Loto Forecast Platform

ミニロト / ロト6 / ロト7 / ビンゴ5 / ナンバーズ3 / ナンバーズ4を対象に、統計モデル、機械学習、深層学習、AutoML、時系列基盤モデル（TSFM）、確率モデルを、時系列リークを防いだ共通契約で比較・検証・運用する研究プラットフォームです。

このREADMEは「何が実装されているか」「どの分母を見ているか」「どこまでruntime/scientific evidenceがあるか」を最短で把握する入口です。

Phase 4A candidate base: main@45bcf60fa04fc3736e3a73760039254573abf4c8 (after GluonTS PR #323, 2026-08-13)
Critical count rule: Broad v1 count != Committed Expanded v2 count != discovered/source count != runtime-certified count
Scientific rule: REGISTERED != ROUTABLE != RUNTIME_CERTIFIED != OOF_EVALUATED != HOLDOUT_EVALUATED != PROSPECTIVE_EVALUATED != PROMOTION_ELIGIBLE
package versionはREADMEへ手書きしません。loto.version.__version__ / installed metadata / loto-build-infoを正本とします。

まず見る資料

知りたいこと	資料
Broad=1と実モデル/Expanded数の違い	docs/INVENTORY_COUNT_BOUNDARIES.md
GluonTS Expanded v2 Phase 3	docs/gluonts/GLUONTS_EXPANDED_V2_PHASE3.md
ライブラリ別モデル・引数・対応機能	docs/LIBRARY_MODEL_COMPATIBILITY_MATRIX.md
current state / open gates	docs/STATUS.md
現在の検証境界	docs/CURRENT_VERIFICATION_REPORT.md
次に作業する人向け引継ぎ	docs/CURRENT_HANDOFF.md
実行・運用機能	docs/CAPABILITIES_AND_OPERATIONS.md
skforecast 0.23.0 operator evidence	docs/SKFORECAST_RUNTIME_CERTIFICATION.md
Darts current state	docs/darts/CURRENT_STATE_DARTS.md
dynamic sklearn	docs/SKLEARN_ALL_MODELS.md
parallel Broad campaign	docs/PARALLEL_UNIFIED_CAMPAIGN.md
LightGBM GPU	docs/LIGHTGBM_GPU_CERTIFICATION.md
TSFM	docs/TSFM_RUNTIME_CAPABILITIES.md

---

1. 現在地

領域	状態	現在確認できること	まだ意味しないこと
6ゲーム geometry	VERIFIED	positions・値域・select/digits契約	全モデル×全ゲーム完走ではない
Broad v1	VERIFIED / FROZEN	174 canonical identities	upstream実モデル総数ではない
Probabilistic v1	VERIFIED / PARTIALLY_VERIFIED	effective catalog 76	Broad plannerへ自動結合されない
Combined accounting	250 identities	174 + 76、6ゲーム換算1,500 cells	current単一campaignが1,500行を生成する意味ではない
Broad campaign planner	VERIFIED CONTRACT	174 × 6 = 1,044 units	probabilistic 76を含まない
Expanded v2	PHASE 4A CANDIDATE / SOURCE-BACKED	AutoGluon 37 + GluonTS 9 + skforecast 27を含むderived total 244	244全件runtime-certified、最終inventory freezeではない
scikit-learn dynamic	VERIFIED / PARTIALLY_VERIFIED	installed-version discovery / smoke / certify	全estimator成功保証ではない
StatsForecast	VERIFIED / PARTIALLY_VERIFIED	Broad 41 / shared 8 / development evidence	Holdout/Prospectiveではない
NeuralForecast fixed	VERIFIED / PARTIALLY_VERIFIED	Broad 37 / shared subset 17	37全件OOF完了ではない
NeuralForecast Auto	VERIFIED / PARTIALLY_VERIFIED	official 36 / Ray / Optuna / GPU evidence path	36×6 formal完了ではない
AutoGluon	EXPANDED 37 / PARTIAL RUNTIME	29 source models + 8 unique ensembles	37全件runtime-certifiedではない
Darts	BROAD 1 / SOURCE SURFACE >1	58 public forecasting exportsをPhase 2で調査、local NLinear/DLinear GPU evidence	58全件standalone/runtime-certifiedではない
GluonTS	BROAD 1 / EXPANDED 9	P6 registryの9 estimatorをExpanded identityへ統合。Draft #309 exact-headは2 lanes × 9 = 18/18 CPU lifecycle VERIFIED	18 unique modelsではない。#309 main統合/GPU/OOFではない
sktime	BROAD 1 / REGISTRY 141	141 discovered/importable、53 core-compatible、88 optional、formal P1=4	141全件runtime-certifiedではない
skforecast	BROAD 1 / EXPANDED 27 CANDIDATE	pinned 0.23.0 sourceからreviewed 27 identitiesを固定	27 runtime-certifiedではない。15 operator-local PASS / 2 BLOCKED / 10 NOT_RUN
ReservoirPy	BROAD 1 / EXPANSION OPEN	current Broad identity + expansion issue #294	final scientifically distinct count未freeze
TSFM	PARTIALLY_VERIFIED	retained 21中19 CERTIFIED / 2 BLOCKED	全19 OOF済みではない
Holdout	CLOSED	explicit authorizationまで閉鎖	development結果から自動解禁されない
Prospective	CLOSED	sealed future predictionのみ	Holdout未承認で自動進行しない
Automatic promotion	FORBIDDEN	human approval前提	runtime PASSだけでchampion化しない

状態語

status	意味
VERIFIED	stated current-code/evidence scopeで確認済み
PARTIALLY_VERIFIED	一部identity/lane/environmentのみ成立
OPERATOR_LOCAL_EVIDENCE	maintainer-host exact-source evidence、current-main retained certificationとは別
EXACT_HEAD_VERIFIED	特定PR/source SHA上の証拠。merge後mainとは別
LOCAL_VERIFIED / MAIN_PENDING	local exact worktreeで成立、main未反映
EXECUTION_PENDING	実装/計画あり、対象分母の完走なし
BLOCKED	dependency/license/runner/policy/artifactで停止
NOT CERTIFIED	成功証拠なし、fail-closed

---

2. Inventory countをどう読むか

2.1 Broad v1 = 174 は凍結分母

次の表は現在のライブラリ総モデル数ではありません。既存Broad campaignとの互換性のため凍結したcanonical identity数です。

Library	Broad v1 count (frozen)
builtin	4
scikit-learn	7
LightGBM	2
XGBoost	1
CatBoost	1
StatsForecast	41
NeuralForecast fixed	37
NeuralForecast Auto	36
MLForecast Auto	8
HierarchicalForecast	10
TSFM	21
AutoGluon	1 umbrella
Darts	1 umbrella
GluonTS	1 canonical identity
ReservoirPy	1 umbrella
sktime	1 umbrella
skforecast	1 Broad identity
TOTAL	174

2.2 umbrellaの「1」と現在確認できる実体数

Library	Broad v1	Expanded v2 candidate	Other denominator	Current gate
AutoGluon	1	37	29 base + 8 unique ensembles	merged
Darts	1	1	58 public forecasting exports	#286 / TAJ-27 open
GluonTS	1	9	9 estimator algorithms, 2 isolated lanes = 18 lifecycle cells	#323 merged; runtime gate separate
ReservoirPy	1	1	final distinct count未freeze	#294 open
sktime	1	1	141 discovered/importable	#289 / TAJ-32 Phase 4B open
skforecast	1	27	27 reviewed source-backed identities	PR #324 Phase 4A; formal repository runtime certificationは別

2.3 current candidate total

Broad v1                                      = 174
Probabilistic effective v1                    = 76
Combined Broad + Probabilistic accounting     = 250
Current Broad campaign planner                = 174 × 6 = 1,044
Combined accounting × six games               = 250 × 6 = 1,500
Expanded v2 current main after GluonTS         = 218
Expanded v2 with skforecast Phase 4A candidate = 244

218 - skforecast Broad copy 1 + skforecast implementations 27 = 244

Darts / ReservoirPy / sktime / Time-Series-Library / BasicTSの分解はまだ244へ入っていないため、244はExpanded v2の最終freeze値ではありません。

詳細: docs/INVENTORY_COUNT_BOUNDARIES.md

---

3. skforecast — Broad 1 / Expanded 27 candidate

Phase 4Aはoperator evidenceだけでなく、固定したupstream source skforecast v0.23.0 を再監査してmanifestを作ります。

package = skforecast==0.23.0
upstream_tag = v0.23.0
upstream_commit = c881d5d350426985c1c31373077b7d5b620f233d
operator_evidence_head = 9fcc1274755dca64c46dc31a9a0f60a9ef1c4ebd

Pinned source confirms:

• recursive exports: ForecasterEquivalentDate, ForecasterRecursive, ForecasterRecursiveClassifier, ForecasterRecursiveMultiSeries, ForecasterStats;
• direct exports: ForecasterDirect, ForecasterDirectMultiVariate;
• deep-learning strategy: ForecasterRnn;
• foundation strategy: ForecasterFoundation / FoundationModel;
• ForecasterStats explicitly supports 7 statistical implementations;
• FoundationModel explicitly lists 8 selectable model IDs including three Chronos-2 IDs.

無限のwrapper × arbitrary sklearn estimator Cartesian productは作りません。

Group	Count	Evidence boundary
Recursive regression families	5	OPERATOR_LOCAL_PASS
Recursive classifier representative binding	1	SOURCE_DECLARED / NOT_RUN
Direct Ridge	1	OPERATOR_LOCAL_PASS
Recursive multi-series Ridge	1	OPERATOR_LOCAL_PASS
Direct multivariate Ridge	1	OPERATOR_LOCAL_PASS
EquivalentDate	1	OPERATOR_LOCAL_PASS
ForecasterStats supported implementations	7	ARAR PASS / other 6 NOT_RUN
RNN LSTM / GRU	2	OPERATOR_LOCAL_PASS
Foundation explicit model IDs	8	mixed PASS / BLOCKED / NOT_RUN
Total	27	15 PASS / 2 BLOCKED / 10 NOT_RUN

全27 rowで runtime_certified=false を維持します。Moirai-2はnormal dependency routeがBLOCKED、TabPFN-TS v3はinvalid/expired authenticationでcheckpoint取得前に停止、source-only追加rowはNOT_RUNです。

---

4. GluonTS — Broad 1 / Expanded 9 / lifecycle cells 18

Phase 3ではsrc/loto/adapters/gluonts/p6_registry.pyをsource of truthとして、Broadのcanonical identityを壊さず9個のlibrary-specific Expanded identityへ分解します。

Broad v1 canonical identity        = 1 (`gluonts-deepar`)
Expanded v2 implementation count   = 9
P6 isolated lane cells             = 9 × 2 = 18

implementation_id	class
gluonts-torch-deepnpts	DeepNPTSEstimator
gluonts-torch-deepar	DeepAREstimator
gluonts-torch-tide	TiDEEstimator
gluonts-torch-simplefeedforward	SimpleFeedForwardEstimator
gluonts-torch-temporalfusiontransformer	TemporalFusionTransformerEstimator
gluonts-torch-wavenet	WaveNetEstimator
gluonts-torch-dlinear	DLinearEstimator
gluonts-torch-patchtst	PatchTSTEstimator
gluonts-torch-lagtst	LagTSTEstimator

Draft PR #309 exact headの18/18 CPU lifecycle VERIFIEDは別evidence classであり、inventory登録だけでruntime-certifiedへ昇格させません。

詳細: docs/gluonts/GLUONTS_EXPANDED_V2_PHASE3.md

---

5. sktime — Broad 1 / registry 141 / formal P1 4

sktime = 1.0.1
Broad v1 umbrella = 1
registry discovered/importable = 141
core-compatible = 53
optional-dependency-declared = 88
formal P1 models = 4

formal P1 4モデルはfit/predict/save-load/formal verification PASSですが、141全件runtime-certifiedではありません。Phase 4Bではexact 141-row manifestを固定し、wrapper/composite/adapterと独立forecasterを分類してからExpanded v2へ統合します。

---

6. Darts / ReservoirPy の「1」

Darts Broad v1は1 umbrellaですが、58 public forecasting exportsをPhase 2 inventory対象として調査中です。ReservoirPyもBroad v1は1ですが、#294でscientifically distinct pipelinesのExpanded化を追跡しています。

---

7. 主要ライブラリ / 実行面

Library	Inventory view	Execution surface	Runtime evidence	OOF
sklearn Broad	7	Broad campaign	tree-specific	未完
sklearn dynamic	installed-version dependent	loto-sklearn	provider/certify surface	未完
XGBoost	Broad 1	resource-aware Broad campaign	CUDA exact-source VERIFIED	未完
CatBoost	Broad 1	resource-aware Broad campaign	GPU exact-source VERIFIED	未完
LightGBM	Broad 2	resource-aware Broad campaign	OpenCL GPU VERIFIED / CUDA tree learner not certified	未完
StatsForecast	Broad 41	shared 8 + campaign	lifecycle + development evidence	部分実行
MLForecast	Auto 8	direct 2 + Auto	backend dependent	未完
NeuralForecast fixed	37	shared subset + dedicated	GPU capable	未完
NeuralForecast Auto	36	AutoModel runner	Ray/Optuna/GPU	未完
AutoGluon	Broad 1 / Expanded 37	isolated	backend dependent	未完
Darts	Broad 1 / source surface 58	provider/campaign	local bounded GPU evidence	未完
GluonTS	Broad 1 / Expanded 9 / lane cells 18	shared + isolated P6 provider	#309 exact-head CPU lifecycle	未完
sktime	Broad 1 / registry 141 / Expanded 1	isolated	formal P1 4 PASS	未完
skforecast	Broad 1 / Expanded 27 candidate	Expanded inventory; routing separate	15 local PASS / 2 blocked / 10 not-run	未完
ReservoirPy	Broad 1	optional/shared	partial	未完
TSFM	21	provider-specific	retained 19/21 certified	未完
probabilistic	effective 76	separate catalog/run/API	backend-specific	combined planner未実装

---

8. Recent implementation/documentation boundary

PR	SHA / status	Scope
#293	f04cd876...	Expanded v2 foundation + AutoGluon 37
#301	3cc73dba...	dynamic sklearn provider
#302	7d75dadc...	parallel Broad campaign orchestration
#307	ed7d6c81...	sktime P1 normalization
#310	4f4f8579...	current-state docs + skforecast operator evidence
#315	770d5b97...	GluonTS count clarity
#316	eb988d29...	umbrella count boundary clarification
#323	45bcf60f...	GluonTS 9 Expanded identities merged
#324	open	skforecast 27 Phase 4A candidate; CI gate pending
#309	Draft	GluonTS P6/P7 runtime repair exact-head evidence

---

9. Scientific contract

Primary metric: Hit@±1.

Required companions: MAE / MSE / RMSE / position-wise Hit@±1 / all-position Hit@±1.

Required baselines: Random / fixed / mean / median / last/recent / frequency / statistical.

Train-only preprocessing / scaler / encoder / feature selection / HPO
-> chronological Validation / OOF
-> all configured seeds + mean / variance / worst
-> prediction SHA-256 seal + timestamp before actual read
-> explicit Holdout authorization
-> Holdout
-> sealed Prospective prediction
-> actual arrival / scoring
-> human promotion decision

Holdout=CLOSED. Prospective=CLOSED. Automatic promotion/retraining/registry write=FORBIDDEN.

---

10. Common commands

uv run loto3 games
uv run loto3 catalog --counts
uv run loto models list

# Broad-only plan: 174 × 6 = 1,044
uv run loto3 campaign --output unused --plan-only

# Expanded v2 Phase 4A candidate must report 244
uv run python -c 'from loto.models.implementation_catalog import expanded_inventory_counts; print(expanded_inventory_counts())'

# sktime P1
ROOT="$PWD" SKTIME_NO_PAUSE=1 bash scripts/run_sktime_p1_matrix_certification.sh

---

11. Source of truth

1. current code/configuration;
2. tests/workflows/repository-retained evidence;
3. exact PR/source evidence;
4. exact-source operator/local evidence with provenance;
5. merged PR/commit history;
6. live GitHub Issues / Linear state;
7. current documentation;
8. historical snapshots.

Count authorityは catalog_full.py（Broad v1）、implementation_catalog.py + pinned manifests（Expanded v2）、framework-specific registry（別分母）を分離して読みます。