Player — Stay On Floor

Map — player-seasons placed by the 64-d embedding, coloured by cluster

MAP — 64-D POSITION, COLOURED BY CLUSTER PWA OFFLINE-FIRST

◎ hover a dot to name it · click to search its twins · drag to rotate

Offline: the map file is not in the service worker's cache — sw.js bypasses .json, .f32 and .bin on purpose, so a stale model asset can never be served. What keeps this drawing without a network is the browser's own HTTP cache, which holds /assets/* as immutable. This line used to say the worker cached it, which the worker's own DATA pattern has always contradicted.

Tap Jordan → position only, not auto-start — deliberate. That's respect, not autoplay.

CLOSER vs EXPLOITABLE ARCHETYPE YELLOW SIGNAL

CLOSER FIT

CLOSER FIT

🟢

Stays on the floor late. The site carries no closer score, so no number is quoted for it.

#E7F6EA GREEN

EXPLOITABLE

EXPLOITABLE

🔴

Targeted late. Same — a colour, not a measured score.

#FFD8D5 RED

DEPENDENT

DEPENDENT

🟡

Needs someone else to create. Same — a colour, not a measured score.

Viz: a sample of the 12,966 player-seasons, placed by the 64-d embedding. Measured on the v5 baseline: archetype purity@20 0.67, adjacent-season top-1 retrieval 0.51.

—

STAY-ON-FLOOR PIDPLAYOFF MINUTE SECURITYPLASTICITY → REAL MINS

Team payroll and wins, from front_office_lite.json

TEAM PAYROLL AND WINS MEASURED

WINS PER $M OF PAYROLL

—

the season’s wins divided by payroll in $M

PAYROLL

—

A fit score and a contract forecast stood here. Both were invented — the fit multiplied a Math.random() term and printed three decimals, and the forecast scaled a league median typed into the source. No contract model is committed to this repo, so these are the figures that are: payroll, wins and the front-office grade, per team.

Where you actually stay on — scarce archetype × cap clarity, not vibes.

WHAT IS HERE, AND WHAT IT RESTS ON GLASS BOX

FREE

FREE

The map, the nearest-neighbour search and the team figures are all free, with no account.

SHAP + CLOSER FIT

SHAP + CLOSER FIT

Per-player attribution is on the explainability page, computed from the committed tensors.

API + WHITE-LABEL

API + WHITE-LABEL

API and white-label. The model is 64-d over 12,966 player-seasons.

Loading the prop summary…

The map holds no minutes played, so no minutes are quoted here.