns |
| Realized volatility | Quantitative finance | Fast-reacting variance complement to GARCH |
| Lyapunov bound | Dynamical systems | Live stability guarantee of SDE parameters |
| StableDRL clipping | RL stability (Li et al. 2026) | Unconditional ratio clipping + self-normalization prevents correction feedback loops |
| Heston Full Truncation | Numerical methods | Eliminates absorption bias in variance simulation — clamps inside drift/diffusion terms |
| CIR/OU scale normalization | Signal processing | Normalizes CIR/Heston/Vasicek/SABR paths to zero-mean unit-variance before drift detection — prevents false drift events from scale mismatch |
Core equation:
dε(t) = a(t)ε(t)dt + b·dW_t
a(t) = (α + β_p·sin(ωt)) / (1+κ)
κ = 0.444 (Hudson Constant) or 0.500 (Standard) — user-selected
dW_t = b · √dt · z · η
η = √(1 + 1/(2Δ))
Δ (MTJ_DELTA) is the thermal stability factor from magnetic tunnel junction physics — Neel-Brown relaxation model (Brown 1963; Koch et al. 2000). Default Δ=25. Toggle ON/OFF in FEATURES tab.
Honest framing: The Langevin math is physically grounded. The direct empirical link between MTJ parameters and generative output coherence is theoretical — same mathematical family, not yet co-validated against actual hardware.
| Feature | What it does |
|---|---|
| AutoTune | Detects session context per turn, selects optimal generation parameters automatically |
| Feedback Loop | +1/−1 per response. EMA learning personalizes AutoTune profiles. Persists across sessions. |
| Reflexive Analysis | Sends session volatility fingerprint for analysis → returns prioritized config suggestions |
| Knowledge Anchors | Domain vocabulary (Medical, Legal, Engineering, Finance, Research) calibrates signal detection |
| Persistent Doc Slots | 3 pinned documents — injected every turn, never pruned, never forgotten |
| Session Memory | Auto-compresses history at turns 10/20/30. Solves long-session context loss. |
| META Panel | Second analysis chat with full VECTOR architecture + live session data embedded |
| Quick Tools Drawer | CALC (SDE/GARCH calculator), VERIFY (15 live session checks), EXPORT (CSV/JSONL/TXT) |
| Demo Mode | Run any prompt with and without VECTOR correction side by side. Baseline scored against user messages only — clean independent comparison. |
| RLHF→SDE Bridge | -1 ratings on drifted turns feed back into the SDE sigma parameter — the engine learns from confirmed correction failures. |
| Causal Delta (A1) | After every correction turn, VECTOR logs ΔC for k=1..5 subsequent turns against a bin-stratified baseline. Measures whether pipe injection actually improved coherence. Passive Phase B data collection — every session generates labeled evidence. |
| Advanced Math Sidebar | Live readout of Lyapunov stability bound, PID output, Realized Volatility, Mutual Information, Fisher Information, LZ Complexity, Berry Phase, SHE Torque. |
| Feature | Claude artifact | Vercel |
|---|---|---|
| TF-IDF + JSD scoring | ✓ | ✓ fallback |
| Semantic embeddings (all-MiniLM-L6-v2) | — | ✓ |
| Unscented Kalman Filter (UKF) | ✓ | ✓ |
| Linear Kalman (SDK only) | — (UKF used) | — (UKF used) |
| GARCH(1,1) + jump-diffusion | ✓ | ✓ |
| Monte Carlo SDE bands | ✓ | ✓ |
| Langevin/MTJ noise model | ✓ | ✓ |
| AutoTune | ✓ | ✓ |
| Feedback loop (EMA learning) | ✓ | ✓ |
| Reflexive session analysis | ✓ | ✓ |
| Knowledge Anchors | ✓ | ✓ |
| Persistent Document Slots | ✓ session | ✓ cross-session |
| Strategic Session Memory | ✓ session | ✓ cross-session |
| META Panel | ✓ | ✓ |
| Quick Tools (CALC/VERIFY/EXPORT) | ✓ | ✓ |
| Display preferences (themes, font, compact) | ✓ | ✓ |
| H-signals + B-signals | ✓ | ✓ |
| Session rewind, RAG, bookmarks | ✓ | ✓ |
| Integrity Floor | ✓ | ✓ |
| Multi-provider (OpenAI, Grok) | — | ✓ |
| API key persistence | — | ✓ |
| 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, precision-critical |
| CIRCUIT | 0.140 / 0.080 / 0.050 | Logic verification, tightest tolerance |
| CUSTOM | user-defined | Fully configurable |
Confirmed: SDE math · Kalman filter · GARCH(1,1) · TF-IDF+JSD scoring · pipe injection · behavioral signal detection · per-preset GARCH tuning · epsilon parameterization · post-audit dual Kalman · Langevin/Neel-Brown math · EDM parallel discovery (Science Advances April 2026 independently confirmed same 45° angular gate VECTOR uses for drift detection)
Active (V1.8.0): Causal delta measurement — ΔC_policy vs ΔC_baseline logged per session across k=1..5 lags post-injection. Bin-stratified to eliminate selection bias. Building Phase B labeled dataset passively.
Requires validation: C-score vs human judgment (target r > 0.6) · H-signal false positive rate (need 100 labeled examples, AUC > 0.80) · GARCH on semantic drift (heuristic, not proven) · 623.81 Hz physical anchor · Langevin/MTJ empirical co-validation · cross-domain applicability beyond LLM conversations
Roadmap — Phase B: Collect N ≥ 100 labeled sessions (normal / degraded / failure). Fit state-space model via EM algorithm. Run logistic regression: P(failure | C_t, σ²_t, k_t, H_t, B_t). Accept only if AUC > 0.80. See CONTRIBUTING.md for full spec.
Honest system classification: Observability engine + heuristic soft controller. Not a provably safe control system. Signals are uncalibrated proxies until Phase B completes. Do not use for autonomous decisions.
Toggleable in TUNE → FEATURES. All default OFF — enable only what you need.
| Module | Toggle | What it does |
|---|---|---|
| Lévy Flight Noise | FEATURES tab | Replaces Langevin with α-stable heavy-tail noise (α=1.7). Models rare large behavioral jumps. |
| Extended Kalman Filter | FEATURES tab | Nonlinear Jacobian linearization. More accurate than linear Kalman for OU dynamics. |
| Particle Filter | FEATURES tab | 200-particle Sequential Monte Carlo. Handles non-Gaussian, multimodal drift. Blends with Kalman. |
| StableDRL Mode | FEATURES tab | Unconditional score clipping + self-normalizing injection. Prevents over-correction feedback loops. Default ON. (Li et al. 2026) |
| RLHF→SDE Bridge | FEATURES tab | Human -1 ratings on drifted turns nudge adaptive σ upward. Independent of sigma adaptation toggle. Default ON. |
Selectable in TUNE → ADVANCED → Alt SDE Model: CIR · Heston · Vasicek · SABR · DEFAULT (OU)
All sidebar math metrics (Lyapunov, PID, Realized Vol, Mutual Info, Fisher Info, LZ Complexity, Berry Phase, SHE Torque) always computed and displayed — no toggle needed, no performance cost.
- Alt SDE Models (CIR, Heston, Vasicek, SABR) — fully implemented, selectable in Advanced tab
- Custom behavioral rails
- Stability convergence panel (RESONANCE_ANCHOR = 623.81 Hz)
- Edit constants (κ, ε, GARCH params)
- MHT Study (Metatron-Hudson Theory SDE)
- Poole Manifold CA Simulator (3D cellular automaton, full adder verification)
- Integrity Floor (DRIFT vs INTEGRITY BREACH detection)
The sdk/ directory is a fully standalone TypeScript engine — all VECTOR math available without the React UI. Use in Node.js pipelines, agents, offline scoring, or any TypeScript project.
Note: The SDK is updated alongside VECTOR.jsx releases but may occasionally lag one version. The definitive source is always VECTOR.jsx.
What the SDK exports:
- SDE simulation: OU, CIR (with Feller check), Heston (Full Truncation), Vasicek, SABR, Lévy noise, path normalization, Lyapunov bound
- State estimation: Kalman, EKF, Particle Filter (200-particle SMC)
- Variance modeling: GARCH(1,1), PID controller, Realized Volatility, StableDRL clipping
- Signal detection: H-signals, B-signals, Mutual Information, Fisher Information, Kolmogorov Complexity, Berry Phase, SHE Torque, EWMA
- Coherence scoring: TF-IDF + JSD with exponential Bayesian blending
- Engine: Pipe injection, RAG, session health, context pruning
import{computeCoherence,kalmanStep,updateSmoothedVariance,buildPipeInjection,simulateCIR,ekfStep,computeMutualInformation,computeLyapunovBound,stabledrlClipScore,PRESETS}from'./sdk/index';constcfg=PRESETS.CIRCUIT;constscore=computeCoherence(response,history);constnewVar=updateSmoothedVariance(scoreHistory,prev,cfg);constkalman=kalmanStep(state,score,turn*(2*Math.PI/12),SDE_PARAMS);constpipe=buildPipeInjection({smoothedVar: newVar,kalmanX: kalman.x,kalmanP: kalman.P, calmStreak, driftCount,harnessMode: 'audit', turn,hSignalCount: 0,bSignalCount: 0,adaptedSigma: null});VECTOR.jsx ← paste into Claude
components/VECTOR.jsx ← same file, used by Next.js
pages/api/proxy.ts ← multi-provider proxy (Anthropic · OpenAI · Grok)
pages/index.tsx ← Next.js entry
public/embedder.worker.js ← neural embedding Web Worker (Vercel only)
sdk/*.ts ← TypeScript math library
evals/VECTOR_EVALS.md ← 15-check release checklist
ai/knowledge/
DIALOGUE_BASELINES.md
HALLUCINATION_REFERENCE.md
VECTOR_CODING_RULES.md
DOCUMENT_INTELLIGENCE.md
VECTOR's Reflexive Analysis and offline optimization tools are adapted from the Meta-Harness framework by the Stanford IRIS Lab.
Lee, Nair, Zhang, Lee, Khattab & Finn (2026). Meta-Harness: End-to-End Optimization of Model Harnesses. arXiv:2603.28052. https://arxiv.org/abs/2603.28052
What we borrowed:
- 3-candidate structured proposal format (exploitation + exploration)
- Exploitation/exploration axis taxonomy (A–F)
- Anti-parameter-tuning rules (parameter sweeps almost always regress — change mechanisms)
- Frontier tracking (
vector_frontierlocalStorage key — best config per context type) - Evolution summary JSONL format (compatible with Meta-Harness tooling)
- Offline harness pattern for
tools/vector_harness.pyandtools/meta_loop.py
What's different:
- Meta-Harness is an offline optimizer that searches over static harness configurations.
VECTOR is a live real-time controller that runs turn-by-turn during active sessions. - Meta-Harness uses Python + Claude Code. VECTOR runs entirely in the browser.
- VECTOR's mathematical engine (SDE/GARCH/Kalman/PID) has no equivalent in Meta-Harness.
The two are complementary: Meta-Harness can auto-optimize VECTOR's preset parameters offline, then feed the winning config back into live sessions.
Perry, D. & Hudson, D. (2026). VECTOR: Volatility-Sensitive Correction Engine.
Hudson & Perry Research. @RaccoonStampede · @Prosperous727
github.com/Myth727/VECTOR
© 2026 Hudson & Perry Research — Experimental R&D. All outputs are proxy indicators.