This site talks about cosine, axis drift, era-z, archetypes and purity as if you already agreed to them. Each entry below gives the plain-English meaning first, the exact definition second, and then names the committed file it comes from — so you can check it rather than trust it. Terms with no file behind them are listed too, at the bottom, and marked.
A list of numbers standing in for a player-season, arranged so that players who played alike end up near each other.
A 64-dimensional unit-length vector produced by the MTNN from 130 input features. Because every vector has length 1, the dot product between two of them is their cosine.
assets/mtnn_arch.jsonHow close two players are, on a scale where 1.00 means identical and 0 means unrelated. It is the number behind every comparison on this site.
The cosine of the angle between two embedding vectors. On unit-length vectors this is the dot product. It ignores magnitude and measures direction only — which is the point: two players can be similar in shape without being equally productive.
computed in the browserStats adjusted so a 1997 season and a 2025 season can be compared without the pace of the league doing the talking.
Every rate is computed per 100 possessions, then z-scored within its own season — so a value says how far from that season's average a player was, not how the eras differ in raw volume. Clipped at ±4 standard deviations.
assets/season_norms.jsonCaveat the file states itself: three features are not invertible back to a raw rate, because they were empirical-Bayes shrunk toward the league mean by attempts before z-scoring. For those, a percentile is honest and a reconstructed percentage is not.
One of eight play-style groups the model sorts every player-season into. They are named after what separates them, not after positions.
Clusters over the embedding, frozen as a 14-dimension game contract so the game scores against a stable target. The eight, read live from the model file:
A slice of the network that only looks at one family of stats — shooting, or defense, or durability — before anything gets combined.
The MTNN reads its 130 features in family groups, each through its own
small sub-network, and fuses the results into the single 64-d embedding. A tower reads
cat([x*m, m]) — the values and a mask saying which were measured.
How much the whole map of the league twisted between one season and the next. A big angle means the game re-sorted who counts as similar to whom.
Orthogonal Procrustes alignment between consecutive seasons, computed on the players who appear in both — so roster turnover cannot manufacture a rotation. Reported as the mean principal angle of the rotation matrix against identity.
assets/drift.jsonHow much one particular stat stopped meaning what it used to mean.
1 − |Qii|, the extent to which that feature's own
axis had to move during the alignment. High drift on three-point accuracy in 2019-20 is
the league re-sorting itself around shooting, not shooters getting better overnight.
What the rotation could not explain — the part of the change that is not a twist of the map but a genuine reshuffle of players.
Normalized Frobenius norm of the difference remaining after alignment. No scaling is applied, because the spaces are already z-normalized.
assets/drift.jsonThe player from a different decade who played most like this one.
Nearest career in another decade by cosine between signature seasons, restricted to careers with at least four charted seasons.
assets/eratwins.jsonCaveat: that file was built in the older 48-dimensional embedding, as its own method line says. The model the game ships now is 64-dimensional. The twins are a snapshot of an earlier space.
The test: hand the model one season of a player's career and ask it to find that same player's next season out of ~13,000 candidates. If the model understands play style, the same person a year later should be the nearest thing to them.
Hit if season N+1 ranks in the top k, self excluded, ties counted against the target. Measured over … eligible pairs.
Loading measured scores…
assets/eval_scoreboard.jsonThe only scores that really count — seasons the model never saw while learning.
Target season ≤2021 is train, 2022-23 is validation, ≥2024 is test. Train-split pairs were training positives, so their score is inflated by construction and should never be quoted as accuracy.
assets/eval_scoreboard.json → protocol.splitsThe dumb version, kept around on purpose. If the neural net cannot beat plain era-adjusted stats, it has not earned its place.
Cosine over the frozen 14-dimension era-z game profile, evaluated on the identical pairs and protocol.
Loading…
assets/eval_scoreboard.jsonWhich input stats moved a given prediction, and in which direction.
aj = xj · ∂(target)/∂xj over
the raw input features. A local linearization — it says how the output responds
to a nudge, not what would happen if a feature were removed. It is not SHAP.
This entry used to add "and is not labelled as such anywhere on this site", which was wrong — and wrong in the direction that matters, because it told you not to look for something that is there. SHAP is on this site, for a different model: /methods reports mean|SHAP| for the draft model zoo — global 1245.3, log_overall 398.7, round 187.2, cap_growth 6.3 — and those figures are in assets/data/model_zoo_eval.json under glass_box.shap_global_mean_abs, computed as coefficient×(x−mean) for the linear models and an ablation approximation for the DeepMLP. Two different models, two different methods. The bars on the model page are the MTNN's local linearization; the SHAP figures on /methods belong to the draft zoo. Neither is the other.
assets/mtnn_attr_pop.json → method · assets/data/model_zoo_eval.json → glass_boxSome stats were not tracked in older seasons. Where a stat is missing, the model is told it is missing — so it contributes nothing, which is not the same as mattering nothing.
A tower reads cat([x*m, m]), so a masked feature has exactly
zero gradient. A zero attribution means NEVER MEASURED, not "no effect." Read it
against the coverage figure, which is the share of seasons where the feature exists.
A blend of real player-seasons that the puzzle asks you to identify.
A composite point in the embedding built from real player-season vectors, scored by cosine against your guesses.
assets/vectors.jsonThe one season used to stand for a whole career when comparing players.
The season selected as representative when matching careers across decades.
assets/eratwins.jsonWhy everyone gets the same puzzle today, and why refreshing will not give you an easier one.
A linear congruential generator seeded from the UTC date:
(seed × 1103515245 + 12345) & 0x7fffffff. Same date in, same puzzle out,
for every visitor — and a shared ?pack= link reproduces an identical set.
Kept honest on purpose. These appear in copy somewhere on the site but are not produced by any committed asset, so they cannot be checked. Treat them as claims.
Usually: of a player's k nearest neighbours, the share sharing their archetype label.
A specific value, purity@10 0.7057, sat on the map on
play.html as “pulp 0.7057” until 2026-08-10, when it was removed for the
reason this entry gives.
It is not in assets/eval_scoreboard.json, which is the committed
evidence file and reports adjacent-season retrieval instead. It may come from an older
48-d evaluation. Until a scoreboard ships it, it is a claim.
How many times better than a baseline something is.
A specific value, lift 6.32, has appeared in site copy with no
stated denominator. The ratio that is computable from the scoreboard —
model top-5 over the transparent 14-d baseline on the held-out test split — does not come
to 6.32. Stated as a claim until its definition ships with it.