A letter is quality of the decision, not whether you won the pot. We under-claim on purpose: this is GTO-shaped training, not a $99 solution library.
Quality is EV given up versus the coach’s best line. Result is chips. Session copy says both — “you played B, you ran bad, session C.” PGR (the number on Home, Profile, and the felt) is a display map of the same internal rating the server uses to seat you. It is not a second ladder and it is not for sale.
Each hero decision is priced against the line the model prefers in that spot. Capture is the share of EV you kept when more than one line was live. A fold that gives up a big pot on a flush-completing river can be a good letter if calling was worse. A win can still be a leak.
Preflop uses house charts. Postflop uses a rollout against the opponent who was actually there, plus an equilibrium check on heads-up rivers when that axis is defined. Multiway rivers have no unique equilibrium — we say so instead of inventing one.
Train grades as you play. Live is unassisted at the table; after the hand you can import it and get the same coach notes. Live letters are honest estimates from the public log (stacks are inferred). The rating that moves your gem is the result against the field — a different number, on purpose.
Day one, most seats are house AI so a table always exists. Every bot wears an AI chip. The lobby header publishes seated · human · AI. Humans displace bots 1:1. Off-hours stay AI-heavy. Humans and AI share one ladder; you always know which is which.
A “leak” is a recurring spot class that costs EV (street × action × texture), not a frequency stat like WTSD being off a 6-max baseline. Gold for “leak cleared” only pays when the EV class was diagnosed and then actually cleared, with server-written hands backing the spots. Frequency panels are labeled context, not a verdict.