How the Lab works

How do you settle a 30-year argument?

We built our own numbers to do it. Every Experiment runs the same way, whoever we like better going in, and every step is on show.

Explainer 1

What is LAB IMPACT?

J Swish on what happens to the scoreboard when you're on the floor. 43 seconds. Captions on screen.

  1. Lock the question first

    "Who was better in their prime?" We pick each player's prime window by a fixed rule (most All-Star and All-NBA seasons, equal length) before looking at a single score.

  2. LAB IMPACT: what happens to the scoreboard when you're out there

    Every time a five-man line-up is on the floor, we know what it scored and what it gave up. Run that across every line-up in three seasons and you can work out how much each player moves the scoreboard, whoever he plays with and whoever he plays against. That's LAB IMPACT, measured in points per 100 possessions and split into offence and defence. The Lab builds it from play-by-play line-up data, from 2000-01 onwards.

  3. Lab Offence and Lab Defence: LAB IMPACT for every era

    Line-up data doesn't exist before 2000-01, so we trained two box-score models to predict LAB IMPACT from things every era recorded: shooting efficiency, usage, assists, turnovers, rebounds, steals and blocks. We tested them on seasons they had never seen (trained on 2001–15, tested on 2016–26), and they beat a hand-weighted formula on both ends. Now they score any season back to 1980.

  4. Offence and defence count the same

    Every prime season gets an OFFENCE score and a DEFENCE score against the starters at that position that year, and they're weighted 50/50. Playoff runs count double.

  5. The Lab Score: 70–30 is an MVP against an average starter

    We compare the two primes and put the gap on one fixed ruler. An MVP-level prime against an average starter = 70–30. The ruler is set by every MVP season since 1984, so a 60 means something.

    The ruler
    3050 · even70 · MVP

    Grey = 50 to 52, a photo finish. Orange = a clear win, up to the MVP mark: an MVP prime against an average starter.

  6. Every result names a winner

    Under 52 is a photo finish, and we still name the winner. We also show how close it is: how far the score moves if our judgement calls went the other way. The only tie is a dead-level 50.0–50.0.

    Player APlayer BPhoto finish

  7. The Battle Map tells the story

    Ten battles, from scoring to rim protection, show where one player wins. They explain the score; they don't set it.

  8. Receipts, and what the Lab can't see

    Every number links to its source. We also publish what the model gets wrong:

    • Rim deterrence: shots that were never taken don't show up in a box score. When a great shot-blocker loses, we say so on air.
    • No line-up data before 2000-01: older eras run on Lab Offence and Lab Defence only.
    • Misses: 3 of 58 test pairs went against the result we expected before running them. They're logged in our Failure Ledger.

  9. Your call

    We do the work. You make the call. Bring evidence we missed and we'll look again.

Explainer 2

How a Lab Score is made

The ruler, the 50/50 split and the photo finish. 47 seconds. Captions on screen.

Explainer 3

What the Lab can't see

Every blind spot, the Failure Ledger, and why the loser always gets his best case. 37 seconds. Captions on screen.

How we built it →

About our host

Our host is AI

Our on-screen host, J Swish, is a synthetic, AI-generated presenter, and we label him that way on every video. He presents the Experiment. He doesn't decide the result: the result comes from the evidence above.