Armchair Hockey Scout

Ask the book

2026-27 edition Ask a direct question about this edition of the book — who, how many, how common, which group more — and it is answered with a table computed on this page from the edition’s index: the season line, height and age, the on-ice percentiles, the grades and the traits read out of every report. Ask for a view on named players — what a comparison comes down to, what makes a man — and you get the book’s reading of their own pages, and only those. The model turns your question into a query and phrases the result, or reads the pages it is handed; it never originates a figure, and an answer carrying a figure that is on neither the table nor the pages is withheld. It does not forecast, and it says so when a question is outside what the book holds.

Tables are computed on this page from the 2026-27 edition’s index; readings rest on the players’ own pages; the model never counts. Twenty questions an hour.

The other half of the experiment

Your picks, scored

Every game you call is stored with the model's probability, its lean on each area, your call on each area, the tally, your final answer and — when you went against it — your reason. This is what all of that was worth.

Reading your picks…

What would count as an answer

Not your record. A record on its own says almost nothing: picking the home side every night wins somewhere near 54% of the time and requires no judgment at all. Both baselines are printed above yours for that reason, and the model's number is computed on the games you chose to pick rather than on the whole season, because those are not the same set and the difference would flatter whichever one you looked at second.

The games where you and the model parted company. Every night you both call the same way tells us nothing about which of you is better. Only the disagreements carry information, and there are far fewer of them than there are picks — which is why the number of wins it would take to beat a coin is printed beside the count rather than worked out afterwards.

What you overrule. The areas are tagged by the kind of evidence they rest on: measured is a full season of it, situational is real data on a short sample — form, rest, a goaltender's fortnight — and scouted is the grades in this book and nothing statistical at all. Someone who repeatedly overrules the measured areas in favour of the scouted one is saying what he trusts. Whether he is right is a question with an answer, and it is the last block above.

What this cannot do

A season is not a sample. Eighty picks gives an interval on your accuracy about eleven points wide in each direction. Almost nothing you will see here in one winter is going to be distinguishable from chance, and the page says so in numbers rather than leaving you to feel it.

A pick made after the puck drops is not a pick. Nothing in the app prevents one and the timestamp is recorded, so any pick saved after the start time is set aside and counted separately rather than quietly included.

Only this season. The site carries the live schedule and only the live schedule, so a pick on a game from a previous season has nothing to be graded against. Those are counted and named rather than dropped.

Nobody else can see this. Your rows are yours by row-level security — the database will not return them to anyone else, or theirs to you. That also means the arithmetic on this page cannot be validated against real picks by anyone but you, which is why npm run check:grade checks it against a constructed season instead.