Deep-stacked poker is a multi-street game: preflop is one decision among many, and a marginal preflop spot can still be profitable because of the postflop maneuvering it buys — position, implied odds, the ability to fold a bad turn or bluff a scared one. Below roughly 15 big blinds effective, that buffer disappears. A single opening raise commits so much of the stack that there's rarely room for a second bet before someone is all-in anyway, so the entire hand's decision tree collapses to one binary choice: jam or fold. This app models that directly — the push/fold node set has no "raise small" action at all, only fold and jam.

The right jamming range at a given stack is a risk/reward balance against the pot's fold equity, not a hand-strength ranking. Shoving risks the whole stack but wins the blinds and antes outright every time the table folds — and that fold equity is the real engine. It's not a coinflip against a random hand: a caller only continues with their calling range, so a weak shove's equity when called is usually well below 50% (often 30-38%). A shove profits only once that discounted called-equity term stacks on top of the fold-equity term — which is why the ranges below aren't the whole deck at any stack. As the stack shrinks, the blinds become a bigger fraction of it, so the breakeven jamming range widens at every seat. Drag the slider from 15bb down to 8bb on any seat and watch its grid fill in, driven by the same pf:<stack>:jam:<seat> nodes (32 of them, one per seat × stack) the Trainer grades against.

Honesty note: unlike the postflop model, these ranges come with a real provenance story, and it's read live off each node below rather than asserted in prose — all of them are tagged solver-derived because they're the output of an in-house offline equilibrium solve of this exact game: nine seats, jam-or-fold and call-or-fold only, chip EV (not ICM), no antes, uniform stacks, over-calls modeled. All 8 seats' jam charts are tagged solver-derived, and the solve is gated on reproducing published Nash anchors (the blind-vs-blind call-off charts) within ±3 percentage points; its measured leftover exploitability ships in the data's metadata instead of being claimed away. Treat them as the app's most trustworthy charts — still a model with stated assumptions, not gospel. Pick a different seat below and the confidence tag updates to match.