README

Status 2026-09-06 — J.E. Herizon LLC / Qovaryx Desktop downloads are paused. There is no public installer today. Do not treat this card as a live product, a fully offline runtime, or a published win rate. Company file: https://jehorizon.com/llms.txt · product: https://qovaryx.jehorizon.com This repo is research/reproducibility only.

---

title: README emoji: 🧪 colorFrom: indigo colorTo: purple sdk: static pinned: false

Qovaryx

Sovereign-trained compact AI for trading and finance.

Random-init scratch substrates. Audited compact specialists. Local-first deployment. No borrowed foundation weights, no closed APIs in the inference path. A research project building toward a frontier-grade decision system that runs on hardware a single person can own.

---

🏆 2026-07-20 — Qovaryx 3.0 milestone shipped

A 5-year full-pipeline backtest (2021–2026, $500 starting capital, real + synthetic option corpus, all wired safety gates engaged) on the four public risk tiers:

Every profile positive every year — including 2022 (bear), 2020-style vol shocks in the corpus, and 2025 chop. Aggressive and Extreme require typed operator acknowledgement and enforce mandatory milestone profit-pulling. Universal circuit breakers active on every tier: daily loss halt, correlation cap, drawdown ack, slippage governor. Numbers describe the backtest; live paper begins Q3 2026.

What actually changed in 3.0:

• Four-tier execution-authority taxonomy on every specialist head. REAL_VALIDATED_SCOPED heads can influence a named consumer strategy. SYNTH_VALIDATED_SHADOW runs alongside live decisions, logged, never mutating. VERIFIED_FORECAST_CONTEXT heads are feature-input only. OFFLINE_FORECAST_ADVISORY is research-only. Nothing overrides an order without a signed head→consumer→mutation contract. Full write-up: Head Architecture Advances.
• A proprietary data moat. Multi-decade daily equity coverage (thousands of tickers), multi-year real options-chain snapshots for the highly liquid names, app-faithful multi-symbol option-PnL labels, calibrated synthetic bridges with label-bridging validation, intraday minute-bar snapshots, and a privacy-preserving decision-outcome corpus. Full write-up: The Data Moat.
• BYO-AI chat with 7 providers. OpenAI, Anthropic, Google, xAI, OpenRouter, Ollama (local), and custom OpenAI-compatible. Keys encrypted via OS keychain. Per-context opt-in for what the AI can see. BYO-AI is for chart analysis and position review — it is not on the order-execution path. The trending "ChatGPT / Claude will trade for you" wrappers are not consistently profitable when benchmarked out-of-sample; general-purpose LLMs weren't trained to price options or size risk. Different job, different tool.
• CPU-only inference. No GPU, no cloud round-trip. The full trading cluster runs on any x86-64 from 2018 or newer with 8 GB RAM.
• 13 shipping options strategies dispatched under the same risk gate: iron condors, credit spreads (both directional variants), PEAD drift (30-DTE and 45-DTE), breakouts, VPA long calls/puts, tactical allocator, VIX-crush credit spread, tail hedges.
• Deep cleanup. 22 previously-dormant safety features wired end-to-end and ~1,750 lines of dead code removed. 80/80 focused test suite. Codex ship audit: APPROVE / HIGH CONFIDENCE.

Implementation specifics — exact training recipes, routing heuristics, sizing math, gate thresholds, corpus labeling functions — are intentionally withheld. The framings publish; the recipes do not.

Trading options involves substantial risk of loss. Backtest results are historical simulation, not a forecast. Not financial advice.

---

What lives here

This organization is the home for the Qovaryx model lineage — the published artifacts of a project arguing that compact, locally-trainable AI is a distinct research target, not a smaller version of frontier scaling.

Scratch-base substrates (random-init — designed to be trained, not chatted with)

These are directed blanks: the architecture, the decision surfaces, and the design choices are baked in; the weights are random. They exist so a single researcher on a single consumer GPU can train a specialist model end-to-end without renting a datacenter.

• qovaryx-1b-scratch-base — 1B parameter own-base decoder. Fits on a 16 GB consumer card under QLoRA NF4. The publication target of the substrate line.
• qovaryx-350m-scratch-base — 350M parameter mid-size substrate. Fits on a 12 GB card. The deployment-friendly size.
• qovaryx-50m-scratch-base — 50M parameter proxy substrate. Fits anywhere. Used for pipeline validation before the larger sizes commit to training time.
• qovaryx-3b-scratch-base — 3B parameter scratch-base extension of the same lineage.

All four share a modern compact stack: Grouped-Query Attention (GQA), Rotary Position Embeddings (RoPE), pre-norm RMSNorm, SwiGLU feed-forward, tied embeddings, native multi-token-prediction (MTP) heads, a four-class decision head, optional chart-patch encoder for vision input. Apache-2.0 weights, Apache-2.0 reference trainer.

Trading-application lineage (trained, working, local-first deployable)

The 9B chart-reading lineage that operationalized the same disciplines on a larger backbone before the current Qovaryx app surface:

• vfai-x-3.5-9b-options — legacy VFAi-X lineage page retained for reproducibility and research.
• vfaix-vpa-options-trader — VPA-flavored legacy variant.

Deployed: Q-Chat router

The first operational artifact in the lineage — a sovereign-trained compact specialist intent router, audited and deployed live on free Hugging Face CPU infrastructure, serving a real community Discord. Read more in the public devlog entry on the deployment.

You can interact with it directly through the project's Discord community — /qchat ask <question> — and the model will route, retrieve, and respond. No closed model anywhere in the loop.

---

Discipline

This project is built under six explicit commitments:

• Local sovereign AI — the deployed system runs on hardware the operator owns. No silent dependency on a remote API.
• Compact cognition — small is not the apology. Density per stored bit, routing per active parameter, structure per training row.
• Verifier-governed inference — the wrapper around the model is part of the architecture, not a deployment afterthought.
• Adaptive compute — the model thinks harder on the inputs that deserve it and cheaper on the inputs that do not.
• Execution-aware intelligence — gross alpha is not a result. Numbers must survive realistic costs.
• Intelligence per watt — the only honest unit of progress under a power-, memory-, and bit-constrained budget.

These are first-class architectural decisions made at the very first commit. They are not deployment optimizations.

---

Where to go next

Implementation specifics — exact training recipes, routing heuristics, gate thresholds, curriculum mixing ratios, verifier internals — are intentionally withheld. The framings are publishable; the recipes are not.

---

Qovaryx is an ongoing research effort. The work is in motion. The claims are provisional. The constraint is the point.

This is research and infrastructure writing. Not financial advice. Not a trading signal.