A high-throughput, multi-league sports research engine with exact replay, self-tuning analytical ensembles, drift-aware online calibration, and content-addressed champion governance.
Research release: mechanics, integrity, autonomy, and performance are verified. Real-world predictive supremacy is not claimed without licensed point-in-time data.
Universal Sports Engine treats every simulation and model improvement as evidence, not theater.
- Exact by construction — immutable event streams, deterministic random namespaces, SHA-256 event chains, and complete result-envelope verification.
- Fast at two fidelities — analytical F0 distributions for massive Monte Carlo and event-level F1 paths for interpretable game flow.
- Autonomously adaptive — bounded population search, multiplicative expert weighting, self-calibrating intervals, champion/challenger gates, and prospective shadow updates.
- Fail-closed — drift alerts quarantine adaptive simulation; failed challengers remain in the ledger; existing checkpoints are never silently overwritten.
- Portable — dependency-free runtime, clean Python SDK and CLI, deterministic JSON replay, and a domain-neutral stochastic-control boundary.
| Dimension | Version 0.3 evidence |
|---|---|
| Supported leagues | 7 |
| Tests | 67 passed |
| Negative and mechanism controls | 49 passed, 0 failed |
| Randomized stress | 14,000 games / 2,003,304 events / 0 failures |
| F0 throughput | 111K–127K games/second |
| F1 single-worker throughput | 347–1,011 games/second |
| F1 four-worker scaling | 1.42×–1.82× |
| Recursive autonomy arena | 4 promotions / 3 guarded retentions |
| Calibration and shadow weight changes | 7 / 4 |
| Shadow drift alerts | 0 |
Machine-readable results live in release evidence, performance evidence, stress evidence, and the autonomy audit.
flowchart TD
CLI["CLI / Python SDK"] --> GOV["Champion governance"]
GOV --> CAL["Recursive calibration + online weights"]
CAL --> F0["F0 analytical distributions"]
GOV --> F1["F1 event simulation"]
F0 --> PACKS["Seven isolated league packs"]
F1 --> PACKS
PACKS --> CORE["Deterministic event / rule / RNG kernel"]
CORE --> PROOF["Replay + hashes + evidence bundles"]
PROOF --> RED["Stress, controls, drift, rollback"]
The neutral kernel contains no player, inning, possession, ball, puck, or scoreboard type. Sport semantics remain inside league packs; model governance remains outside simulation truth.
| Family | League packs | Primary F1 unit |
|---|---|---|
| Bat and ball | MLB | Pitch / plate appearance / inning |
| Court invasion | NBA, WNBA, NCAA MBB | Possession / period / game |
| Gridiron | NFL, NCAA Football | Play / drive / game |
| Ice invasion | NHL | Event / shift / period |
Each pack owns its rules, scoring, clocks, stochastic transitions, and provisional reference anchors. One sport is never treated as a rescaled copy of another.
flowchart LR
T["Point-in-time train"] --> S["Bounded population search"]
S --> C["Chronological calibration"]
C --> W["Expert reweighting + interval adaptation"]
W --> G{"All gates pass?"}
G -- Yes --> P["Promote + freeze checkpoint"]
G -- No --> R["Retain incumbent + record failure"]
P --> H["Prospective shadow outcomes"]
R --> H
H --> D{"Drift?"}
D -- Healthy --> W
D -- Alert --> Q["Quarantine / recalibrate / rollback"]
Profiles, ensembles, adaptive states, champion bundles, and promotion ledgers all carry independent content hashes. Promoted bundles embed the complete incumbent model and state, so rollback is self-contained rather than merely documented.
Read the full autonomous improvement protocol.
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[dev]"
.\.venv\Scripts\sports-engine.exe list-leagues
.\.venv\Scripts\sports-engine.exe simulate --league nba --seed 42
.\.venv\Scripts\sports-engine.exe simulate-fast `
--league nfl --seed 42 --home-strength 1.0 --away-strength 1.0.\.venv\Scripts\sports-engine.exe auto-improve `
--csv data\point-in-time\historical.csv `
--config AUTONOMY_CONFIG.json `
--output reports\improvement-run-001
.\.venv\Scripts\sports-engine.exe verify-autonomy `
--bundle reports\improvement-run-001\champion_bundle.json `
--ledger reports\improvement-run-001\improvement_ledger.jsonl.\.venv\Scripts\sports-engine.exe simulate-adaptive `
--bundle reports\improvement-run-001\champion_bundle.json `
--league nba --seed 42 --home-strength 1.0 --away-strength 1.0| Document | Purpose |
|---|---|
| Architecture | Kernel, fidelity, and governance boundaries |
| Autonomous improvement | Search, weighting, drift, promotion, and rollback |
| Quality gates | Machine-readable non-negotiable release rules |
| Release evidence | Current validation and reproducibility snapshot |
| Upgrade audit | Version 0.3 implementation and results |
| Known limitations | Exact boundaries of demonstrated capability |
| Prior art | Primary references and defensible distinctions |
The bundled autonomy arena is deliberately synthetic. It proves that the engine can discover, reject, promote, reweight, checkpoint, reload, monitor, and roll back models; it does not turn synthetic outcomes into a claim of real-world accuracy.
Licensed point-in-time game, roster, injury, event, and tracking data are required before held-out predictive certification. F2 spatial agents and F3 physics remain blocked rather than fabricated.
Build measurable worlds. Preserve every decision. Believe only what survives.