Quantitative developer & independent researcher
Time series across domains · AI agents · verification & reproducibility · data infrastructure
neozork.github.io · QWC evidence atlas · Live Lab · ORCID 0009-0000-3773-5789 · Zenodo · OpenAIRE · LinkedIn
Testing what data and AI systems really support, and publishing what fails.
| Time series, any domain | Regimes, thresholds, change detection: markets and non-market signals alike |
| Quantitative research | Backtest validation, market-data pipelines, MetaTrader / MQL for 15+ years (research and paper trading only) |
| Econophysics & quantum-inspired modelling | QWC, below |
| AI agents & local LLM infrastructure | Local-first agents on Apple silicon, MLX, pipelines with validators and human gates |
| Verification & reproducibility | Evidence cards, frozen protocols, hashed sources: Monte-Neo, ClaimBound Evidence |
| Knowledge engines | Content-addressed claims in a typed graph, immutable snapshots, agents that only propose |
| Data infrastructure | Collectors, Postgres / Redis, dashboards, monitors, backups with restore drills |
A quantum-inspired, wave-based framework for time series, developed independently and audited the way any other claim should be: 26 named laws, 5 axioms, 27 pre-registered experiments, hash-frozen before they ran. The public record includes what failed.
- One replicated advantage: an unmodified alarm threshold stays usable on 34.2 % of 237 source→target pairs from 23 public sources, against 19.8 % for a range check. Replicated twice; a blind third run was inconclusive.
- Refuted or not supported: directional prediction, shape-change early warning, wave rules, a published density normalisation. All published.
- Where measurement disagrees with theory is on the page, not hidden.
Mechanics, formulas and parameters stay private; the audit is public: explore the atlas →
- Monte-Neo · MIT · PyPI · MCP registry. The independent verifier for trading strategies written by AI agents and humans: look-ahead bias, hidden costs, overfitting. CLI, GitHub Action, MCP server, signed certificates. In development, not yet released: a phantom-profit check, planned alongside my first paper (not yet published).
- ClaimBound Evidence · Apache-2.0 · PyPI · conda-forge. Evidence cards for narrow public AI, ML and data claims: frozen protocol, hashed sources, exact status, claim boundary. Started 29 April 2026 as OpenWave, renamed ClaimBound on 1 May 2026 (git history). Most public cards are single-operator, not independently reproduced, and the project says so.
Neo-Planet (public events → causal chains → Monte Carlo, paper only) · Alpha Cascade Factory (a fail-closed hypothesis factory: 4,054 hypotheses generated, none validated) · Canonis (evidence-bound knowledge engine) · NeoArc (verification-first code generation). Described, not open-sourced: platforms.
- Register the hypothesis and the test before looking at results.
- Freeze protocol and data with hashes. Never revise thresholds afterwards.
- Test out of sample, net of realistic costs.
- Publish failures with the same care as successes.
- Ship the experiment so a stranger can recompute the number.
Since 2024 new code is written AI-assisted (Claude Code, Cursor, local models). I set the questions, design the architecture and validate every result.
ClaimBound is independent of EviBound (R. Chen, arXiv, 28 Oct 2025), which I found after development started: EviBound gates the execution of research agents, ClaimBound records what one public claim supports and where it stops. More: related work.
NeoZorK is my research and engineering handle. It is not affiliated with the gaming channel of a similar name on YouTube, and ClaimBound Evidence is not affiliated with claimbound.com or any commercial product using a similar name.
© 2026 Rostyslav Shcherbyna (NeoZorK). All rights reserved. Nothing here is investment advice. Ukraine · remote (UTC+2/+3).