CourtVision
Calibrated multi-sport forecasting, audited in public -- with the rejects and retractions left in.
What it is. A four-sport forecasting system (NBA, MLB, soccer, tennis) with a leak-guarded walk-forward
harness and a paper-only execution layer. Pregame probabilities match or trail the devigged close; the in-game
model is behind the live market; honest rejects and six retractions are published. Nothing here is a
betting-edge claim. The engine is private; this repository is the evidence surface.
Start here
Claim-by-claim reproduce index: EVIDENCE.md.
Who it is for. Quant research, pricing and sports-analytics teams, and ML or data hiring managers who want
to see a forecasting system measured against the market with its negative results kept in.
Reading this as a trading desk or quant team? Start with the one-page
desk brief: pricing quality against the devigged close and the live market,
what exists and what it is evidence of, and what is not shown.
Results at a glance
The central result is deliberately unglamorous: against real closing lines the market is efficient.
Every row links to its proof in EVIDENCE.md (claim -> number -> verdict ->
artifact -> reproduce command).
| Question | What the record says | Verdict |
|---|
| How does the pregame model compare with the devigged close? | NBA held-out Brier 0.1735 vs 0.1666 for the devigged close (n=372 scored, CI includes 0, underpowered); MLB moneyline and soccer O/U trail by 0.0039 and 0.0076 Brier with intervals excluding 0; totals and ATP trail by more (recomputed 2026-09-24) | TRAILS_CLOSE on MLB moneyline and soccer O/U; MATCHES_CLOSE (underpowered) on NBA; no row scores better than the close |
| Does in-game state help? | vs a static pregame prior: NBA Brier 0.209 to 0.159, MLB 0.241 to 0.126 | sharper than static |
| How does it compare with the live market? | NBA end-Q1: market sharper (n=1,592, as of 2026-07-22). Later checkpoints: underpowered. Kalshi paired ticks (2026-09-14): model behind | BEHIND on the Kalshi ticks; MARKET_SHARPER_PROVISIONAL at NBA end-Q1 -- and the page says so |
| How many discovered signals shipped? | 0 of 60 candidate classes; 513 recorded reject / defer verdicts | the gate rejects, by design |
| What got retracted? | six headline figures, each root-caused to a leak or a grading artifact by the project's own harnesses | the retraction record |
| Player-prop accuracy | PTS MAE 4.83, REB 1.92, AST 1.39 on a 20,354-row chronological holdout (re-measured 2026-07-20) | accuracy only |
Brier: mean squared error of a probability forecast, lower is better. Devigged close: the closing market price with
the bookmaker margin removed.
The strongest signal in this repository is not a metric. It is that the same person who built
the system also built the instruments that caught his own overclaims, and published the result.
How far it goes
One person directing an agentic build pipeline. This public repository holds 6,556 commits from 2026-03-09 to
commit 4f033b5a5 (git rev-list --count 4f033b5a5, counted 2026-10-21; 1,620 of them, 24.7%, carry a Claude or Codex Co-Authored-By trailer), most of them agent-authored under direction; the audited count in the
private repository was 5,504 at 2026-09-07 (packet section 4) and has grown since. Counts below are read
from committed artifacts or the audited evidence packet; the layer-by-layer
map is docs/CAPABILITIES.md.
| Layer | Scale |
|---|
| Sports covered by one kernel + per-sport adapters | NBA, MLB, soccer, tennis |
| Analytics site (public, static, CI-gated) | 29 research papers, 77 measurement modules, 16 findings (as of 2026-10-21), 10 interactive inspectors, 1,549 entity cards, 1,653 searchable records |
| Intelligence layer between raw data and models | 151 files; a 291,625-pair player-vs-player matchup matrix; 1,249 player dossiers across 28 categories |
| Signal discovery under a gate built to refute | 60 candidate signal classes, 0 shipped; 513 recorded reject / defer verdicts; 197 hypothesis-ledger rows with every NULL kept (the ledgers are public) |
| Fact-claims corpus | 103,048 generated claims, 101,864 re-verified from their declared source and formula |
| Simulation | possession-level Monte Carlo where the negative teammate correlation emerges from a shared scoring pie instead of a hand-tuned matrix |
| Engine (private repository, not runnable here) | FastAPI serving layer, about 99 routes across 12 routers (API reference, capability map), 9 long-running daemons under a watchdog, a 430-module codebase, about 7,400 tests; read access on request |
| Computer vision (paused) | Broadcast video to court coordinates (YOLOv8n, SIFT homography, Kalman + Hungarian tracker, OSNet re-ID). Status: paused since 2026-09-17; the frozen audit accepted no game as coordinate-quality evidence, and tracked features carry no measured forecast lift. Code is in the private repository; the audit and design notes are TRACKING.md and CV_TRACKING.md |
| Execution | paper only, behind six pre-registered go-live gates; the order lifecycle has no reachable live path; a daily forward paper desk has run since 2026-10-06 (130 games served on in-game markets through 10-22, night receipts under , every scored cell still below its 30-game floor -- ) |
What is on the site
- Calibration -- every reliability bin with a game-cluster bootstrap interval.
- Papers -- research write-ups whose evidence paths are machine-checked against the published artifacts before deploy.
- Inspectors -- state reliability, residual anatomy, blowout timing, pitch sequencing, score decomposition, cross-sport comparability.
- Experimental metrics, the measurement lab, Compare, the entity library (1,549 cards), and Ask Scout -- every answer cites its artifact and
as_of date, or returns NO_DATA.
- The retraction finding -- what was claimed, what was wrong, how it was caught.
The site is a static export built from webapp/ on every push; published JSON is
scrubbed and receipt-checked in CI before it ships. Page-by-page map, data lineage and the CI gates:
docs/ANALYTICS_SITE.md.
Papers, model cards and the desk record
- Pregame vs the close -- the pregame Brier comparison against the devigged close, with the 2026-09-24 recompute.
- In-game vs the market -- Gate A-0, the cleaned MLB corpus, the six rejected Model 2.0 probe families and the staleness backstop: the model trails a fresh venue mid in every corpus.
- In-play capture latency -- capture-leg latency recomputed from 20 committed nightly ledgers (no venue latency, no pricing claim).
- Model cards -- one card per forecaster the pages cite: intended and not-intended use, inputs, validation, measured performance with as-of dates, failure modes and the retraction record;
scripts/platformkit/eval_gate/model_card_check.py fails on a missing card, field or hash.
- Desk uptime -- how many scheduled nights the paper desk actually scored, from tracked receipts only.
Verify it yourself (no private data, about 5 minutes)
git clone --depth 1 https://github.com/neeljshah/court-vision.git
cd court-vision
pip install "numpy>=1.24" "pandas>=2.0" "matplotlib>=3.8"
python scripts/platformkit/analytics_showcase/check_all.py
Each showcase module re-verifies its own committed artifact; 0 FAIL is the bar. Details and
scope limits: REPRODUCE.md. Generated receipts, one row per measurement:
RECEIPTS.md.
Audit trail of retractions and fixes: git log -i --grep="retract\|revert" --oneline
Five more cross-page checks (figure consistency, devig, fee schedule) are in the
desk brief, section 6; the methods are defined on CLV_AND_MARKOUT.md and KALSHI_NOTE.md.
How it is built
flowchart LR
D["DATA<br/>keyless leak-free ingest<br/>4 sports, as-of stamped"] --> S["SIGNALS<br/>ratings, per-player rates<br/>151-file intelligence layer"]
S --> M["MODELS<br/>one calibrated win prob per sport<br/>count distributions for props"]
M --> E["ENGINES<br/>possession Monte Carlo<br/>in-game repricer"]
E --> P["PREDICTIONS<br/>moneyline, spread, total, props<br/>pregame and in-game, coherent"]
P --> X["EXECUTION - paper only<br/>no reachable live order path"]
P --> V["EVIDENCE SURFACE<br/>analytics site, receipts<br/>EVIDENCE.md"]
L["AGENTIC LOOP<br/>propose -> walk-forward gate<br/>-> permutation control -> ship or reject"] -. re-validates every stage .-> S
L -.-> M
V -. retractions and rejects recorded .-> L
classDef pub fill:#0b3d2e,stroke:#2ea043,color:#ffffff;
classDef priv fill:#1f2937,stroke:#6b7280,color:#ffffff;
class V pub;
class D,S,M,E,P,X,L priv;
Green is what this repository holds; grey runs in the private engine. The full capability map, layer by
layer with scale numbers, status and a public place to verify each one, is
docs/CAPABILITIES.md.
One calibrated win probability per sport anchors the moneyline, spread, total and the in-game
reprice, so the markets are coherent reads off one engine rather than independent models that can
disagree. A sport-blind kernel/ holds the validated machinery; each sport is an
adapter. Candidate signals are proposed automatically and must survive walk-forward folds, a
null-shuffle permutation control, ablation against the full model, and a multiple-comparisons
correction before they ship. Most do not.
The code that enforces this is readable here: the walk-forward harness, leak guard, calibration
and deflated-metric modules in scripts/platformkit/eval_gate/,
and the conformance and golden-set kit in kernel/testing/.
Longer reads: ARCHITECTURE.md -
the full system tour -
how the honesty gates work -
how the agentic build pipeline works -
the MCP answer-engine interface
Repository map
This repository is the public evidence surface. The production engine, sport adapters, API,
operational tooling, bulk raw corpora, model artifacts and ledgers are kept in a private repository;
reviewers can request read access. What is public is chosen by one reviewed allowlist
(scripts/hooks/public_allowlist.txt) that the push guard enforces.
What this is not
- Not a betting-edge or ROI product. No dollar figure on this repository is a result.
- Not third-party reproduced.
check_all.py is self-serve artifacts evaluated -- functional;
the private corpora behind recorded rows are not in a fresh clone.
- Not copyleft-clean: the computer-vision lineage depends on Ultralytics YOLO (AGPL-3.0). The
repository's own code is proprietary and published for evaluation (LICENSE).
Related public repos
Single-concept repos distilled from this platform, each runnable on its own: shin-devig (four de-vig methods incl. Shin 1992; the Shin-devigged NBA close used in EVIDENCE.md section A is the 0.1666 figure), calibration-gate (Brier decomposition, ECE, recalibration, two-corpus acceptance), walkforward-guard (leak-proof time-series CV; the demo plants two leaks and catches both), prereg-cards (pre-registration workflow), sdv-calibration-audit (a calibration audit of a published sports-model probability column).
Contact
Built by Neel Shah, who has three years of data-science coursework at the University of Iowa, directing an
agentic build pipeline. Open to quant research, ML
engineering and founding-engineer roles, and to conversations about the system itself.
All numbers are calibration and sharpness measures (Brier, CRPS, MAE, ECE). Retracted figures
appear only in explicit retraction context in docs/JOB_EVIDENCE_PACKET.md
and docs/KNOWN_LIMITATIONS.md, never here.
