Quantitative Developer & Independent Researcher
Testing what data and AI systems really support, and publishing what fails.
An independent research and engineering programme across time series, AI agents, verification and data infrastructure. Built on pre-registered experiments: 27 frozen protocols, 46 logged findings, one threshold-transfer result replicated twice, and a public record of refuted claims.
Registry snapshot, October 2026. Every row is on the QWC page.
Domains
What the work covers
Time series run through almost all of it, in markets and outside them, alongside AI agents, verification and the infrastructure that research needs.
Domain 01
Time series, any domain
Regimes, thresholds and change detection across markets and non-market signals. One threshold-transfer test used 23 public sources, among them spacecraft telemetry, cardiac signals and water treatment.
Domain 02
Quantitative research & trading systems
Research-grade backtesting, validation and market-data pipelines; MetaTrader / MQL tooling for 15+ years. Research and paper trading only.
Domain 03
Econophysics & quantum-inspired modelling
QWC: a framework of named laws, audited by pre-registered experiments, with superstatistics and heat-kernel ideas tested against measurement.
Domain 04
AI agents & local LLM infrastructure
Local-first agents and coding assistants on Apple silicon, MLX tooling, multi-step pipelines with validators, retries and human gates.
Domain 05
Verification & reproducibility
Evidence cards, frozen protocols, hashed sources, backtest verification: Monte-Neo and ClaimBound Evidence, both open source.
Domain 06
Verification-first code generation
NeoArc: cache, then symbolic derivation, then a small model, every result checked by a real verifier.
Domain 07
Knowledge engines
Canonis: content-addressed claims in a typed graph, immutable snapshots, agents that may only propose.
Domain 08
World events & probabilistic forecasting
Neo-Planet: public event data, causal chains and Monte Carlo on GPUs, paper trading only.
Domain 09
Data infrastructure & operations
Collectors, Postgres and Redis, dashboards, monitors, multi-tier backups with restore drills; 50+ scheduled services.
Domain 10
Native & GPU computing
Swift, Metal, C++ and MLX for compute-heavy research code on Apple silicon.
Domain 11
Developer tooling & automation
Packaging, CI-style checks, command-line tools and MCP servers.
Domain 12
Quality assurance & project delivery
Years as QA team lead, project manager and system administrator: the discipline behind the experiments.
Flagship research programme
QWC: Quantum Wave Core
QWC stands for Quantum Wave Core: an independently developed, quantum-inspired and wave-based analytical framework for time series, developed and audited first on financial markets. Audited the way any other claim would be: against standard methods, out of sample, with the failures published.
Held up
Measured structure
Heavy tails, volatility clustering, near-martingale increments, a persistent business-time clock: measured on public data. Mostly confirms the literature, and says so.
One advantage
Thresholds that transfer
An alarm threshold applied unchanged to a new source stays usable on 34% of pairs against 20% for a range check. Replicated twice; the blind third run is inconclusive.
Failed
Direction, shape, waves
Directional prediction, shape-change early warning, wave rules and a published density were tested and did not survive. They are published.
26 named laws, 5 axioms, 17 experiments
Each law is a measurement with a verdict, grouped by what it is, searchable on the atlas. Each experiment was frozen before it ran.
Where theory and measurement disagree
- Schrödinger-type evolution and the Brownian heat kernel
- A published density that does not normalise
- Directional predictability against market efficiency
- Elliott wave rules as a source of prediction
- Brownian motion and Gaussian increments
Selected findings
What the record shows
T1 · Cross-domain transfer
An unmodified threshold stays usable more often across 23 independent public sources
237 source-to-target pairs from spacecraft telemetry, cardiac signals, water treatment, cloud and web-service metrics. Share of pairs where the…
T3 · Blind replication
Third, blind replication on sealed families: same direction, not confirmed
Pre-registration frozen with a SHA-256 hash and data hashes recorded before the run; thresholds not revised afterwards. 5 of 20 pairs usable against…
N2 · Direction
Directional timing from proprietary wave indicators: no pass in forty attempts
Forty pre-specified timing experiments across direction, filtering, sizing and risk overlays: none passed. Direction was closed on four independent…
N7 · Automated search
An automated hypothesis factory produced 4,054 hypotheses and none validated
4,004 rejected, 48 blocked, 0 validated. 99.3 percent of rejections carried no recorded reason, so the largest corpus of negative results taught…
What the experiments run on
Research platforms
Event-driven research, a hypothesis factory, a knowledge engine and verification-first code generation.
Research platform · paper trading only
Neo-Planet
Collects world events from free public sources, normalises them, and builds causal chains from event to consequence to market.
Research platform · fail-closed
Alpha Cascade Factory
A hypothesis factory: intake, point-in-time dataset binding, cheap triage, historical proof (walk-forward, combinatorial purged validation, deflated Sharpe, overfitting probability, doubled costs), then evidence gates.
Proprietary · pre-alpha
Canonis
A local-first, deterministic engine for evidence-bound knowledge: claims as content-addressed objects in a typed graph, evidence with explicit statuses, immutable snapshots, and AI agents that may propose but never commit..
R&D · local-first
NeoArc
A verification-first alternative to slow LLM-only code generation: cache, then symbolic derivation, then a small model as the last resort, every result checked by a real verifier..
Tools anyone can run
Open source
MIT · PyPI · MCP registry
Monte-Neo
The independent verifier for trading strategies written by AI agents and humans.
Apache-2.0 · PyPI · conda-forge
ClaimBound Evidence
Where is the evidence? Evidence cards for narrow public AI, ML and data claims.
Interactive
Run it in your browser
Six working demonstrations on live public data, frozen registry snapshots or clearly labelled synthetic data. Free, static, and nothing leaves your browser except the public API calls you can see.
Live Lab
ClaimBound Evidence
Live repository and package data, plus a claim you can freeze and watch drift.
Live Lab
Monte-Neo
Live release data, plus a simulator that shows how the best of many strategies is just luck.
Live Lab
QWC
The 26 laws as a clickable map; every verdict, every experiment.
Live Lab
NeoArc
Cache, symbolic derivation, model, verifier: the routing principle, running.
Live Lab
Neo-Planet
This week's real earthquakes, normalised and pushed through a Monte Carlo chain.
Live Lab
Canonis
Claims as content-addressed objects, sealed snapshots, agents that may only propose.
Seven letters, seven pillars
The seal
- Name the claim before you test,
- Evidence will do the rest;
- Open the data, open the code,
- Zero faith in a hidden road;
- Out of sample, or it isn't true,
- Rerun it, so that others do,
- Kindly share what you've been through.
Seven letters, seven pillars of the work: verification-first code, evidence, open source, a fail-closed factory, the observed world, research and a knowledge engine.
The person
Rostyslav Shcherbyna
Since 2024 new code is written AI-assisted (Claude Code, Cursor, local language models). I set the questions, design the architecture and validate every result; the tools type faster than I do.