Vector Hoops

Daily NBA chimera puzzle over an era-honest player embedding space. 12,392 player-seasons (1996-2026), per-100-possession, z-scored within season; PCA(3) map; 8 auto-named archetypes. Static, zero backend, free.

Picking up in-progress work? Start at docs/HANDOFF.md — current branch state, the four dormant data tracks and how to activate them, verification commands, and open follow-ups.

Skills Lens (skills.html, docs/SKILLS_LENS.md)

Every charted player-season graded 0-99 on twelve skills — fixed linear composites of the era-z contract, percentile within season pool, badges at 90+. pipeline/build_skills.py emits assets/skills.json (grades, order-aligned with vectors.json), assets/skill_probe.json (weights + pooled quantile knots so the client grades the fused daily chimera), and MTNN training targets. pipeline/test_skills.py gates every rebuild; pipeline/update_dataset.py is the growth loop (fetch best-effort → rebuild → gate → ledger at pipeline/cache/dataset_ledger.json). train_mtnn.py (v4) adds a per-skill tower bank on the fused embedding with held-out per-skill R² in mtnn_report.json — research lane only; the site ships the transparent composites.

Two dormant data tracks extend the MTNN with distinct tower families, each cache-ready and gated on a committed fixture until one operator fetch on a non-datacenter IP (stats.nba.com blocks this environment):

• Pedigree (Track H) — draft slot, entry expectations, team-fit prior + pedigree_expectation head. Activate: bash pipeline/operator_fetch_pedigree.sh.
• Playoffs (Track I) — postseason as a distinct regime: playoff-vs-regular-season deltas (minutes, usage, scoring, efficiency)
  • team wins/rounds + playoff_riser head, plus a transparent Playoff Lens (RS vs PO splits + riser/fader) on skills.html. Activate: bash pipeline/operator_fetch_playoffs.sh.
• Wide skills (Track J) — post / transition / motor as masked skills (2015-16+) from synergy + hustle feeds; Skills Lens bars + MTNN skill-tower targets (per-skill mask matrix). Activate: bash pipeline/operator_fetch_wide_skills.sh.

The Skills Lens also carries a Steals of the Draft board (draft expectation vs actual peak skill grade) that lights up with the Track H draft data.

OKF LLM-Wiki (knowledge/)

One interlinked, machine-editable markdown page per charted player (2,293), plus archetype and position hubs. Two-layer contract: an AUTO block regenerated from the data, and a CURATED layer below the marker that humans/LLM agents extend and the generator never touches. Nearest- neighbor wikilinks make the graph walkable; the game's reveal card links each chimera component to its dossier. Contract: knowledge/OKF.md.

• python pipeline/build_wiki.py — idempotent regeneration

Enrichment (pipeline/enrich_vectors.py)

Adds to assets/vectors.json without touching the game contract:

• proj — exact affine recovery of the build-time PCA+minmax map, so the client projects the fused Chimera vector into the 3D map honestly
• axes — PC1/PC2/PC3 interpretations verified against feature correlations (paint vs perimeter / scoring load / ball in hand)
• p + positions — per-season PG/SG/SF/PF/C from Basketball-Reference (pipeline/fetch_positions.py, 99.7% coverage), coloring the 3D map

Data pipeline (v2)

python pipeline/build_vectors.py builds two artifacts:

• assets/vectors.json — the frozen 14-dim game contract (+ optional sal salary-z per player where payroll coverage exists)
• pipeline/data/train_matrix.npz + feature_manifest.json — the wide matrix: Base + Advanced + shot-mix (Scoring) + bio + player-tracking (2013-14+, masked before) + salary, all era z-scored with missing masks, grouped into tower families

Sources: stats.nba.com (leaguedashplayerstats Base/Advanced/Scoring, leaguedashplayerbiostats, leaguedashptstats), basketball-reference current contracts, and an optional pipeline/cache/salaries_history.csv drop-in (name,season,salary) for full 1996+ payroll history.

Every season/endpoint response is cached under pipeline/cache/; stats.nba.com throttles hard, so re-running resumes where it left off. --offline rebuilds from cache only.

Cleaning: dedupe on (PLAYER_ID, season), NaN -> season mean with masks, attempt-weighted empirical-Bayes shrinkage of FG3%/FT%/FG%, z-clip ±4.

Embedding v2 (multi-tower net)

python pipeline/train_towers.py (torch) trains per-family MLP towers (volume / playmaking / rebounding / defense / efficiency / shot-mix / tracking / bio / market) fused into a 32-dim contrastive embedding (InfoNCE; positives = same player in adjacent seasons + augmented views; auxiliary salary-regression head). Outputs pipeline/data/embedding_v2.npz + tower_report.json with a same-player-next-season recall@10 sanity metric. The game keeps the transparent 14-dim profile until v2 demonstrably beats it — promotion is a deliberate, separate step.