A market instrument that mostly says no.

Qovaryx runs on your own machine and reads the market for you. The thing that makes it worth installing is not how often it finds something — it is how much it throws away first.

Join the release list Watch how it is built Discord

The Qovaryx dashboard: account state, open positions and current signals, running locally on a Windows desktop — Qovaryx 3.0 — dashboard **runs entirely on your machine**

Candidate ledger — what happens to an idea most do not survive

Worked beautifully on the years it was designed against. Fell apart on an era it had never seen.

Cut

Twenty variations, all agreeing. Correlation check showed they were one idea wearing twenty hats.

Collapsed to one

Strong on paper. The result could not be reproduced by a checker that never saw how it was produced.

Cut

Held up on decades it was never shown, survived the duplicate check, and reproduced independently.

Kept

We publish no performance figures. Any return, win rate or accuracy number attributed to Qovaryx did not come from us. What we will describe is the method — and the method is mostly rejection.

Ch. 01 — The application

What it looks like on your machine.

Screens from the build currently in evaluation. Everything runs locally — no market data leaves your machine for a cloud model to read.

Qovaryx charting view with indicator overlays and price history — Charts

Qovaryx options chain with strikes, expiries and pricing — Options chain

Qovaryx strategy builder for composing and testing a position — Strategy builder

Qovaryx order management showing pending and filled orders — Orders

Qovaryx fundamentals view for a selected symbol — Fundamentals

Qovaryx settings including risk limits and broker connections — Settings & risk limits

Questions, or want to follow along? The Discord is where release news lands first and where we answer things that are not worth an episode.

Join the Discord

Ch. 02 — Method

Built to break its own ideas.

A trading tool is only worth the testing behind it. Four things every candidate signal has to get past before it is allowed anywhere near the product.

Validated on eras it has never seen

Ideas are developed on one span of market history and judged on a completely different one. Anything that only worked by memorising the past fails here, loudly, instead of quietly later.

Near-duplicates are collapsed

Highly-correlated candidates are grouped and reduced to one. Twenty variants of a single idea cannot line up and pretend to be twenty independent confirmations.

Nothing certifies itself

Results are accepted by a deterministic checker that reads the finished state, not by the component that produced them. A stage claiming success proves nothing on its own.

Decades of history, including the bad years

The evaluation runs against the long record — thousands of symbols across tens of thousands of trading days — not a flattering window. The years markets broke are the years that matter.

Ch. 03 — Research log

The reasoning, written down.

This has been running close to continuously for a month — evaluation runs, audits, and the arguments behind each design decision. The write-ups below are the ones we can defend in public. Read them before deciding whether to trust the thing.

Currently in the lab updated 2026-08-09

Large-scale candidate evaluation. Ninety-nine candidate signals put through the same held-out-era test, with near-duplicates collapsed before anything is counted.

Running

Evaluation engine rebuilt for speed. One ranking step was doing tens of millions of redundant operations per pass; rewriting it cut that stage from roughly thirteen hours to about one, with results confirmed bit-for-bit identical to the slow version.

Shipped

Audit of our own defect list. Every claim we re-measured turned out wrong in some material way — including a module described as ready to drop in that did not compile.

Published

Counting audit. A summary total was double-counting, because a roll-up row was being summed into a table that already contained its own members. Per-item figures were correct; the headline was not.

Fixed

Every item above is a real incident with a receipt behind it. The ones worth watching end up as episodes on the channel.

Evaluation The brutal evaluationWhy the test is designed to be hostile to our own ideas. Doctrine Abstention is a decisionDoing nothing is a position, and it has to be measured like one. Doctrine Calibrated conviction beats raw confidenceA confident wrong answer is worse than an uncertain one. Data The data moatWhat the long record contains, and why a curated window would be cheating. Architecture Why a cluster of small heads beats one big modelSpecialists that can disagree, instead of one model that cannot. Architecture Head architecture advancesWhat changed in the decision heads, and what it cost. Engineering CPU, not cloudWhy inference stays on your machine — a design constraint, not a limitation. Engineering Sovereign, not borrowedOwn weights, own data, own audit trail.

Ch. 04 — Build log

We publish the failures too.

A documentary series about running this thing, cut from our own incident records. Not a marketing channel — most episodes are about something going wrong and what the receipts showed afterwards.

youtube.com/@TheAgentForeman

New episode every Tuesday, 1:30 PM ET

Episode

Our benchmark went up six times. The model got worse. Both true. The reconciliation is the episode.

Episode

The drive hit zero and the dashboard stayed green. Every number on it was accurate. Every number was measuring the wrong thing.

Episode

We audited our own defect list. Every claim we measured was wrong in some way — wrong file, a module that would not compile, a switch whose fix was the defect.

Ch. 05 — What happens next

No date. A standard instead.

The next build ships when it has been through the evaluation described above, and not before. We would rather be late than ship a trading tool we cannot defend. Join the list and you will hear when it is ready.

Join the release list Discord