Today's largest AI models are built by optimizing millions of parameters at a time against a loss function. An AI-native hedge fund works the same way. The Tuatara hedge fund's AI system uses proprietary vector embeddings, algorithms and triangulated data, and it keeps changing as it learns and finds new edges and alpha in global financial and information markets. Until now, that work was invisible from the outside. The Fund Optimizations page puts it on the record, live.
What the page is
Fund Optimizations is a datestamped ledger of the fund's AI system at work. Every optimization, fine-tuning run, training run, post-training promotion, calibration and hyperparameter search lands on it as a record with a timestamp, a stage, a market domain and the objective it was aimed at. New records stream in as the work happens, and the totals for each stage update while you watch.
At publication the ledger held more than 20,000 records going back to the end of December 2025, over 5,600 of them from the last 30 days. Those numbers will be out of date by the time you read this, which is the point: the live page is the source.
The stages
Each record belongs to one stage of the learning cycle:
- Training run: a full fit of a model on labelled history, from a cold start.
- Fine-tuning: a warm-start fit, a parameter re-tune, or an accepted parameter sweep on a model or rule set that already exists.
- Post-training: promotion, activation or retirement of a candidate after out-of-sample review.
- Calibration: mapping raw model scores onto outcome probabilities, checked on held-out samples.
- Hyperparameter search: a population search over a bounded parameter space, behind an out-of-sample overfitting gate.
- Optimization: a proposed or applied change to a live control lever, measured against its baseline.
- Evaluation: a backtest or replay of a rule set over a fixed window against a benchmark.
The page charts how each stage's work grows over time as a race between stages, so you can see where the system is spending its effort: mostly evaluation and calibration, with far fewer searches and promotions.
What boosted returns
Activity alone proves nothing, so the page also has a section called What boosted returns. It shows which parameters and which optimizations raised returns, each measured against the system's own baseline: a search's best candidate against the parameter set it was run against, a basket against its benchmark over the same window, and a fine-tuned model against its cold start. Wherever held-out data exists, the held-out result is the one that counts.
That section reports failures as well as wins. At publication, the best search candidates beat the parameter set in use on held-out data in a little over half of searches, and some promotions scored below the set they replaced once the dataset had grown. Those results stay on the page with everything else.
Why some names read "BlackBox"
The ledger is public, but the algorithms behind it are proprietary. Algorithm families and parameters are already shown as opaque ids (AF-xxxx, P-xxxxxx) and plain role words such as "holding period" or "basket width". We have now gone one step further: a randomly selected 20% of the optimization names on the page (models, parameters and control levers) are shown as BlackBox-NNN, with the number drawn at random between 101 and 1001.
The selection was drawn once and is fixed, so a BlackBox name means the same thing every time you see it. You can follow BlackBox-611 or BlackBox-642 through the ledger and watch its results change, without the page revealing what it is. Drawing a new selection on every page load would eventually reveal every name, so we don't. Measurements, datestamps and outcomes are unchanged. Only the names are hidden.
How to read it
Start with the running totals at the top to see the scale. Then look at the race chart to see which stages are growing fastest. Finally, open What boosted returns to see which work actually lifted returns on data the system had not seen. The ledger below that is the raw record: filter it by stage or domain and page back as far as you like.
See the ledger live: every optimization as it lands, running totals per stage, the race chart, and what boosted returns.