Graduation Test

What this paper can tell you

This is the last course and it teaches nothing new. It is a paper: 24 questions pulled from the whole book, across 7 topics — preflop decisions, nut awareness, counting outs, price, the stack-to-pot ratio, planning the turn, and multiway discipline. Every one of them is a question you could have met in the course it came from. None was written specially for this paper, because a graduation test that invents its own questions is testing a course you never took.

Now read the arithmetic of the paper itself, honestly. 7 topics over 24 questions is between 3 and 4 questions each. 3 answers cannot produce an accuracy rate for anything: get one wrong and the topic reads as a third of a disaster; get all three right and it reads as mastery. Both readings are noise wearing a percentage sign.

A score from a single sitting is a score, not a diagnosis.

Applies to any block of questions small enough that one answer moves the rate by a visible amount — which is every per-topic slice of a 24 question paper. It does not apply to the total, which has all 24 behind it.

The mirror-image mistake is refusing to report anything at all. A total out of 24 is a real number and worth reading; so is the list of exactly which questions you got wrong. What is not worth reading is a per-topic percentage built on 3 answers — so we show one and refuse the other, rather than hiding both.

So the graduation report is in two halves. The paper gives you the total and every question back to reread. The picture of where you are weak comes from somewhere else entirely: every answer you have given anywhere in this trainer, grouped by the leak tag on the question. A tag with fewer than 8 answers behind it gets no percentage — it gets the words "not enough data" and the count so far. The 8 is one of the rules of this course, not an equilibrium solution; we would rather hand you a visible gap than a number built on 3 answers.

After you graduate

What you have finished is a book whose every percentage can be traced to a command, a seed or an enumeration in this repository, and whose thresholds are all rules of this course, not an equilibrium solution. That is the whole claim. It is smaller than what a solver claims and larger than what most trainers can show you.

What is still open, in one paragraph, so that nobody has to go looking for it. Judgement coverage in full-street play was 53.3% on our probe; the 194 points where you led a turn or a river, or raised, got 0 verdicts between them, and a flop where you have already made a hand gets none either — the flop rule we calibrated is a rule about draws. Your leak picture needs 8 answers behind a tag before it says anything. The opponent is a scripted opponent, so beating it is not evidence about a real table — it is a way of checking your own decisions against the criteria that exist.

Trust a number in this trainer exactly as far as the page that produced it says you should.

Every number here carries its method: exact enumeration, a seeded sample with a stated tolerance, or a reference range that gives an action and never a frequency. When a page tells you which of the three it is, that is the range of the claim, not a disclaimer.

The failure mode this course worries about is the opposite of scepticism: a plausible-looking number with nothing behind it. If you ever find one on these screens with no method attached, it is a bug in our work, and the report button on the feedback screen is the right place for it.

Still uncovered after graduation, listed rather than glossed over: turn and river leads and all raises get no verdict; the stack-to-pot bands are calibrated for heads-up only, so multiway pots get no threshold from us; implied odds get a qualitative boundary and no number; postflop range advantage has no data in this project at all, so the sizing rule refuses to answer instead of guessing. There is no date on any of them, and each only happens when there is something real to calibrate against.