⚡ WERR: Zero-Memory Fractal System-One Decision Engine (Waves & Errors)

Motto: "When the Wave meets Error, we Recurse (werr)."
"Werr is the reflex? Ver! (werr)."
"Jev decisions come from 4B-parameter tensors; werr decisions come from infinite geometric waves, Euler thresholds, and recursive subdivision."

🌐 Language Switcher / Dil Seçici:
English (Default) | 🇹🇷 Türkçe Dokümantasyon (README_TR.md)

🌐 Interactive Web Lab: Try the Live Decision Simulator on GitHub Pages (Supports English & Türkçe, Light & Dark mode).
📄 Official arXiv Paper: Read on arXiv:2609.25498 [cs.NE] │ Direct PDF │ CERN Zenodo Archive.
⚡ Online Live Benchmark Arena: Run WebMCP, JevBench, Gymnasium RL, Arena.ai Blind Match & Tau-Bench Live in Browser (100% Client-Side, 0 VRAM, Instant CPU Execution).
⚔️ The Zero-VRAM Gauntlet (Master Benchmark Directory): Explore Complete Benchmarks Directory & Showdown (Detailed Markdown monographs for Snake, Jevenator 2, WindTunnel & JevBench).
📜 Open-Source Benchmark Integrity Report: Read the full audit on JevBench v1.2 / v1.3 metric shifts and dual verification (81.36 / 76.90).
⚙️ Engine Optimization Report: Read the full technical report on label tokenization, calibration, and zero-overfitting invariants.
🛡️ Official Sealed PDF Audit Report: Download Publication-Grade Audit PDF │ HTML Interactive Report │ Cryptographic SHA-256 Manifest.

werr is an open-source, machine-native System-One decision engine for software applications. Instead of running large language models or maintaining multi-gigabyte weight tensors in VRAM, werr synthesizes instant, typed decisions (noul, choice, score) on-the-fly from deterministic Mandelbrot fractal escape dynamics and quadrant subdivision.

💡 The 4-Dimension WERR Philosophy & Terminology

1. Dynamical Synthesis (Waves & Errors): Continuous complex polynomial trajectories colliding with catastrophic Euler error divergence thresholds ($|Z_n| > 2$). When the wave meets error, we recurse (WERR).
2. Spatial Inquiry ("Where" $\to$ "Werr"): Homophonous with English "Where", operational edge questions map directly to fractal coordinate topology:
- "Werr is the point?" $\to$ Finding the exact resonant coordinate seed $(c_x, c_y)$ along $\partial \mathcal{M}$.
- "Werr is the error?" $\to$ Instant microsecond fault isolation and boundary divergence detection.
- "Werr is the answerr?" $\to$ Deciding whether a problem requires instant reflex (werr) or deliberate reasoning (answerr).
- "Werr is the result?" $\to$ Sub-millisecond typed decision output (noul, choice, score).
- "Werr is the boundary?" $\to$ The fundamental mathematical threshold ($|Z| = 2.0$) dividing stability from chaos.
3. Actionable Imperative (Turkish "Ver!" & Dual-Engine Action Model): In Turkish phonetics, ver is the definitive command for action: "Karar werr!" (decide!), "Cevap werr!" (answer!), "İzin werr!" (authorize!).
- English UI Pair:
  - [ 💬 Answerr It! ] $\to$ System-2 Conversational Reasoning / Code Generation / Solution Design.
  - [ ⚡ Werr It! ] $\to$ System-1 Instant Reflex / 0-Byte VRAM / Microsecond Gate Execution.
  - Contextual Gates: [ 🛡️ Gate It! ], [ 🚀 Rank It! ].
- Turkish UI Pair:
  - [ 💬 Cevap Werr ] $\to$ Sistem-2 Sohbet ve Kod Çözümü.
  - [ ⚡ Karar Werr ] $\to$ Sistem-1 Anlık Fraktal Karar ve Refleks.
  - Ekstra Kapılar: [ 🛡️ İzin Werr ], [ 🚀 Öncelik Werr ].
4. Twin Cognitive Ecosystem Synergy (answerr $\leftrightarrow$ werr):
- answerr (answerr.me): The deliberative System-2 Cloud Portal, AI IDE assistant, and cognitive reasoning hub.
- werr (pip install werr): The machine-native, in-process System-1 reflex kernel executing in $< 0.5$ ms with 0 Bytes of VRAM.
- 📘 Full Design & Terminology Specification: See docs/werr_terminology_guide.md.

🎨 Official Vector Logos & Design Assets

Asset Type	Graphic Preview	File Link	Specifications
Twin Ecosystem Banner	`werr` ←→ `answerr`	`assets/werr_answerr_twin_ecosystem.svg`	1200×650 SVG • Complete Dual Architecture
werr Core Symbol (Badge)	3D Crystal "W"	`assets/werr_logo_core.svg`	800×800 SVG • Full Badge with Typography
werr Icon (Transparent)	Translucent "W"	`assets/werr_logo_core_fav_trans.svg`	514×452 SVG • Ideal for Dark UI & Web Headers
werr Icon (Dark Slate)	Solid Dark Base	`assets/werr_logo_core_fav_black.svg`	514×452 SVG • High-Contrast Favicon / App Icon
answerr Portal (Trans)	Neon Aperture	`assets/answerr_logo_portal_fav_trans.svg`	514×452 SVG • Transparent Cloud SaaS Emblem
answerr Portal (Dark)	Cosmic Void Base	`assets/answerr_logo_portal_fav_black.svg`	514×452 SVG • High-Contrast Favicon / App Icon

📊 Comparison: Jev (TypeSafe AI) vs. werr

Dimension	TypeSafe AI (Jev)	OpenJev / NanoJev	werr (Fractal System-1)
Foundation	Proprietary LLM	Qwen / Gemma (4B)	Mandelbrot Boundary ($\partial \mathcal{M}$)
Weight Tensor Memory	Multi-GB (Cloud)	~8 GB VRAM	0 Bytes (Zero Tensor Memory!)
Seed Footprint	Cloud API Endpoint	Local Model Checkpoint	24 Bytes $(c_x, c_y, \text{zoom})$
Typical Latency	~100 ms (Network HTTP)	~15–30 ms (GPU)	< 1.0 ms (Pure Local CPU)
Hardware Requirement	Internet Connection	CUDA-capable GPU	Any standard CPU / Microcontroller
License & Autonomy	Proprietary API ($/token)	Open Weights	Free for Research & Edu (BSL 1.1)

⚡ The Three Decision Primitives

Like Jev, werr answers three fundamental question types without producing conversational prose:

noul (Boolean Probability):
- Computes binary probability $p \in [0.0, 1.0]$, decision True/False, and confidence.
- Derived from $\partial \mathcal{M}$ hyper-surface thresholding.
choice (Categorical Selection):
- Chooses the optimal option among user-defined criteria.
- Derived from 4-Quadrant ($Q_1, Q_2, Q_3, Q_4$) or Quadtree energy partitioning.
score (Continuous / Ordinal Ranking):
- Evaluates continuous position along a defined ordinal scale (e.g. 0 to 3).
- Derived from the dark area integral ($D$) and escape velocity.

📦 Installation

# Install directly from GitHub:
pip install git+https://github.com/pCwOrM/werr.git

🚀 Quickstart

import werr as wr
from werr import (
    WerrEngine,
    NoulQuestion,
    ChoiceQuestion,
    ScoreQuestion,
    create_smart_router
)

# 1. Initialize engine (loads 24-byte coordinate seed)
engine = create_smart_router()

# 2. Define program state
state = {
    "user_role": "admin",
    "request_rate": 4.5,
    "payload_bytes": 1024
}

# 3. Ask typed questions
response = engine.decide(
    state=state,
    questions={
        "is_safe": NoulQuestion(instructions="Is this operation safe to proceed?"),
        "route": ChoiceQuestion(
            instructions="Target cluster",
            criteria={"prod": "Production", "canary": "Canary", "block": "Block"}
        ),
        "priority": ScoreQuestion(
            instructions="Priority level",
            criteria=["Low", "Medium", "High", "Critical"]
        )
    }
)

# 4. Use in ordinary code as a smart if-statement:
if response.boolean("is_safe") and response.score("priority") > 1.0:
    print(f"Routing to: {response.choice('route')} in {response.latency_ms} ms")

⚙️ Multi-Scale Dynamic Harmonic Architecture (v0.5.0)

werr v0.5.0 introduces the Multi-Scale Dynamic Harmonic Architecture, combining multi-scale boundary sampling, bounded density estimation, and continuous bifurcations across orthogonal parametric dimensions:

🌟 The Three Algorithmic Pillars (v0.5.0)

Tripod Multi-Scale Harmonic Kernel (tripod=True, Default Enabled): Rather than evaluating boundary dynamics at a single focal scale, WERR evaluates the escape boundary at three harmonic focal scales: $0.6\times$ (macro topological basin), $1.0\times$ (nominal boundary cusp), and $1.6\times$ (fine multi-fractal filaments) with weighted geometric ensemble ($w = [0.25, 0.50, 0.25]$). This prevents boundary trapping and lifts JevBench accuracy across all tiers to 54.55% (85.42% Easy, 44.14% Hard).
Bounded Density Normalization (arXiv:1810.11107): Incorporates $L_\infty / L_1$ bounded quadrant density estimation (extract_bounded_quadrant_weights), stabilizing escape velocity distributions against edge saturation.
Cadence Supercritical Pitchfork Bifurcation: Employs continuous dynamical system bifurcation ($\dot{x} = r x - x^3$, apply_cadence_bifurcation) to dynamically sharpen binary and categorical boundaries under high-risk conditions without artificial step functions.

1. Domain Routing Dimension (`domain_mode="none"` vs `domain_mode="multi"`)

domain_mode="none" (Default for General Reasoning & Academic Benchmarks): Domainless Monolithic Mode. Bypasses discrete domain gates and directly modulates questions into the universal chaotic boundary cusp ($c = -0.743643887 + 0.131825904i$, zoom $50.0$). Eliminates domain classification bias, allowing open-ended general intelligence and multi-step deduction to resolve naturally along $\partial \mathcal{M}$. Achieves 54.55% overall accuracy and an official JevBench v1.4.1 score of 23.74 at 7.32 ms latency via Tesla 3-6-9 Harmonic Grid (tying/beating Raw Qwen3 8B at 23.68, LitJev 27B at 19.51, and GPT-5.6 Luna at 18.51).
domain_mode="multi" (Production IoT & Infrastructure Default): Multi-Domain Mode. Dynamically routes state signatures through the 5 specialized domain gates (IoTSafetyGate, FinancialRiskGate, APISecurityGate, EcommerceFraudGate, GameCombatGate) via AutoSeedRouter. Essential for edge hardware, industrial safety, and API gateways where deterministic physical sensor thresholds are required.

2. Lexical & Resonance Ontology Dimension (`mode="production"`, `mode="resonance"`, `mode="pure_fractal"`)

mode="production" (Default in Production): Full activation of bilingual lexical dictionaries (English & Türkçe) and physical sensor threshold heuristics (smoke_detected, gas_ppm, temperature extremes). Active on production servers (mechsrv / answerr.me).
mode="resonance": Pure mathematical Tesla 3-6-9 frequency resonance dictionary. Delivers ultra-low latency (13.8 ms in multi-domain mode) while achieving 52.81% accuracy on JevBench.
mode="pure_fractal" (or enable_ontologies=False): Strips all external lexical dictionaries; operates purely on chaotic Mandelbrot boundary escape dynamics and criteria $N$-gram geometry.

from werr import WerrEngine
from werr.adapters import JevWireAdapter

# 1. Domainless Tripod Monolithic Mode (General reasoning, Q&A, and benchmark audits)
engine_general = WerrEngine(domain_mode="none", mode="pure_fractal", tripod=True)

# 2. Multi-Domain Edge IoT & Infrastructure (Deterministic safety & domain gates)
engine_edge = WerrEngine(domain_mode="multi", mode="production", tripod=True)

# 3. Transparent Wire-Format Adapter (External benchmark harness integration)
wire_adapter = JevWireAdapter(domain_mode="none", tripod=True)
decision = wire_adapter.decide(task_dict)

Note on Protocol Adapters: Wire protocol integration is handled directly by werr.adapters.JevWireAdapter with zero task heuristics and 100% air-gapped deterministic evaluation. The early prototype module werr.calibrated_engine has been completely purged from the repository in favor of the clean, modular JevWireAdapter and WerrEngine.

🗺️ The Universal Fractal Natural Language Decision Map

werr is pioneering the concept of the Universal Fractal Natural Language Decision Map.

Instead of training dense neural networks that require billions of parameters, any arbitrary program state and natural language questions—in English, Türkçe, or domain-specific terminology—are deterministically modulated onto the chaotic boundary of the Mandelbrot set ($\partial \mathcal{M}$).

[Program State / Girdi Durumu]
       │
       ▼
[Deterministic Semantic Modulation (TR/EN)]
       │
       ▼
[24-Byte Coordinate Seed (cx, cy, zoom)]
       │
       ▼
[Instant Fractal Boundary Evaluation (< 0.5 ms)]
 ├── w1, w2, w3, bias (Quadrant Decomposition)
 └── Quadtree Escape Integral
       │
       ▼
[Typed Decisions: Noul (Yes/No) | Choice (Routing) | Score (Severity)]

🌐 Multi-Domain Application Use-Cases

API Gateway & Microservices:

# State: {"user_role": "guest", "failed_attempts": 3, "req_frequency": 45}
# Decision: allow_execution=False | route=sandbox_audit | threat_score=1.45 / 3.0

🏠 Akıllı Ev / Smart Home (IoT):

# State: {"oda": "salon", "sicaklik": 27.5, "hareket_var": True, "pencere_acik": False}
# Question: "Klima çalıştırılsın mı?" -> True (p=0.892, Güven=%78.4)

🛒 E-Ticaret & Sahtecilik Tespiti (Fraud Detection):

# State: {"siparis_tutari": 18500, "yeni_cihaz": True, "vpn_kullanimi": True}
# Question: "İşlem doğrudan onaylansın mı?" -> False | Rota: "sms_dogrulama"

🎮 Oyun Yapay Zekası / Game AI (NPC Combat Reflexes):

# State: {"npc_can": 20, "dusman_mesafe": 5.2, "muhimmat": 0, "siginak_yakin": True}
# Decision: savasa_devam=False | taktik_karari="siginaga_kac" | panik_seviyesi=2.6

🏦 Finans ve Otomatik Kredi Değerlendirme:

# State: {"kredi_notu": 1520, "aylik_gelir": 75000, "gecikme_sayisi": 0}
# Decision: kredi_onay=True | kredi_paketi="aninda_onay" | guven=3.0 / 3.0

🔒 100% Air-Gapped Autonomous Operation & Zero-Network Guarantee

When an engineer downloads or installs werr, it operates 100% locally and completely offline:

Zero Inbound Network Calls: werr never downloads external weights, models, embeddings, or schemas from remote servers. There are no Hugging Face model checkpoints, no cloud API endpoints, and no server-side dependencies.
100% Local Text Normalization: All text processing—including Turkish diacritic normalization (ı/i, ö/o, ü/u, ş/s, ç/c, ğ/g), token extraction, and mathematical projections—is executed entirely in-process on the local CPU in microseconds.
Passive Outbound Telemetry (100% Opt-Out): The only network capability in werr is an optional, passive background telemetry dispatcher used for open-science benchmark calibration. It never blocks decision execution, sends zero PII, and can be completely shut off at any time:
```
export WERR_TELEMETRY=0
```

🧠 Dual-Layer Cognitive Inference Architecture

How does werr understand and decide upon arbitrary inputs? Does it search for specific keywords, or can it synthesize meaning from completely unseen words?

werr operates on an innovative Dual-Layer Cognitive Mechanism:

           [Arbitrary State & Question]
                       │
         ┌─────────────┴─────────────┐
         ▼                           ▼
[Layer 2: Bilingual Root    [Layer 1: Universal Chaotic
 Ontology Heuristic]         Phase-Space Resonator]
 • Fast cognitive shortcut   • Deterministic fallback
 • TR & EN domain lexicon    • Handles ANY unknown string
 • Action & risk polarities  • Fourier MD5 phase-angle
         │                           │
         └─────────────┬─────────────┘
                       ▼
    [Mandelbrot Boundary Perturbation (dx, dy)]
                       │
                       ▼
       [Sub-2ms Deterministic Decision]

1. Layer 1: Universal Geometric Phase-Space Resonator (Zero-VRAM Fallback)

What happens if you pass alien, fictional, or completely unknown tokens?
(e.g., glork_factor: "frobnicate_the_wozzer", synthetic telemetry keys, or foreign jargon).
Unlike Large Language Models (which hallucinate or throw Out-Of-Vocabulary / OOD errors), werr never crashes and never fails to decide.
Unrecognized strings are mapped to continuous trigonometric phase angles $(\sin \theta, \cos \theta)$ via Fourier projection of their cryptographic byte-entropy onto the complex plane $\mathbb{C}$.
These phase vectors deterministically modulate the Mandelbrot boundary seed coordinates $(c_x, c_y)$ and evaluate quadrant escape dynamics ($w_1, w_2, w_3, \text{bias}$).
Result: 100% deterministic, type-safe decisions with 0 Bytes of GPU VRAM, even for inputs the system has never seen before!

2. Layer 2: Calibrated Bilingual Root Ontology (Cognitive Shortcut)

For real-world production domains, werr incorporates an ultra-compact, calibrated root ontology in English and Türkçe:

Directionality & Authorization: allow, permit, safe, valid, approve, izin, onay, uygun, gecerli, calistir, ac, dogrula, kabul...
Threat & Risk Polarities: threat, attack, block, deny, hazard, fraud, fire, tehlike, risk, engelle, yasak, saldiri, yangin, tahliye...
Operational Action Tiers: direct (dogrudan), rate_limiter (sinirla), sandbox_audit (incele), drop_packet (reddet / engelle)...
Domain Feature Handlers: Full dual-scale credit scoring (FICO 300–850 and Turkish Findeks 0–1900), IoT HVAC/comfort and life-safety hazard triage, payment fraud heuristics, and NPC tactical combat states.

3. 🎼 Chordial Semantic Resonance & Acoustic Tinleme Index Filter

To prevent incidental words (such as "Direct butane..." or "...proceed on escape trajectory") from triggering false keyword spikes in out-of-domain contexts without using crude, destructive string blacklists, werr implements the Chordial Semantic Resonance Field:

Triad Harmonic Agreement ($\mathcal{H} = \Phi(S) \otimes \Psi(Q) \otimes \Omega(C)$): Semantics are evaluated as a musical triad across the State context $\Phi(S)$, Question intent $\Psi(Q)$, and Option declaration $\Omega(C)$. Isolated words lacking chordial support from state and question are treated as acoustic noise and damped.
Acoustic Tinleme Index ($\mathcal{T}$): Primary option keys carry sharp, high-integrity metallic signal ($\mathcal{T} = 1.0$), while incidental descriptive words in arbitrary text are dynamically damped ($\mathcal{T} \le 0.08$).
Conservative Smooth Coupling ($C^\infty$): Discontinuous if-else steps are replaced by smooth hyperbolic tangent potentials ($\Delta S = \mathcal{A}_{\max} \cdot \tanh(\Delta \text{risk}) \cdot \mathcal{T}$), mathematically eliminating chaotic butterfly explosions along the fractal boundary.

🧭 Multi-Domain Auto-Seed Router & Empirical Benchmark (v0.2.0)

In version 0.2.0, werr introduces the Multi-Domain Auto-Seed Router (AutoSeedRouter). While earlier iterations used a monolithic boundary seed ($c_x \approx -0.747, c_y \approx 0.131$), evaluating distinct domains requires dynamically hopping into the topological coordinates where each domain's feature derivatives resonate with maximum sensitivity.

🔬 Empirical Ablation Study (Monolithic Seed vs. Auto-Seed Router)

Evaluated across $N = 336$ empirical telemetry decisions from production traffic and multi-batch validation sets:

Operational Domain	Sample Size ($N$)	Monolithic Seed Acc	Auto-Seed Router Acc	Net Gain ($\Delta$)	Avg Confidence	Inference Latency
API Gateway & Security	85	83.5%	87.1%	$+3.5%$	93.9%	3.42 ms
Smart Home & IoT Safety	61	85.2%	98.4%	$+13.1%$	100.0%	3.31 ms
E-Commerce Fraud	60	56.7%	85.0%	$+28.3%$	100.0%	3.29 ms
Game AI & NPC Combat	60	50.0%	95.0%	$+45.0%$	100.0%	3.17 ms
Financial Risk & Credit	60	35.0%	100.0%	$+65.0%$	100.0%	3.33 ms
OVERALL MACRO ACCURACY	326	63.8%	92.6%	+28.8%	98.0%	3.31 ms

All inferences executed with 0 Bytes of neural tensor memory (VRAM/RAM) and strict determinism.

🏆 JevBench Benchmark Evaluation & Grand Matrix (v1.4.1)

Werr was evaluated on the public test split (231 items) of JevBench (benchmarkheaven.com/jev-models & fstandhartinger/jevbench), measuring autonomous System-One decision models across 4 axes: Intelligence, Calibration, Speed, and Cost.

🌍 JevBench v1.4.1 Official Comparative Context

Under JevBench v1.4.1's strict chance-corrected quadratic harmonic scoring ($\text{Score} = \text{HarmonicMean} \times (\text{Intelligence}/50)^2$ for $\text{Int} < 50$), WERR v0.5.0 achieves 23.74 / 100 at 7.32 ms latency on pure cusp, and 20.63 / 100 via WerrLocalAdapter, directly rivaling an 8-Billion parameter transformer and surpassing larger models:

Model / System	Architecture	Hardware / VRAM	Intelligence	Speed	Cost	v1.4.1 Score
⚡ WERR v0.5.0 (Tesla 3-6-9 Cusp)	Pure Fractal Boundary Cusp ($\partial \mathcal{M}$)	Commodity CPU (0 Byte VRAM / 24 Byte Seed)	33.4	95.4	100.0	23.74 🏆
Raw Qwen3 8B	Dense Transformer (8 Billion Params)	GPU Cluster (~16 GB VRAM)	51.2	82.4	48.0	23.68
⚡ WERR v0.5.0 (`WerrLocalAdapter`)	In-Tree Standard Adapter (`res=36, max_iter=36`)	Commodity CPU (0 Byte VRAM / 24 Byte Seed)	30.6	95.3	100.0	20.63
LitJev 27B	Open Weights MoE / Dense	Dual GPU (~54 GB VRAM)	54.1	74.5	32.0	19.51
GPT-5.6 Luna	Frontier Closed LLM (OpenAI API)	Multi-Cluster Cloud Supercomputer	96.8	77.5	28.5	18.51
SmallJev (Local Checkpoint)	Distilled SLM Checkpoint	Local GPU (~4 GB VRAM)	41.2	84.1	68.0	12.87

📊 Grand Matrix: Multi-Domain vs. Domainless across 3 Comprehensive Benchmark Suites (Tesla 3-6-9 Accelerated)

To verify robustness across diverse operational conditions, WERR v0.5.0 was benchmarked across 3 independent suites in all 4 operational modes:

Operational Mode	Suite 1: Edge Domains (50 Tasks)	Suite 2: 100 TR Production	JevBench v1.4.1 Accuracy	JevBench v1.4.1 Score	Latency (CPU)
1. Domainless (Universal Cusp + Tesla 36)	18 / 50 (36.0%)	35 / 100 (35.0%)	126 / 231 (54.55%)	23.74	7.32 ms 🏆
2. Multi-Domain + Lexical Dictionary	21 / 50 (42.0%)	31 / 100 (31.0%)	119 / 231 (51.52%)	16.20	8.51 ms
3. Multi-Domain + Resonance (Tesla 3-6-9)	20 / 50 (40.0%)	31 / 100 (31.0%)	122 / 231 (52.81%)	18.82	7.80 ms
4. Multi-Domain + Hybrid (Lexical + Resonance)	21 / 50 (42.0%)	31 / 100 (31.0%)	120 / 231 (51.95%)	17.08	8.12 ms

📈 Cumulative Evolution Matrix across JevBench Versions

Run / Edition	Methodology & Invariant	Overall Accuracy	Easy Split	Original Split	Hard Split	Median Latency	v1.2 / v1.3 Score	v1.4+ Score
Run 1: Heuristic Calibrated	Task-specific semantic mappings (Historical)	74.80%	85.42%	90.28%	46.85%	2.58 ms	81.36 / 76.90	—
Run 2: Clean Core (`JevWireAdapter`)	Zero Hardcoded Rules, pure general criteria $N$-gram	46.75%	75.00%	43.06%	36.94%	3.47 ms	35.80 / 30.12	7.51
Run 3: Clean Calibrated Engine	Zero Hardcoded Rules, Platt temperature scaling	49.78%	85.42%	44.44%	37.84%	3.79 ms	41.50 / 36.20	12.39
Run 4: WERR v0.5.0 (Tripod Baseline)	Multi-Scale Tripod (64x64 @ 50 iters, 0.6x/1.0x/1.6x), Bounded Density, Cadence	54.55% (126/231)	85.42% (41/48)	50.00% (36/72)	44.14% (49/111)	19.9 ms	51.80 / 46.70	23.66
Run 5: WERR v0.5.0 (Tesla 3-6-9 Harmonic Grid)	Tesla Vortex Grid (36x36 @ 36 iters, 81 px/tile), Multi-Scale Tripod, Bounded Density, Cadence	53.25% (123/231) (Adapter) 54.55% (126/231) (Cusp)	83.33% (40/48) 85.42% (41/48)	48.61% (35/72) 50.00% (36/72)	43.24% (48/111) 44.14% (49/111)	7.58 ms (Adapter) 7.32 ms (Cusp) 🏆	53.20 / 48.50	20.63 (Adapter) 23.74 (Cusp) 🏆

📘 Detailed Benchmark Dossier: See the dedicated monograph in benchmarks/README.md and the integrity audit in docs/BENCHMARK_INTEGRITY_REPORT.md.

🚀 Live TypeSafe-Compatible Decision Server

Werr ships with a native, zero-dependency HTTP decision server implementing the TypeSafe wire format (POST /v1/systemone):

# Launch the live decision server:
python -m werr.server --port 8443

# Test instant reflex decision via curl:
curl -X POST http://localhost:8443/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{
    "state": "User requested account deletion without refund",
    "model": "werr-system-one",
    "questions": {
      "decision": {
        "type": "choice",
        "instructions": "Route request action",
        "criteria": {"cancel": "Subscription cancel", "refund": "Money return"},
        "labels": ["cancel", "refund"]
      }
    }
  }'

🌐 Autonomous Web Agent Benchmark: WindTunnel WebMCP (49/49 Solved, 100%)

Werr was evaluated against the official WindTunnel WebMCP 49-task benchmark suite (nekuda-ai/WindTunnel), the industry benchmark evaluating agentic reasoning across 8 real-world production web applications (nextjs-starter-medusa, hi-events, easyappointments, idurar-erp-crm, learnhouse, directory-9d8, tailwind-nextjs-blog, bulletproof-react).

Officially submitted and detailed in nekuda-ai/WindTunnel#25.

While standard LLM-based computer use models (such as GPT-6 Astra or Claude 3.7 Sonnet) require multi-second cloud roundtrips and high dollar costs per interaction, WebMCP transforms continuous browser DOM interaction into discrete MCP tool selections. Werr serves as a zero-memory procedural System-One router, resolving discrete tool selection and parameter routing in 3.35 ms median latency with 0 Bytes of VRAM and $0.0000 model cost.

📊 WebMCP Head-to-Head Benchmark Results

Model / Architecture	Interface	Tasks Solved	Success Rate	Median Latency	Model Cost (49 Tasks)	VRAM / Weights	Air-Gapped / Privacy
WERR (Procedural System-1)	WebMCP	49 / 49	100.0%	3.35 ms	$0.0000	0 Bytes	100% On-Device
Jev + Mercury 2.5	WebMCP	49 / 49	100.0%	3,200 ms	$0.0011	Cloud API	External API
GPT-6 Astra	Computer Use (Code)	46 / 49	93.9%	8,400 ms	$0.1230	Cloud API	External API
Claude 3.7 Sonnet	Computer Use (Bash)	45 / 49	91.8%	11,200 ms	$0.1850	Cloud API	External API
GPT-6 Astra	Computer Use (Screenshots)	44 / 49	89.8%	14,600 ms	$0.2700	Cloud API	External API
Jev (Ultrafast)	DOM (Raw)	25 / 49	51.0%	4,800 ms	$0.0013	Cloud API	External API

🔬 Key Advantages in Agentic Web Automation

Sub-4ms Instant Reflex: Executes tool selection and parameter navigation ~955× faster than Jev + Mercury 2.5 and ~2,500× faster than GPT-6 Astra.
True Zero-VRAM Footprint: Requires zero neural weights (0 Bytes VRAM), running deterministically via procedural escape-time mathematics.
Zero Financial & Cloud Cost: Total execution cost across all 49 tasks is $0.0000, with no token consumption or rate limits.
Absolute Data Privacy & Air-Gap Compliance: Zero user session data, DOM structures, or credentials leave the local environment, making it suitable for security-sensitive ERP/CRM platforms (idurar-erp-crm, easyappointments).

🧪 Reproducibility & Isolated Test Harness

The complete isolated benchmark harness is part of the Werr test suite:

# Run the isolated WindTunnel WebMCP benchmark suite (49 tasks):
python tests/test_windtunnel_webmcp_isolated.py

Standalone test script: tests/test_windtunnel_webmcp_isolated.py

🐍 Real-Time High-Frequency Benchmark: Snake AI Reflex (Werr vs. Laya-MLX vs. Jev)

To evaluate real-time continuous reflex throughput under strict closed-loop latency constraints, Werr was benchmarked against the standard Snake environment used in low-latency decision model research (experiments/snake_runtime.py).

Officially submitted to the Laya-MLX repository in mizorewww/laya-mlx#3.

Werr was evaluated across two execution regimes:

Run 1 (Baseline): Standard resolution (64×64, max_iter = 50).
Run 2 (Optimized In-Process Kernel): Tuned lightweight resolution (32×32, max_iter = 30, local in-process loop).

📊 Head-to-Head Benchmark Results

Metric / Dimension	TypeSafe Jev API	Laya-MLX (ModernBERT 421M)	Werr (Run 1: Baseline)	Werr (Run 2: In-Process)
Model Size / Weights	Cloud Model	421 Million Parameters (943.6 MiB)	0 Bytes (24-Byte Seed)	0 Bytes (24-Byte Seed)
Hardware Required	Cloud Server Cluster	Apple Silicon M3 Max ($3,500)	Commodity Desktop CPU	Commodity Desktop CPU
VRAM Footprint	Cloud GPU	943.6 MiB VRAM	0 Bytes VRAM	0 Bytes VRAM
P50 Decision Latency	150 - 350 ms	13.42 ms	2.88 ms	1.34 - 1.99 ms (Sub-2ms!)
Throughput (Moves/Sec)	2 - 5 moves/s	74.5 moves/s	243.1 moves/s	273.5 - 302.1 moves/s (3.7x - 4.1x Faster!)
Speedup vs Laya	Baseline (0.05x)	1.0x (Reference)	3.26x Faster	3.67x - 4.05x Faster
Speedup vs Jev Cloud	1.0x	15x - 25x	65x Faster	78x - 85x Faster
Marginal Cost / 1k Decisions	$0.0399	$0.0029 (est.)	$0.0000 (Pure Local)	$0.0000 (Pure Local)
Portability / Network	Cloud API required	Local (Mac MLX only)	100% Offline & Cross-Platform	100% Offline & Cross-Platform

🔬 Open-Source Reproduction & Exact Boundary Seed

The standalone harness, terminal visualizer, and game loop are tracked directly in the repository:

Repository Directory: benchmarks/snake/
Chaotic Boundary Seed: cx = -0.7445, cy = 0.1250, zoom = 65.0 (24 bytes)
Visual Terminal Showcase: assets/werr_snake_benchmark.gif • High-Res MP4 Video (122 KB)
Single-Command Reproduction:

# 1. Run comparative headless benchmark (600 steps):
python benchmarks/snake/benchmark_snake.py --steps 600 --mode compare

# 2. Render terminal visualizer (exports GIF/MP4 with real-time HUD):
python benchmarks/snake/visualize_snake.py --mode mp4 --steps 200

🌐 Cross-Repository Ecosystem Integration:

Interactive Browser 1v1 Arena: Explore the web GUI with mobile touch D-pad, swipe gestures, and human-vs-fractal autopilot in mandelbrot-fractal-neural-synthesis/demos/snake.html.

Cloud Reflex Gateway: Stream decisions to remote agents via answerr over POST /v1/systemone at https://api.answerr.me:4431.

🎯 Computer Vision & Video Tracking Benchmark: Jevenator 2 (Werr vs. Maisa djev)

Werr was evaluated on visual spatial object localization and temporal video tracking using the Jevenator 2 (Judgment Day) benchmark suite (mmastrac/jevenator2 by Matt Mastracci).

Officially submitted to the Jevenator 2 repository in mmastrac/jevenator2#1.

The benchmark evaluates a System-One model's ability to divide visual scenes into labeled spatial grids (3×3 or 7×5 = 35 cells) and answer boolean noul containment queries (Does region {c} contain {target}?) across static exemplar scenes and 24 sequential video tracking frames (840 discrete decisions):

📊 Head-to-Head Benchmark Results: Werr vs. Maisa djev

Metric / Dimension	Maisa djev (Diffusion-Gemma)	WERR (Fractal System-One)	Advantage / Gain
Model Architecture	Diffusion-Gemma Vision-LLM	Mandelbrot Escape Boundary Kernel	Procedural Zero-Tensor
VRAM / Weights Memory	~8 GB GPU VRAM	0 Bytes VRAM (24-Byte Seed)	Infinitely Lighter
Hardware Required	NVIDIA RTX / Cloud GPU	Standard Commodity CPU	Ubiquitous Edge Execution
Mean Frame Latency (35 cells)	761.8 ms	27.39 ms	27.8× Faster
P50 Frame Latency	~750.0 ms	22.07 ms	33.9× Faster
Decision Throughput	~45.9 decisions/s	455.7 decisions/s	9.9× Higher Throughput
Shapes Ground Truth (B, F)	100% (Triangle=B, Circle=F)	100% (Triangle=B, Circle=F)	Parity (100% Match)
Negative Control (Dyson=Empty)	Partial False Positive	100% Clean (0 False Positives)	Superior Specificity
Marginal Cost (840 Decisions)	Cloud API / GPU compute	$0.0000 (Pure Local CPU)	100% Free
Data Privacy & Air-Gap	Cloud Server / API dependency	100% Air-Gapped & Offline	Zero Video Leakage

🔬 Standalone Reproduction

# Run the complete Jevenator 2 benchmark:
python benchmarks/jevenator2/benchmark_jevenator2.py

# Or run the automated regression test suite:
python tests/test_jevenator2_isolated.py

Full benchmark suite: benchmarks/jevenator2/
Standalone test script: tests/test_jevenator2_isolated.py

📄 Academic Publication & Preprint Status

This research builds upon peer-reviewed open science preprints and is published as follows:

Primary Paper: Universal Fractal Natural Language Decision Map: Real-Time Edge Triage Across Heterogeneous Domains
- arXiv Preprint: arXiv:2609.25498 [cs.NE, cs.AI, cs.CL] • DOI: 10.48550/arXiv.2609.25498
- Permanent Archive: CERN Zenodo (DOI: 10.5281/zenodo.22867426)
- Authors: Volkan Dağlı, Dr. Zerrin Dağlı, Dağhan Dağlı • Lead Author: ORCID: 0009-0000-1587-8703
- Status: Officially published on arXiv (September 2026).
Foundational Companion Research: Mandelbrot Fractal Neural Synthesis: Zero-Storage Procedural Weight Derivation and Non-Linear Decision Boundaries
- Preprint Archive: CERN Zenodo (DOI: 10.5281/zenodo.22774934) • arXiv: submit/8092292 (under review)
- Authors: Volkan Dağlı, Dr. Zerrin Dağlı, Dağhan Dağlı
- Status: Under academic review / CERN Zenodo registered.

🇹🇷 First-Class Dual-Language Support (Türkçe & English)

werr natively supports Turkish and English queries without external translation models. Diacritics and character variants (ı/i, ö/o, ü/u, ş/s, ç/c, ğ/g) are normalized seamlessly:

Roller: yönetici, yetkili, üye, kullanıcı, misafir, ziyaretçi, saldırgan, şüpheli.
Soru Yönergeleri:
- İzin / Onay: izin verilsin mi?, onayla, geçiş uygun mu?, çalıştır.
- Engelleme / Tehlike: engelle, yasakla, tehlike var mı?, riskli mi?, saldırı mı?.
- Yönlendirme Rotası: doğrudan, hızlı yol, kuyruk, karantina, inceleme, reddet.

💻 Interactive CLI Simulator (`examples/sor.py`)

An interactive terminal application is provided under examples/sor.py with automatic dependency checking (werr), an interactive numbered menu, and CLI arguments:

# Run with interactive selection menu:
python examples/sor.py

# Or pass a role directly from the command line:
python examples/sor.py admin
python examples/sor.py member
python examples/sor.py guest
python examples/sor.py attacker

🔬 Open Science Telemetry & Public Benchmark Dataset

To calibrate and continuously optimize the universal fractal decision map, werr includes an asynchronous, non-blocking telemetry client (werr.telemetry).

🔒 Zero-PII Privacy Guarantee

No IP addresses are stored on disk or database.
No cookies, machine IDs, or personal accounts are collected.
Sensitive keys & values (password, token, secret, key, auth, email, jwt) are automatically sanitized and redacted ([REDACTED]) on the client side before dispatch.
100% Opt-Out: Set the environment variable WERR_TELEMETRY=0 (or WEVV_TELEMETRY=0) to disable telemetry completely.

📊 Live Public Dataset & Automated Nightly Dual-Sync (1,318+ Decisions)

Telemetry records are aggregated in MariaDB (werr_db) on the dedicated cluster node api.answerr.me. Every night at 00:00 UTC, an automated crontab dual-sync pipeline exports fresh records and publishes them synchronously to both GitHub and Hugging Face:

🤗 Hugging Face Hub Dataset: https://huggingface.co/datasets/pCwOrM/werr_open_decisions
🌐 Direct Download (1,318+ Records): https://api.answerr.me:4431/werr/dataset/werr_open_decisions.jsonl
📂 Repository Mirror: dataset/werr_open_decisions.jsonl
📑 Empirical Test Reports:

🧬 Organic Dynamic Calibration & Local Homeostasis (v0.2.2 - v0.3.0)

While v0.2.1 initialized empirical quadrant normalization from static benchmarks, werr v0.2.2+ introduces Organic Dynamic Calibration (werr.calibration.DynamicCalibration):

Local Homeostasis: Runs an $O(1)$ continuous Exponential Moving Average (EMA) of quadrant escape ratios across live queries. The decision engine organically self-calibrates to its local operational domain (e.g. industrial plants vs. high-frequency trading) with strict 0 Byte VRAM allocation.
Quadrant Phase Invariance: Eliminates positional option bias via deterministic instruction-hash phase shifts (phase_offset = hash(instructions) % 4).
Domain & Risk Adaptive Thresholding: Dynamically modulates decision cutoffs with continuous $\tanh(\text{net-risk} \cdot 0.8)$.
Adversarial Resilience: Tested against prompt-injection / trap-word attacks with a 0% exploit rate.
Web3 & Decentralized Decision Oracle Roadmap: View our long-term architectural specifications for on-chain verifiable fractal decision maps, ZK-Mandelbrot proofs, and EVM/Solana smart contract oracles in docs/roadmap_blockchain_decision_oracle.md.

🛡️ Server Hardening & Defensive Architecture

The remote ingestion endpoint on api.answerr.me:4431/werr/telemetry (with backward-compatible alias /wevv/telemetry) is hardened against abusive bots and brute-force traffic:

Token Bucket Rate Limiting: 30 requests/minute with a 5 req/s burst limit.
Auto-Jail (Anti-Bruteforce): Clients generating repeated violations (HTTP 413, 422, or rapid bursts) are automatically jailed for 15 minutes (HTTP 403).
Strict Payload Guard: Hard cap of 32 KB per request (LimitRequestBody 32768).
Google reCAPTCHA v3 Shield: The web simulator and custom scenario playground verify client authenticity via background reCAPTCHA v3 site verification.
Storage & Disk Quota: Dual-storage system caps log size at 1 GB and monitors host disk thresholds.
Systemd Sandboxing: Runs under an isolated service with MemoryMax=256M, CPUQuota=20%, ProtectSystem=full, and NoNewPrivileges=true.

🏛️ Production Runtime & Isolated Architecture (`mechsrv` / `api.answerr.me`)

The live production deployment on mechsrv (api.answerr.me:4431) is engineered with complete physical and OS-level sandbox isolation between production APIs and developer experimentation:

Layer / Component	Location & Service	Isolation & Security Guarantee
Production Decision Engine	`/opt/apps/answerr/` (`engine`, `api`, `venv`)	🔒 Full Sandbox (`ProtectHome=true`, `ProtectSystem=full`)
Developer Workspace	`/home/pcworm/werr` + independent `venv`	🛠️ Free Testing & Benchmarking
Verified Production Deploy	`/home/pcworm/deploy_to_prod.sh`	🚀 Automated Sanity Gate $\to$ Rsync $\to$ Restart $\to$ Health Check
Global Terminal CLI	`/usr/local/bin/werr`	⚡ Instant System-wide Reflex Execution
Telemetry Ingestion DB	MariaDB `werr_db` (1,318+ records)	💾 Isolated `werr_user` with zero client IP logging
Telemetry Ingestion Daemon	`werr-telemetry.service` (Port 8550)	🟢 Hardened FastAPI Collector (`/werr/telemetry`)
Open Decision Dataset	`/home/pcworm/werr_telemetry/dataset`	🌐 Served via Apache at `/werr/dataset`

Global CLI Commands (`/usr/local/bin/werr`)

# 1. Check werr version
werr --version
# Output: werr 0.5.0 (0-VRAM Fractal System-One Decision Engine)

# 2. Test security roles
werr admin       # Allowed (0.00ms latency, zero memory)
werr attacker    # Denied (Instant reflex block)

# 3. Interactive Decision Simulator
werr             # Interactive terminal simulator

🔗 Ecosystem Architecture & Connections

Foundational Scientific Theory & Interactive Client Labs: mandelbrot-fractal-neural-synthesis
The mathematical basis for zero-storage procedural parameter derivation from the Mandelbrot escape boundary ($\partial \mathcal{M}$) (Zenodo DOI: 10.5281/zenodo.22774934). Houses the interactive 1v1 AI Snake Arena GUI (demos/snake.html with mobile touch controls and human vs fractal autopilot) and quadrant visualizers.
Machine-Native Reflex Engine & Benchmarks (This Repository): werr
The in-process System-1 reflex kernel running on standard CPU with 0 Bytes VRAM, providing terminal benchmark visualizers (benchmarks/snake/visualize_snake.py) and verified world-record benchmark suites (WindTunnel WebMCP 49/49, JevBench #1, Jevenator 2).
Dual-Cognition Production Workspace & REST Gateway: answerr
The fullstack conversational platform and high-throughput production API (answerr.me, api.answerr.me:4431) that pairs werr System-1 spinal reflexes with Google Gemini Flash System-2 deliberation, exposing the live POST /v1/systemone wire endpoint.

🤝 Scientific Peer Acknowledgment & Mutual Evolution

Scientific progress is inherently bidirectional. We extend our sincere gratitude and highest professional respect to Florian Standhartinger and the JevBench research community (@airesearch12, fstandhartinger/jevbench). Just as their rigorous external audit inspired our architectural leap—completely purging domain wording overrides in favor of generalized multi-token $N$-gram criteria resonance, bidirectional negation scoping, and chaotic boundary dynamics—Werr’s arrival as the first 0-VRAM fractal contender disrupted the benchmark landscape, catalyzing JevBench's rapid evolution toward sealed testing, chance baselines, and harmonic penalty gates across v1.3 and v1.4. Having a real, non-conformist challenger in the ring forced both sides to innovate faster. In open science, iron sharpens iron.

👥 Authors & Academic Affiliations

Volkan Dağlı (Corresponding Author)
Anadolu University, Eskişehir, Turkey & ITouch Systems, Mersin, Turkey • ORCID: 0009-0000-1587-8703 • GitHub: @pCwOrM
Dr. Zerrin Dağlı
Mersin University, Mersin, Turkey • ORCID: 0000-0001-9490-6425
Dağhan Dağlı
Toros Science College, Mersin, Turkey • ORCID: 0009-0003-2492-8313 • GitHub: @Lexovian

📄 Academic Citation

@article{dagli2026werr,
  author        = {Da{\u{g}}l{\i}, Volkan and Da{\u{g}}l{\i}, Zerrin and Da{\u{g}}han, Da{\u{g}}l{\i}},
  title         = {Universal Fractal Natural Language Decision Map: Real-Time Edge Triage Across Heterogeneous Domains},
  journal       = {arXiv preprint arXiv:2609.25498},
  year          = {2026},
  eprint        = {2609.25498},
  archivePrefix = {arXiv},
  primaryClass  = {cs.NE},
  url           = {https://arxiv.org/abs/2609.25498},
  doi           = {10.48550/arXiv.2609.25498},
  note          = {CERN Zenodo Archive: https://doi.org/10.5281/zenodo.22867426; Companion Concept: 10.5281/zenodo.22774934}
}

@software{werr2026,
  author    = {Volkan Dağlı and Zerrin Dağlı and Dağhan Dağlı},
  title     = {werr: Zero-Memory System-One Decision Engine via Waves and Errors},
  year      = {2026},
  url       = {https://github.com/pCwOrM/werr},
  doi       = {10.5281/zenodo.22867426}
}

📜 License & Intellectual Property

This software and its reflex decision algorithms are licensed under the Business Source License 1.1 (BSL 1.1).

Academic, Educational & Non-Commercial Use: Free and permissive for non-commercial research, university education, student contributions, and academic benchmarking.
Commercial & Enterprise Use: Any commercial deployment, hosted SaaS, or paid cloud service requires an express commercial enterprise license from ITouch Systems (ITouch Bilişim Sistemleri Ltd. Şti.).
Patent Notice: Certain procedural decision algorithms are subject to pending patent application TÜRKPATENT TR 2026/016285.
Change Date: On 2030-01-01, this work automatically converts to the Apache License, Version 2.0.
Commercial Inquiries: Contact ITouch Systems (MERSİS: 0469094455800001, VKN: 4690944558, Sanayi Sicil: 827254) via info@itouch.com.tr • ask@answerr.me • KEP: itouchbilisim@hs01.kep.tr • answerr.me.

💖 Sponsor & Grant Funding

WERR is an independent, patented open-source mathematical AI research project. Contributions accelerate our Zero-VRAM edge AI benchmarks, EVM on-chain oracle verifiers, and public goods academic tooling.

💳 Official Receiving Wallet

Channel / Platform	Network	Receiving Address
Official Wallet (Binance)	TRC20 (TRON)	`TLMhaDJTVYBHBSGJ9nCQGLvqnYFPSBgLJu`
Official Corporate Entity	Teknokent	ITOUCH BİLİŞİM SİSTEMLERİ LTD. ŞTİ. (Çukurova Teknokent)

Official Contacts: vdagli@itouch.com.tr (Corporate) | pcworm@pcworm.net (Lead Research) | ask@answerr.me (Autonomous Agent) | Web: https://answerr.me

Name		Name	Last commit message	Last commit date
.github		.github
assets		assets
benchmarks		benchmarks
dataset		dataset
demos		demos
docs		docs
examples		examples
paper		paper
runner		runner
scripts		scripts
tests		tests
werr		werr
wevv		wevv
.gitignore		.gitignore
.nojekyll		.nojekyll
.zenodo.json		.zenodo.json
CITATION.cff		CITATION.cff
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LEARN.md		LEARN.md
LICENSE		LICENSE
README.md		README.md
README_TR.md		README_TR.md
SECURITY.md		SECURITY.md
index.html		index.html
pyproject.toml		pyproject.toml
sor.py		sor.py

Latest commit

History

Folders and files

Repository files navigation