Which formats should I use, HOLDEM or HOLDEM_binary?
Use HOLDEM if you need per-hand (1326×2) EV and optional per-node strategy matrices. Use HOLDEM_binary for compact post-processing (flat arrays) when you don't need the nested tree.
Do you support multi-way trees?
No, multi-way is not supported.
Can I control exact paths (e.g. check → bet X → raise Y → call)?
Yes. You can craft the pot_fractions grid to include those raise levels, and/or build a tree from an explicit leaf path list if you need a mirror of an external abstraction.
How do I know when the solver is "done"?
You set iters. During solving we compute a convergence metric and, at the end, a Nash distance. You can early-stop client-side based on those values.
What about hand order?
We return per-hand EV/EQ in the engine's native 1326 order. If your consumer needs a different order, apply a reindex on your side (we can provide the mapping).
How many jobs can I submit?
As many as you want, 10,000+ jobs queue fine. They'll run as GPUs free up.
What's my parallelism?
Each job uses 1 GPU. With N GPUs you'll run up to N solves concurrently.
How long are results retained?
For the full duration of your active subscription. You can also opt to store results in your own S3/GCS.
Can I stream partial strategies during solving?
Yes, set requested_nodes + requested_attrs (e.g., "strategy") and we'll push partials periodically.
Authentication details?
Bearer token in headers; optional IP allow-listing. All traffic over HTTPS.
Regional deployment?
Yes, separate infra per client. Choose a region (or closest available); we deploy workers there.
Can I resell or white-label the API?
Yes, no restrictions.
Any usage restrictions?
No real-time assistance at live tables (fair-use policy). Otherwise standard research and product use is welcome.
Error handling & retries?
Treat 404 on the result endpoint as "not ready yet." Use exponential backoff on 5xx. task_id is idempotent for results; scheduling is safe to retry with a client-side idempotency key.