Blackjack RL

Four studies, one experimental design: give a learning system a domain where the truth is exactly known, ask whether it can rediscover what we can prove — and then ask again, each time with a piece of the answer key taken away.

Blackjack is the rare domain that makes this honest. Optimal play is exactly computable, the house edge is published to decimal places, and card counting provides a genuine, precisely measurable edge. Every claim in the series can be checked against ground truth — never “the agent seems good,” always a measured distance from provably optimal.

How the arc unfolds

It starts with the instrument, because if the measuring device is wrong, nothing built on it means anything: a from-scratch simulator validated against published blackjack mathematics across ~80 million hands, which every later result is measured on.

Then the question, three times. With the full answer key on the table, tabular Monte-Carlo control rediscovers ~93% of optimal play — and the audit traces the missing 7% to coverage, not method. With the key held at arm’s length, a neural network takes on the same game: generalization repairs the table’s coverage gaps and pays for the smoothness exactly at the policy’s sharp boundaries — capability as a measured trade, not an upgrade. And finally with no key at all, an agent must learn bet sizing over whole sessions, where the optimum is derivable (Kelly, 1956) but was never learned — and proving why it fails, with an oracle control and two falsified explanations, became the result.

What it demonstrates end-to-end

The series is less about blackjack than about a method: validate the instrument before believing its output; audit against a reference instead of a vibe; treat capability as a trade to be measured; and when something fails, test the failure like a success — controls, falsifiable predictions, and the honest gap between “the effect is real” and “the effect is learnable.”