Launch autonomous universal sports engine

.gitattributes

@@ -0,0 +1,5 @@
+* text=auto eol=lf
+
+*.bat text eol=crlf
+*.cmd text eol=crlf
+*.ps1 text eol=crlf

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,32 @@
+## What this delivers
+
+Launches Universal Sports Engine 0.3.0: a deterministic, high-throughput,
+multi-league simulation platform with exact replay, bounded recursive calibration,
+online expert weighting, drift-aware shadow monitoring, and content-addressed
+champion governance.
+
+## Why it matters
+
+- Unifies MLB, NBA, WNBA, NCAA MBB, NFL, NCAA Football, and NHL behind a neutral kernel.
+- Preserves sport-specific rules and mechanics in isolated league packs.
+- Separates fast analytical distributions from interpretable event simulation.
+- Makes every autonomous promotion, rejection, weight update, and rollback auditable.
+- Fails closed when integrity, calibration, drift, or reproducibility gates fail.
+
+## Verified evidence
+
+- [ ] Ruff passes.
+- [ ] Strict MyPy passes.
+- [ ] All 67 integration and invariant tests pass.
+- [ ] Release evidence gates pass across all seven leagues.
+- [ ] The 49-control negative and mechanism suite reports zero failures.
+- [ ] The 14,000-game stress audit reports zero failures.
+- [ ] The autonomy audit verifies champion, ledger, shadow-update, and rollback integrity.
+- [ ] A clean wheel rebuild reproduces the expected package contents.
+
+## Evidence boundary
+
+This release verifies mechanics, determinism, replay integrity, autonomy governance,
+and measured local throughput. It does not claim real-world predictive supremacy.
+That requires licensed point-in-time game, roster, injury, event, and tracking data
+plus held-out and prospective validation.

.github/workflows/quality.yml

@@ -0,0 +1,54 @@
+name: Quality
+
+on:
+  pull_request:
+  push:
+    branches:
+      - main
+
+permissions:
+  contents: read
+
+concurrency:
+  group: quality-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  verify:
+    name: Python 3.12 verification
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+
+    steps:
+      - name: Check out source
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: pip
+
+      - name: Install project
+        run: python -m pip install --upgrade pip && python -m pip install -e ".[dev]"
+
+      - name: Lint
+        run: python -m ruff check .
+
+      - name: Type-check
+        run: python -m mypy src scripts
+
+      - name: Test
+        run: python -m pytest -q
+
+      - name: Build wheel
+        run: python -m pip wheel . --no-deps --wheel-dir dist
+
+      - name: Verify release gates
+        run: >-
+          python scripts/build_release_evidence.py
+          --games 500
+          --benchmark-games 100
+          --seed 20260718
+          --workers 2
+          --output "${{ runner.temp }}/release-evidence"

.gitignore

@@ -0,0 +1,18 @@
+.venv/
+.repro-venv/
+.repro-v3-venv/
+__pycache__/
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+*.egg-info/
+build/
+dist/
+reports/archive/
+reports/COPY_PROOF.json
+reports/HANDOFF_PROOF.json
+reports/COMPREHENSIVE_AUDIT_V2.md
+reports/PERFORMANCE_AUDIT_V2.json
+reports/PERFORMANCE_AUDIT_V3.json
+reports/RECOVERY_AND_COMPLETION_REPORT.md
+reports/STRESS_AUDIT_V2.json

AGENTS.md

@@ -0,0 +1,10 @@
+# Universal Sports Engine Rules
+
+- Keep the simulation truth independent from rendering.
+- Preserve deterministic replay: identical versions, inputs, and seed must produce identical events.
+- Keep sport semantics in league packs; the core API must remain domain-neutral.
+- Fail explicitly on invalid state, unknown league IDs, corrupted event chains, and invalid rules.
+- Do not claim real-world calibration from synthetic fixtures or reference anchors.
+- Do not add wagering, sportsbook, brokerage, or live-capital behavior.
+- Keep changes typed, functional, minimal, and covered by integration or replay tests.
+- Never weaken a quality gate because a candidate or league pack fails it.

ARCHITECTURE.md

@@ -0,0 +1,64 @@
+# Architecture
+
+## Decision
+
+The kernel uses immutable event drafts and a content-addressed event log as the
+source of truth, pure reducers for replay, fixed phase ordering, and label-addressed random draws. This
+combines the strongest parts of the two architectures recovered from the stalled
+Kimi build:
+
+1. Event-sourced state plus hierarchical domain state.
+2. Explicit `PERCEIVE -> DECIDE -> ARBITRATE -> RESOLVE -> APPLY -> PUBLISH` phases.
+3. Independent random namespaces and common-random-number counterfactuals.
+4. Game-granularity parallelism; a single game timeline remains deterministic.
+5. A result-envelope hash that commits to league, final score, seed, event chain,
+   and evidence metadata.
+
+## Layers
+
+```text
+CLI / SDK / domain-neutral bridge
+          |
+Champion governance, recursive calibration, adaptive weights, drift quarantine
+          |
+F0/F1 routing, batch, historical evaluation, counterfactuals
+          |
+League packs and sport-family resolvers
+          |
+Domain-neutral event log, state, rules, RNG, replay
+```
+
+League packs own scoring, timing, possession/inning/drive/shift semantics, and
+reference anchors. The kernel contains no player, ball, puck, inning, possession,
+or scoreboard type.
+
+## Fidelity
+
+- F0: analytical/batch score distribution (implemented; no event path).
+- F1: event-level paths (implemented for every league).
+- F2: spatial-agent contracts and escalation interface (API complete; no tracking calibration).
+- F3: rendering/physics adapter boundary (not implemented without tracking/physics evidence).
+
+The fidelity router selects the lowest evidence-supported fidelity and rejects
+unsupported F2/F3 requests. It never uses visual
+plausibility as a substitute for held-out calibration.
+
+## Integrity and performance
+
+League packs accumulate immutable drafts locally and content-address them once in
+linear time. External replay import recomputes the event chain and result envelope.
+The RNG maps semantic labels directly to BLAKE2b uniform draws, so unrelated model
+changes do not perturb other random namespaces.
+
+## Recursive model governance
+
+The adaptive layer is outside the neutral kernel. Profiles, expert mixtures,
+online calibration state, champion bundles, and decision ledgers are immutable
+and content-addressed. Search may mutate only bounded analytical parameters.
+Promotion is decided on chronological calibration forecasts made before each
+outcome is revealed. Test data, rule packs, evaluators, and quality gates are not
+mutable search inputs.
+
+Online shadow adaptation changes expert weights and interval scale, while two
+loss EWMAs monitor drift. Drifted checkpoints remain auditable but are rejected
+by adaptive simulation until a healthy successor or rollback is selected.

AUTONOMOUS_IMPROVEMENT.md

@@ -0,0 +1,80 @@
+# Autonomous recursive improvement
+
+Version 0.3 adds a bounded model-improvement system around the deterministic
+simulation kernel. It changes analytical model parameters and ensemble weights;
+it cannot edit source code, official rule packs, historical outcomes, split
+boundaries, evaluators, promotion gates, or live-action interfaces.
+
+## Evidence loop
+
+```text
+point-in-time train block
+        -> deterministic population search
+        -> diverse bounded expert pool
+        -> chronological calibration predictions
+        -> outcome revealed once
+        -> expert weights and interval scale updated
+        -> challenger compared with incumbent
+        -> promote or retain
+        -> content-addressed checkpoint and chained ledger
+        -> shadow outcomes update weights and drift state
+        -> healthy, quarantine, recalibrate, or roll back
+```
+
+Training outcomes select bounded score-model parameters. Calibration outcomes
+are consumed prequentially: the ensemble predicts first, then receives the
+outcome exactly once. Test outcomes are excluded from tuning and promotion.
+Shadow outcomes may update an already selected champion but cannot rewrite its
+lineage or erase prior decisions.
+
+## Adjustable state
+
+Each expert contains league-specific home and away score locations, total and
+margin dispersion, and strength-response exponents. Recursive search uses
+deterministic log-space mutations, retains a diverse elite set, and reduces the
+mutation radius by generation.
+
+The selected ensemble applies multiplicative expert weighting using joint score
+likelihood and home-win Brier loss. A separate online calibration state adjusts
+the 90% interval radius. Short- and long-horizon loss estimates produce a drift
+score. A drift alert blocks adaptive simulation until a new governed checkpoint
+or explicit rollback is selected.
+
+## Promotion gates
+
+A challenger must satisfy every preregistered requirement in
+`AUTONOMY_CONFIG.json`:
+
+- minimum composite-loss improvement;
+- no excessive Brier deterioration;
+- no excessive interval-coverage deterioration;
+- finite scores and normalized non-negative weights;
+- healthy calibration state;
+- sufficient non-overlapping train and calibration evidence;
+- valid profile, ensemble, adaptive-state, bundle, and ledger hashes.
+
+Synthetic fixtures can validate learning mechanics but can never certify
+real-world accuracy. Real promotion claims require licensed point-in-time data
+and an untouched test audit after the champion is frozen.
+
+## Commands
+
+```powershell
+.\.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 simulate-adaptive `
+  --bundle reports\improvement-run-001\champion_bundle.json `
+  --league nba --seed 42 --home-strength 1.0 --away-strength 1.0
+
+.\.venv\Scripts\sports-engine.exe adapt-online `
+  --csv data\point-in-time\shadow.csv `
+  --config AUTONOMY_CONFIG.json `
+  --bundle reports\improvement-run-001\champion_bundle.json `
+  --league nba --output reports\nba-shadow-checkpoint-001.json
+```
+
+Output paths are fail-closed: existing improvement artifacts and adaptive
+checkpoints are not overwritten.

AUTONOMY_CONFIG.json

@@ -0,0 +1,22 @@
+{
+  "brier_weight": 2.0,
+  "coverage_weight": 2.0,
+  "drift_threshold": 0.12,
+  "elite_count": 6,
+  "experts_per_ensemble": 6,
+  "generations": 8,
+  "interval_learning_rate": 0.04,
+  "long_decay": 0.03,
+  "maximum_brier_deterioration": 0.01,
+  "maximum_coverage_error_deterioration": 0.03,
+  "minimum_calibration_games": 50,
+  "minimum_drift_observations": 40,
+  "minimum_relative_improvement": 0.01,
+  "minimum_train_games": 100,
+  "mutation_scale": 0.12,
+  "online_learning_rate": 0.25,
+  "population_size": 32,
+  "seed": 20260718,
+  "short_decay": 0.20,
+  "target_coverage": 0.90
+}

DATA_RIGHTS_AND_PROVENANCE.md

@@ -0,0 +1,22 @@
+# Data Rights and Provenance
+
+## Local build inputs
+
+No licensed event feed, tracking feed, private injury feed, sportsbook feed, or
+proprietary roster database is present in this repository.
+
+The shipped engine uses:
+
+- deterministic synthetic events generated by the versioned league packs;
+- transparent league-level reference anchors for broad output sanity checks;
+- official rule-source links as provenance pointers for the minimal rule manifests.
+
+The reference anchors are approximate research fixtures. They are not an ingested,
+point-in-time dataset and are labeled `provisional_reference_anchor` everywhere.
+
+## Prohibited promotion
+
+These inputs cannot support a claim of held-out player, team, game, season, spatial,
+or counterfactual accuracy. Such claims require separately licensed data manifests,
+earliest-available timestamps, revisions, train/calibration/test boundaries, and
+independent reproduction.

DOMAIN_NEUTRAL_ADAPTATION_READINESS.md

@@ -0,0 +1,12 @@
+# Domain-neutral adaptation readiness
+
+No external project was inspected, imported, or modified during this build.
+
+The reusable boundary consists only of immutable state, action, observation,
+scenario, counterfactual, uncertainty, fidelity, and evidence contracts. A future
+consumer must implement its own domain pack and validation program. Sports rule
+packs, reference anchors, league identities, and generated scores must not cross
+that boundary.
+
+Compatibility is demonstrated by the included non-sports inventory-control fixture.
+There is no live-action, capital, brokerage, sportsbook, or wagering adapter.

DOMAIN_NEUTRAL_KERNEL_SPEC.md

@@ -0,0 +1,22 @@
+# Domain-neutral kernel specification
+
+The kernel owns only immutable events, world-state payloads, phase ordering,
+versioned rules, deterministic random draws, replay verification, fidelity routing,
+and evidence metadata. It contains no sports entities or timing assumptions.
+
+## Invariants
+
+- Sequence numbers are contiguous and zero-based.
+- Every event hash commits to its body and prior event hash.
+- Identical root seed and semantic labels produce the same draw.
+- An unavailable fidelity raises instead of silently returning a lower-fidelity answer.
+- A league pack must verify its event stream before publishing a result.
+- Portable replay import verifies the complete chain before returning a result.
+
+## Extension points
+
+- League pack: complete-game resolver and rule manifest.
+- Learned transition: replaces a league transition function without changing evidence contracts.
+- Spatial adapter: produces `SpatialState` values only after league-specific validation.
+- Physics adapter: remains independent of rendering and uses the same event sink.
+- Policy adapter: consumes observations and permitted actions, never omniscient state by default.

KNOWN_LIMITATIONS.md

@@ -0,0 +1,22 @@
+# Known Limitations
+
+- The stalled remote Kimi repository was not exported; this is an independent local reconstruction.
+- F1 event paths are calibrated only to broad reference anchors, not licensed held-out data.
+- Player identities, rosters, coaching policies, officials, injuries, fatigue, venues, and weather are contract surfaces rather than calibrated models.
+- F2 spatial trajectories and F3 physics/rendering are intentionally not fabricated without tracking evidence.
+- Rule manifests cover complete-game flow but are not a complete executable transcription of every official rulebook edge case.
+- Generic F1 games resolve a winner; competition-specific regular-season tie and
+  postseason context is not yet a first-class scenario input.
+- The analytical F0 model matches provisional aggregate moments but does not emit
+  sport-valid event sequences or prove game-level predictive skill.
+- Historical evaluation is implemented, but no licensed point-in-time test manifest
+  was supplied, so CRPS, Brier, and coverage results on real games remain unavailable.
+- Recursive tuning currently adjusts F0 analytical score distributions and ensemble
+  weights. It does not fabricate player, tracking, tactical, F2, or F3 calibration.
+- The bundled autonomy audit uses labeled synthetic fixtures solely to prove search,
+  weighting, gating, checkpoint, and replay mechanics. Its promotions are not
+  real-world champions.
+- Windows multi-process calls require an importable script or CLI entrypoint; an
+  interactive/stdin caller receives an explicit error and should use F0 or one worker.
+- Counterfactual pairing reduces simulation noise but does not by itself establish causality.
+- No live feed, wagering, sportsbook, trading, or capital interface exists.

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 ObtuseAI
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.