The pilot
Your models, our foundation. Eight weeks, at no cost, on training jobs you would otherwise curate blind.
Before the compute is spent, one call tells you whether curating toward your target will pay. Timestamped before you train, checked against what your run does.
- 01
Send two samples
A seeded random sample of whole documents from your target and from your candidate pool, a few million words in all. Our reference client draws them and records the seed and the full corpus size. The corpus never leaves your machines.
- 02
Receive the verdict
In seconds, on CPU, without touching your model. One of three calls, a confidence tier, a scope card, and a forecast id. On Select, a scoring table comes back so your pool is ranked on your own hardware in a single pass.
- 03
Train, then tell us
Where you can, train both arms: the curated sample and a random sample of the same size. Report the held-out result under the forecast id. The receipt ties the call to the outcome, and the receipt is the product.
Each call carries a confidence tier, the scope card, and a fingerprint of the sample it was made on. Every forecast is timestamped before your first training step.
- A key on our hosted service. Nothing to install, nothing of yours to deploy.
- Fifty calls a day, raised on request. Samples travel compressed, far more room than a forecast needs.
- Verdicts with scoring tables, local keep lists through the reference client, and outcome receipts.
- A written forecast-versus-outcome report at the end, for every run you report.
- A direct line to the founder for interpretation and for fitting the call into your pipeline.
- Run the forecast on at least three real target and pool pairs.
- Where you can, train both arms and report the held-out result. A forecast without an outcome teaches neither of us anything.
- Tell us what a verdict must say to be acted on: formats, sizes, latency, wording.
- Samples only, over HTTPS, processed in memory and never written to disk.
- The target needs about 2,600 words in two or more documents. The pool sample at least four times that. More target data sharpens the reading.
- Document and word counts, a fingerprint of the sample so a receipt is tied to exactly the text it was made on, the verdict, timestamps, the sampling seed, and any outcome you report.
- Never your text. Never the scoring table. Keys are stored hashed and revoked on request, one per organisation.
- Text models
- up to 41.7M parameters, from scratch
- Image and text
- one 256M-parameter model, continued training
- Data kept
- a quarter of the pool
- Language
- English, whitespace-delimited text
Every response carries this card. Tell the client the model size, keep fraction and language you plan to use; when your job is outside the range, the confidence drops one tier and the response says why. The call itself does not change. Larger models are exactly what the pilot is for.
- Free for the pilot. Eight weeks from the day your key is issued. Either side can stop at any time.
- No exclusivity. No publication of partner names or data without written consent.
- The engine is a sealed service. Outputs are forecasts, not guarantees; every training decision is yours.
- Aggregated, anonymised forecast and outcome statistics may improve calibration and appear in our publications. Never partner-identifiable, never text.
A few lines are enough. We read every application and answer by hand.
Questions before applying, a scope you are not sure fits, or a pipeline that needs the call in a different shape: write to us. A person answers.
[email protected]Members reach the founder through the same address, and every email from the program can be replied to.