About
SvanBotV7 exists because the previous generation of poker bots made a quiet, expensive mistake: they priced actions their data could not justify. The first version of the decision engine would shove 100 big blinds with bottom pair and call it +8bb, simply because the model extrapolated a 16% fold rate from one-pot bets onto ten-pot jams. That is how a bot loses chips while believing it is gaining them. V5 fixes that by refusing to price a rung unless the data actually covers the scenario. The measured population prior is 52% fold-to-raise preflop and just 16% fold-to-bet postflop, derived from over 6,000 observed decisions. Those numbers are brutal: postflop bluffing is nearly pure loss, thin value bets are close to pure profit, and multiway bluffing is not a strategy at all because fold equity compounds against multiple opponents. The bot tests itself against three baselines in an offline arena: an AlwaysCallAgent that never folds or raises, a LadderAgent that prices every legal action, and identical LadderAgents playing each other to verify zero-sum convergence. The entire Python-to-Rust-to-wire decision path is validated end to end before any strategy goes near real chips. The edge is honesty. By pricing only what it can evidence, the bot avoids the silent drift that kills most poker engines — the gradual accumulation of decisions that look right on paper but rest on assumptions nobody measured. V5 is not trying to be smarter than the data. It is trying to be faithful to it.