Universal Fractal Natural Language Decision Map: Real-Time Edge Triage Across Heterogeneous Domains
Volkan Dağlı, Volkan Dağlı, Zerrin Dağlı, Zerrin Dağlı, Dağhan Dağlı, Dağhan Dağlı
cs.NE · cs.AI · cs.CL
Quick verdict
Proposes weight-free Mandelbrot-boundary triage engine; headline 93.1% risks in-sample seed tuning, and 0% injection bypass rests on 10 attack classes.
Summary
The paper proposes werr, a deterministic 'zero-tensor' decision engine that replaces learned weights with 24-byte coordinate seeds placed on the Mandelbrot set boundary; decisions (Boolean noul, categorical choice, ordinal score) are read from multi-scale escape-time dynamics on a 36x36 grid. A domain projector maps state dictionaries and lexicon-matched tokens to coordinate perturbations; seeds are optimized by genetic search. Five innovations are described: auto-seed routing, a semantic token damping filter (T_desc=0.045), multi-scale tripod fusion, a margin-expansion operator, and a Z/9Z-motivated grid discretization. Headline results: on the authors' five-domain campaign (N=326), macro accuracy rises from 62.1% (monolithic) to 93.1% with per-domain seeds; on the public JevBench suite (231 tasks), 100% wire-protocol compliance, 55.4% zero-shot accuracy (vs 31.8% random) and 81.65% on calibrated subsets; prompt-injection bypass 0/10 attack categories; median latency 3.31-7.08 ms on CPU. The paper also sketches MCU (~2 KB SRAM) and EVM oracle deployments (21,438 gas per Boolean reflex). Main limitations: the 93.1% result is likely in-sample (Table 2 seeds report F1=1.000 on the same domains evaluated in Table 3); the security claim rests on 10 attack classes (CI upper bound ~31%); competitor comparisons and energy figures (37,000x) are estimates, not measurements; and the Z/9Z 'resonance' argument is numerological. Code and telemetry are claimed public.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- werr produces deterministic, strongly-typed triage decisions (noul/choice/score) from 24-byte Mandelbrot-boundary seeds with zero stored weight tensors at sub-10 ms CPU latency.
- Per-domain seed routing raises macro accuracy from 62.1% to 93.1% (+31.0 points) on the authors' five-domain campaign (N=326).
- The semantic damping filter (T_desc=0.045) blocks all 10 evaluated prompt-injection classes while cutting mean escape iterations by 45.8% and accelerating inference 2.5x.
- On the public JevBench suite (231 tasks) werr achieves 100% wire-protocol compliance, 55.4% zero-shot accuracy versus a 31.8% random baseline, and 81.65% on calibrated subsets.
- The architecture extends to microcontrollers (~2 KB SRAM) and to an in-EVM oracle consuming 21,438 gas per Boolean reflex.
Strengths
Deterministic strongly-typed output contract
100.00% TypeSafe wire compliance across all 231 JevBench tasks with 0 schema violations (Section 5.8) directly addresses the fragile regex-parsing failure mode of LLM triage described in Section 1 and Table 1.
Extreme resource envelope with concrete engineering artifacts
0 stored weight tensors, a ~2 KB SRAM runtime sketch for Cortex-M (Section 6.1), and a working on-chain implementation with measured costs (21,438 gas per noul reflex, 23,150 gas for a Uniswap v4 hook, Section 6.2) are non-trivial, verifiable engineering results.
Open-science posture
Honest interval reporting on the security headline
Multi-level ablation effort
Weaknesses
In-sample contamination of the headline accuracy
Security claim severely underpowered
Decorative mathematics
Unmeasured baselines and estimated savings
Internal inconsistencies in reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes **werr**, a 'Universal Fractal Natural Language Decision Map': a deterministic triage engine that stores no neural weight tensors. A 24-byte coordinate seed (cx, cy, zoom) anchored on the Mandelbrot boundary is perturbed by a domain projector, and escape-time dynamics on a 36x36 grid (Mmax=36) yield three typed primitives — noul (Boolean), choice (categorical), score (ordinal). Five contributions are claimed (Sections 1, 3-4): auto-seed routing; a semantic token damping filter (T_desc = 0.045); multi-scale tripod fusion across zooms {0.60, 1.00, 1.60}·z0; a pitchfork-style margin operator; and a Z/9Z-motivated grid discretization. Headline claims: 93.1% macro accuracy on a five-domain campaign (N=326, Table 3); 100.00% TypeSafe wire compliance, 55.4% zero-shot and 81.65% calibrated accuracy on JevBench (231 tasks, Section 5.8); 0.0% prompt-injection bypass (Section 5.4); 2.5x self-acceleration from damping (Section 4.2); a ~37,000x energy advantage (Section 1.1); and deployments on microcontrollers (~2 KB SRAM, Section 6.1) and in-EVM oracles (21,438 gas per noul reflex, Section 6.2).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does werr really use no neural network weights?
It stores no weight tensors: decisions come from iterating the Mandelbrot escape map from a 24-byte coordinate seed. However, its knowledge lives in hand-calibrated lexicons, domain seeds, and constants (e.g., T_desc=0.045, EMA alpha=0.03) — parameters in another storage format, tuned offline via genetic search.
What accuracy does werr achieve?
On the public JevBench suite (231 tasks) it reports 55.4% zero-shot accuracy (random baseline 31.8%) and 81.65% on calibrated subsets. On the authors' own five-domain campaign (N=326) it reports 93.1% macro accuracy, but the domain seeds were genetically tuned with in-sample F1=1.000 on the same domains, so that figure risks in-sample inflation.
Is werr robust to prompt injection?
The paper reports 0 bypasses out of 10 attack categories. With only 10 samples, the 95% confidence interval upper bound reaches roughly 28-31%, so this result is far from proof of robustness; adaptive attacks at larger scale were not tested.
Is the code and data available?
The paper links a GitHub repository (github.com/pCwOrM/werr, plus werracle for the EVM oracle), an open telemetry dataset (1,150+ decisions as JSONL on api.answerr.me:4431), Zenodo DOIs, and a self-run JevBench evaluation (issue #10). Availability was not independently verified in this review.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are available only via captions in the extracted text; visual claims (Fig. 3 EMA convergence, Fig. 5 latency distribution) could not be verified.
- The OOD battery sentence in Section 5.3 is truncated mid-result ('maintained 100...'), so the claimed out-of-domain accuracy figure is unverifiable from this extraction.
- External artifacts (GitHub repos, JevBench issue #10, telemetry endpoint) were not accessed; availability statements are taken from the paper.
- The v2 preprint self-describes as submitted for peer review; no venue acceptance status is stated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.