The data moat

People ask us: "how do you compete with the frontier labs on model quality when your training budget is 1% of theirs?" The honest answer is that model architecture is not where the fight is won. The fight is won on data that nobody else has bothered to collect, curate, and grade. That's the moat.

The corpora that sit behind the specialist cluster

Different heads use different task-appropriate subsets of the corpora assembled for Qovaryx. In broad strokes:

• Multi-decade daily equity corpus. 7,195 US stocks and 1,344 ETFs in the source archive, with daily OHLCV reaching back to 1962 for the longest-lived symbols. Coverage varies by listing lifetime and data quality, but it spans far more than a single recent bull cycle.
• Real options-chain snapshots. Multi-year full-chain history for a targeted set of highly liquid names: SPY, QQQ, AAPL, NVDA, TSLA, and more. Contract-level detail: bid/ask/sizes, IV, all Greeks, open interest. The kind of data that is normally paywalled at $300+ per symbol-year.
• Multi-symbol option-PnL labels. Where real chains and required quotes exist, candidates are replayed through app-faithful strategy consumers with contract, spread, and cost assumptions recorded alongside the result.
• Synthetic bridges. Where older real chains do not exist, synthetic option labels expand hypothesis coverage. Overlap tests quantify their error against real-chain periods. Synthetic evidence remains shadow-only unless a separate real-data review earns a broader role.
• Intraday minute-bar snapshots. Selected symbols with minute-resolution OHLCV going back several years for intraday-sensitive heads.
• Chart composites. Roughly 500 MB of rendered chart images supported earlier vision-head research. Automated chart submission is disabled in the 3.0 release; those research assets are not evidence of execution authority.
• Local decision ledger. The application can retain its own decisions and outcomes for the user's audit and diagnostics. The 3.0 distribution does not upload that ledger or collect a new training corpus from customer machines.

What we're not going to publish: the exact snapshot dates, the specific symbol enumeration, the label functions, or the label-bridging validation code. Those are the moat.

The grading discipline

Data is only a moat if you use it properly. Two rules govern every training run:

• Leakage scrubbing. Every training set is audited for label leakage before a single training epoch runs. When we found in June that a chunk of our earlier synthetic labels was contaminated by split-adjustment glitches, we retracted the affected heads and re-labeled the corpus. The heads that survived the re-audit are the ones running in 3.0.
• Holdouts stay holdouts. Time, symbol, and regime holdout slices are reserved for final evaluation. Training and promotion records document the split and consumer used; a promising training metric cannot substitute for the holdout result.

Why the moat compounds

Every week we spend adding new labeled samples, new regimes, new failure cases into the corpora is a week that a competitor who is fine-tuning a foundation model on public data cannot make up. Public options data ends where somebody paywalled it. Our corpora start with what's public, add what we've paid for, and add what we've labeled ourselves — the last of which is not for sale at any price.

This is why we take the "no recipe leakage" policy so seriously in our writeups. We publish the framing (what we tried, what we retired, why). We do not publish the labeling function, the specific eval slices, or the exact feature sets that survived the funnel. That's the recipe. The recipe is the moat.

Model architecture is a photocopiable commodity. A well-graded corpus of things that actually happened is not.

See the companion post on how the specialist cluster evolved using this data foundation.