What jarvispoker is, and what it isn't

How it works

The pipeline has three stages: a vision parser reads the cards, positions, stacks, and street from a photo; a solver computes a GTO action distribution for the resulting spot; a display layer renders the frequencies, equity, and pot-odds threshold alongside a short explanation. Each stage is independently inspectable on the result page, so you can see what was read, what was assumed, and where to fix the input if something looks off.

The GTO computation uses an open-source MIT-licensed Discounted Counterfactual Regret Minimization solver written in Rust: exinori/DCFR-SOLVER. DCFR sits in the well-studied CFR family of algorithms used by the major published poker AIs. The solver is run offline on precomputed spot keys; results are cached and served from the database, not solved live per request.

What it is for

• Reviewing a hand you played after the session. Photograph the spot, get the action breakdown, study the reasoning.
• Sanity-checking a tricky preflop or river decision against a solver baseline, without paying for or learning a desktop tool.
• Learning what balanced play looks like in common spots, by seeing the mix the solver returns rather than memorising a single “correct” answer.

What it is not

• Not for live play. Using a solver-assist tool while a hand is in progress on a regulated site is against most operators' terms of service and can get accounts banned and balances seized.
• Not a real-time HUD. There is no live overlay, no per-villain tracking, no exploit model. Inputs are static photos of finished or paused spots.
• Not gambling advice. The math is the best answer the solver has. Whether you should be playing the hand, the session, or the game at all is not a question the tool can answer. Bankroll, tilt, table selection, and the legality of play in your jurisdiction are all on you.

Open work

The bet tree, solver configuration, and per-spot validation harness are intentionally narrow at launch: a single bet size on each street, a fixed stack-depth band, and a small set of positions. Bigger trees are not technically harder but get expensive fast; expansion happens as solve-cost data justifies it.

An independent exact best-response oracle has checked the solver's play on river spots, not just its own self-report. That same check has not been run yet on flop and turn solves, and the cache is keyed at the flop, so that is the street every real cache hit rests on. Those solves are not known to be wrong: an earlier worry that they could not tell one runout card from another is resolved, and what is left is proving them exact.