Hindsight

A Brier score of 0.250 and a log loss of 0.693 are what shrugging and saying “fifty-fifty” to everything earns you. Lower is better on both. Log loss is the harsher of the two on purpose — it punishes a confident mistake far more than a cautious one, which accuracy cannot see at all. Void means a bet that never got settled — counted apart, never as a win.

The register

One square per day, shaded by how much of that day survived contact with the truth. A day it got entirely right is plain ink. Only the misses take colour.

Per question

Every question scored on its own record, against the two numbers that decide whether the score means anything. Coin flip is what saying “fifty-fifty” to everything earns. Base rate is how often the event simply happens — a question whose answer is “yes” nine days in ten hands you 90% accuracy for nothing, and without this column there is no way to tell that apart from skill. Sample size is printed on every row because at these numbers it is the most important figure on the page.

Rules and models

Every question is answered twice — once by a hand-written rule, once by a regression trained only on bets already settled. The rule is the control, and it is never retired however badly it does. Judged on shared days only: the model may not bet until it has 25 settled examples, and counting the days it sat out would flatter whoever drew the easier ones.

Brier decides it, not accuracy — accuracy cannot tell the difference between a cautious miss and a confident one.

Calibration

The one figure that cannot be rushed. Each mark is a confidence band: how sure it claimed to be, against how often it turned out right. On the dashed line means honest. Below it means bluffing. This needs months of evidence before it means anything, which is the whole reason the register is kept daily.

Receipts

Most recent first. Losses are not quietly dropped — that is the entire point.

Why this is hard to fake