docs: record covariance estimator decision

README.md

@@ -25,6 +25,7 @@
 - `backtest` — weight-based 多日仿真（rebalance_table / compute_nav / compare_to_benchmark）
 - `portfolio_construction` — 多期因子分数 → Top-K → 等权目标权重表
 - `research_pipeline` — 因子日 → 下一真实交易日 → 显式执行价 → 日末估值 → 成本后绩效（防前视编排）
+- `artifact` — 版本化、确定性、存储中立的完整 research run 事实表与 manifest
 - `attribution` — 基于实际成交后持仓的隔夜 / 日内 / 交易成本逐日收益归因与闭合审计
 - `metrics` — 绝对绩效 + 严格日期对齐的 TE / IR / alpha / beta 基准相对绩效
 - `factor_library` — 通用方法（turnover / winsorize / IC / OLS / jb_test）
@@ -136,6 +137,39 @@ print(attribution.residual)  # 应接近 0；否则说明贡献未闭合到账
 # benchmark_returns 必须与成本后 factor_backtest.returns 使用完全相同的日期索引。
 print(factor_backtest.benchmark_stats(benchmark_returns))
 
+# 下游稳定交付：显式提供代码版本、数据快照和时区，不在核心层写数据库。
+from quant_engine.artifact import build_research_run_artifact
+from quant_engine.risk import estimate_covariance_snapshot
+
+risk_date = factor_backtest.position_weights.index[-1].date()
+risk_snapshot = estimate_covariance_snapshot(
+    daily_asset_returns,
+    as_of_date=risk_date,
+    lookback_sessions=252,
+    min_observations=120,
+    data_snapshot_id="<risk-return-data-snapshot-id>",
+)
+
+artifact = build_research_run_artifact(
+    factor_backtest,
+    run_id="research-run-001",
+    strategy_id="alpha-top20",
+    strategy_name="Alpha Top 20",
+    strategy_version="1.0.0",
+    engine_version="1.2.0",
+    code_revision="<git-sha>",
+    data_snapshot_id="<data-snapshot-id>",
+    calendar="CN-A",
+    timezone="Asia/Shanghai",
+    started_at="2026-08-21T10:00:00+08:00",
+    finished_at="2026-08-21T10:01:00+08:00",
+    parameters={"top_k": 20, "lag_sessions": 1},
+    benchmark_id="000300.SH",
+    benchmark_returns=benchmark_returns,
+    risk_snapshots={risk_date: risk_snapshot},
+)
+print(artifact.manifest())
+
 # run_weight_backtest 是低层算子：只接受收益区间开始前已经生效的持仓权重。
 # 不要把 signal-date 的 factor_scores/decision_weights 直接传给它。
 backtest = run_weight_backtest(

docs/OPEN_SOURCE_REFERENCES.md

@@ -16,6 +16,36 @@
 当前核心不新增依赖。逐日收益归因必须从实际换仓前后持仓、成交记录、执行价和
 收盘估值推导；因子分数与目标权重只是意图，不能作为成交后归因事实源。
 
+## 2026-08-21：研究运行工件
+
+- 借鉴 [Qlib Recorder / RecordTemplate](https://github.com/microsoft/qlib/blob/main/qlib/workflow/record_temp.py)
+  将 signal、portfolio analysis 和 risk analysis 分成稳定事实，但不引入 Qlib 运行时；
+- 借鉴 [MLflow Tracking](https://mlflow.org/docs/latest/tracking/) 的 run / params /
+  metrics / artifacts 分层，但 MLflow 只保留为未来可选 exporter；
+- HTML、PNG 和 tearsheet 是可再生展示物，不能替代 NAV、成交、持仓、归因和绩效事实。
+
+因此 `ResearchRunArtifact` 使用显式 `schema_version`、`config_hash`、代码版本和数据
+快照身份，并提供确定性 JSON / SHA-256 manifest；核心层仍不写数据库或 artifact store。
+
+schema `1.1.0` 将 Qlib 的独立 risk-analysis artifact 思路与 Riskfolio-Lib 的 Euler
+component-risk 语义结合，但只保留本项目需要的轻量合同：协方差快照必须声明
+`snapshot_id`、`as_of_date`、收益频率和年化期数；风险从成交后的实际日末持仓计算，
+component risk 闭合到年化组合波动，percentage contribution 闭合到 1。未来日期、资产
+标签不完整和零方差组合都直接失败，不以默认值伪造结果。
+
+## 2026-08-21：协方差快照估计
+
+| 项目 | 借鉴内容 | 当前决策 |
+|---|---|---|
+| [PyPortfolioOpt risk models](https://github.com/PyPortfolio/PyPortfolioOpt/blob/main/pypfopt/risk_models.py) | 将收益输入、协方差估计器和组合优化解耦；sample / EWM / shrinkage 使用统一标签输出 | 借鉴可替换估计器边界，不引入完整包 |
+| [scikit-learn covariance](https://github.com/scikit-learn/scikit-learn/blob/main/sklearn/covariance/_shrunk_covariance.py) | 维护成熟的 Ledoit–Wolf / OAS shrinkage 实现 | 未来作为可选 adapter；不复制统计公式 |
+| [Qlib structured risk model](https://github.com/microsoft/qlib/blob/main/qlib/model/riskmodel/structured.py) | PCA/FA 结构化协方差和固定随机状态 | 留作因子风险模型阶段，不进入当前 baseline |
+
+当前 `estimate_covariance_snapshot` 只编排 pandas 的 sample covariance：先按 `as_of_date`
+截断，再取固定 session 窗口，使用 complete-case 行并拒绝历史不足；禁止 pandas 默认的
+pairwise 样本集合产生含义不一致的矩阵。snapshot ID 对窗口数据、缺失掩码、上游数据
+快照身份和估计参数做 SHA-256，追加未来数据不会改变历史快照。
+
 ## hikyuu 的定位
 
 [hikyuu](https://github.com/fasiondog/hikyuu) 的 SG / MM / CN / PG 部件化思想、

docs/handoff/2026-08-21-research-artifact-contract.md

@@ -0,0 +1,48 @@
+# Research artifact contract handoff
+
+## Goal
+
+把完整可信研究链固化成存储中立、版本化、确定性的 `ResearchRunArtifact`，供
+`research_results` 持久化和 `research_platform` 查询：
+
+- run identity / schema version / config hash / code revision / data snapshot；
+- signal scores / decision weights / signal-to-execution mapping；
+- NAV / returns / benchmark / costs；
+- trades / realized positions / cash；
+- asset and daily return attribution；
+- performance including Sortino / TE / IR / alpha / beta；
+- reproducible covariance snapshots and annualized Euler component-risk facts；
+- canonical JSON / SHA-256 manifest。
+
+## Branch stack
+
+- 当前：`codex/research-artifact-contract-20260821`
+- 基线：`codex/ledger-attribution-20260821`（Draft PR #4）
+- 下层：Draft PR #3 → Ready PR #2 → `main`
+
+不得绕过堆叠顺序直接合并到 `main`。
+
+## Verification
+
+- `pytest -q`: 529 passed，9 个既有 SciPy warning；
+- risk/artifact focused coverage 87.85%；
+- `mypy --strict src/`: 16 source files passed；
+- changed-scope Ruff: passed；
+- no runtime dependency added；
+- no database, network, broker or filesystem write side effect in artifact builder。
+
+## Current risk contract
+
+- artifact schema：`1.1.0`；
+- `CovarianceSnapshot` 对输入矩阵深拷贝并显式记录截至日、频率和年化期数；
+- `risk_snapshots` 按研究交易日映射，可只生成需要的风险观察日；
+- 使用成交后实际持仓，不包含现金风险资产；协方差资产标签必须与研究资产全集一致；
+- `covariance_as_of_date` 不得晚于 `trade_date`；无正组合方差时拒绝产物。
+- `estimate_covariance_snapshot` 从显式数据快照的日收益生成无前视、complete-case、
+  SHA-256 可复现的 per-period sample covariance；不包含 I/O 或未来行。
+
+## Next action
+
+保持 Draft PR #5，不绕过堆叠顺序合并；下游 `research_results` / `research_platform`
+已在各自 Draft 分支兼容 1.0.0 / 1.1.0，下一阶段让市场数据适配器提供带稳定
+`data_snapshot_id` 的资产日收益，再评估可选 scikit-learn shrinkage adapter。

src/quant_engine/artifact.py

@@ -0,0 +1,577 @@
+"""Versioned, deterministic research-run artifacts for downstream adapters.
+
+This module is deliberately storage-neutral.  It snapshots a completed
+``FactorBacktestResult`` into queryable fact tables but never writes a database,
+starts a service, or talks to a broker.  ``research_results`` owns persistence;
+``research_platform`` owns read models and presentation.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import math
+from collections.abc import Mapping
+from dataclasses import dataclass
+from datetime import date, datetime
+from typing import Any
+
+import numpy as np
+import pandas as pd
+
+from quant_engine.research_pipeline import FactorBacktestResult
+from quant_engine.risk import CovarianceSnapshot, labeled_component_risk
+
+RESEARCH_ARTIFACT_SCHEMA_VERSION = "1.1.0"
+
+RISK_COLUMNS = [
+    "run_id",
+    "trade_date",
+    "asset_id",
+    "weight",
+    "marginal_risk",
+    "component_risk",
+    "risk_contribution",
+    "covariance_snapshot_id",
+    "covariance_as_of_date",
+    "risk_measure",
+    "return_frequency",
+    "periods_per_year",
+]
+
+__all__ = [
+    "RESEARCH_ARTIFACT_SCHEMA_VERSION",
+    "ResearchRunArtifact",
+    "build_research_run_artifact",
+]
+
+
+def _frame_copy(frame: pd.DataFrame) -> pd.DataFrame:
+    return frame.copy(deep=True)
+
+
+@dataclass(frozen=True, slots=True, eq=False)
+class ResearchRunArtifact:
+    """Immutable-by-interface snapshot of one completed research run."""
+
+    schema_version: str
+    _run: pd.DataFrame
+    _signals: pd.DataFrame
+    _nav: pd.DataFrame
+    _trades: pd.DataFrame
+    _positions: pd.DataFrame
+    _attribution: pd.DataFrame
+    _attribution_daily: pd.DataFrame
+    _risk: pd.DataFrame
+    _performance: pd.DataFrame
+
+    @property
+    def run(self) -> pd.DataFrame:
+        return _frame_copy(self._run)
+
+    @property
+    def nav(self) -> pd.DataFrame:
+        return _frame_copy(self._nav)
+
+    @property
+    def signals(self) -> pd.DataFrame:
+        return _frame_copy(self._signals)
+
+    @property
+    def trades(self) -> pd.DataFrame:
+        return _frame_copy(self._trades)
+
+    @property
+    def positions(self) -> pd.DataFrame:
+        return _frame_copy(self._positions)
+
+    @property
+    def attribution(self) -> pd.DataFrame:
+        return _frame_copy(self._attribution)
+
+    @property
+    def attribution_daily(self) -> pd.DataFrame:
+        return _frame_copy(self._attribution_daily)
+
+    @property
+    def risk(self) -> pd.DataFrame:
+        return _frame_copy(self._risk)
+
+    @property
+    def performance(self) -> pd.DataFrame:
+        return _frame_copy(self._performance)
+
+    def table_frames(self) -> Mapping[str, pd.DataFrame]:
+        """Return isolated table snapshots keyed by stable logical table name."""
+        return {
+            "run": self.run,
+            "signals": self.signals,
+            "nav": self.nav,
+            "trades": self.trades,
+            "positions": self.positions,
+            "attribution": self.attribution,
+            "attribution_daily": self.attribution_daily,
+            "risk": self.risk,
+            "performance": self.performance,
+        }
+
+    def canonical_json(self) -> str:
+        """Serialize tables deterministically for checksums and artifact storage."""
+        payload = {
+            "schema_version": self.schema_version,
+            "tables": {
+                name: _frame_records(frame)
+                for name, frame in self._internal_table_frames().items()
+            },
+        }
+        return json.dumps(
+            payload,
+            ensure_ascii=False,
+            sort_keys=True,
+            separators=(",", ":"),
+            allow_nan=False,
+        )
+
+    @property
+    def content_sha256(self) -> str:
+        return hashlib.sha256(self.canonical_json().encode("utf-8")).hexdigest()
+
+    def manifest(self) -> Mapping[str, object]:
+        """Return a compact immutable identity and row-count manifest."""
+        return {
+            "schema_version": self.schema_version,
+            "run_id": str(self._run.at[0, "run_id"]),
+            "config_hash": str(self._run.at[0, "config_hash"]),
+            "content_sha256": self.content_sha256,
+            "tables": {
+                name: len(frame) for name, frame in self._internal_table_frames().items()
+            },
+        }
+
+    def _internal_table_frames(self) -> Mapping[str, pd.DataFrame]:
+        return {
+            "run": self._run,
+            "signals": self._signals,
+            "nav": self._nav,
+            "trades": self._trades,
+            "positions": self._positions,
+            "attribution": self._attribution,
+            "attribution_daily": self._attribution_daily,
+            "risk": self._risk,
+            "performance": self._performance,
+        }
+
+
+def _required_text(value: str, name: str, *, max_length: int | None = None) -> str:
+    normalized = value.strip()
+    if not normalized:
+        raise ValueError(f"{name} must be non-empty")
+    if max_length is not None and len(normalized) > max_length:
+        raise ValueError(f"{name} must contain at most {max_length} characters")
+    return normalized
+
+
+def _aware_timestamp(value: str | pd.Timestamp, name: str) -> pd.Timestamp:
+    try:
+        timestamp = pd.Timestamp(value)
+    except (TypeError, ValueError) as error:
+        raise ValueError(f"{name} must be a valid timestamp") from error
+    if timestamp.tzinfo is None:
+        raise ValueError(f"{name} must include a timezone")
+    return timestamp
+
+
+def _json_value(value: object) -> object:
+    if value is None or isinstance(value, str | bool | int):
+        return value
+    if isinstance(value, float):
+        return value if math.isfinite(value) else None
+    if isinstance(value, np.generic):
+        return _json_value(value.item())
+    if isinstance(value, pd.Timestamp):
+        return value.isoformat()
+    if isinstance(value, datetime):
+        return value.isoformat()
+    if isinstance(value, date):
+        return value.isoformat()
+    if isinstance(value, Mapping):
+        return {
+            str(key): _json_value(item)
+            for key, item in sorted(value.items(), key=lambda pair: str(pair[0]))
+        }
+    if isinstance(value, list | tuple):
+        return [_json_value(item) for item in value]
+    raise TypeError(f"value of type {type(value).__name__} is not JSON serializable")
+
+
+def _canonical_mapping_json(values: Mapping[str, object]) -> str:
+    normalized = _json_value(values)
+    return json.dumps(
+        normalized,
+        ensure_ascii=False,
+        sort_keys=True,
+        separators=(",", ":"),
+        allow_nan=False,
+    )
+
+
+def _frame_records(frame: pd.DataFrame) -> list[dict[str, object]]:
+    return [
+        {str(key): _json_value(value) for key, value in row.items()}
+        for row in frame.to_dict(orient="records")
+    ]
+
+
+def _build_nav(
+    result: FactorBacktestResult,
+    run_id: str,
+    benchmark_returns: pd.Series | None,
+) -> pd.DataFrame:
+    nav = result.execution.ledger_frame.copy(deep=True)
+    nav.insert(0, "run_id", run_id)
+    nav["trade_date"] = pd.to_datetime(nav["trade_date"]).dt.date
+    nav["total_cost"] = [
+        sum(execution.total_cost for execution in daily.executions)
+        for daily in result.execution.daily_executions
+    ]
+    if benchmark_returns is None:
+        nav["benchmark_nav"] = np.nan
+        nav["benchmark_return"] = np.nan
+        nav["excess_ret"] = np.nan
+    else:
+        benchmark = benchmark_returns.astype(float, copy=True)
+        nav["benchmark_nav"] = (1.0 + benchmark).cumprod().to_numpy()
+        nav["benchmark_return"] = benchmark.to_numpy()
+        nav["excess_ret"] = result.returns.to_numpy() - benchmark.to_numpy()
+    return nav
+
+
+def _build_signals(result: FactorBacktestResult, run_id: str) -> pd.DataFrame:
+    columns = [
+        "run_id",
+        "signal_date",
+        "execution_date",
+        "asset_id",
+        "factor_score",
+        "target_weight",
+    ]
+    rows: list[dict[str, object]] = []
+    for signal_date, scores in result.factor_scores.iterrows():
+        execution_date = pd.Timestamp(result.schedule.signal_to_execution.at[signal_date]).date()
+        for asset, score in scores.items():
+            rows.append(
+                {
+                    "run_id": run_id,
+                    "signal_date": pd.Timestamp(signal_date).date(),
+                    "execution_date": execution_date,
+                    "asset_id": asset,
+                    "factor_score": float(score),
+                    "target_weight": float(
+                        result.schedule.decision_weights.at[signal_date, asset]
+                    ),
+                }
+            )
+    return pd.DataFrame(rows, columns=columns)
+
+
+def _build_trades(result: FactorBacktestResult, run_id: str) -> pd.DataFrame:
+    trades = result.execution.trades_frame.copy(deep=True)
+    trades.insert(0, "run_id", run_id)
+    trades["trade_date"] = pd.to_datetime(trades["trade_date"]).dt.date
+    trades.insert(
+        1,
+        "trade_id",
+        [f"{run_id}:{sequence:08d}" for sequence in range(1, len(trades) + 1)],
+    )
+    signal_by_execution = {
+        pd.Timestamp(execution_date).date(): pd.Timestamp(signal_date).date()
+        for signal_date, execution_date in result.schedule.signal_to_execution.items()
+    }
+    trades["signal_id"] = [
+        f"{run_id}:signal:{signal_by_execution[trade_date].isoformat()}"
+        for trade_date in trades["trade_date"]
+    ]
+    trades["total_cost"] = trades["fee"] + trades["slippage"]
+    return trades
+
+
+def _build_positions(result: FactorBacktestResult, run_id: str) -> pd.DataFrame:
+    columns = [
+        "run_id",
+        "trade_date",
+        "asset_id",
+        "asset_type",
+        "quantity",
+        "mark_price",
+        "market_value",
+        "weight",
+    ]
+    rows: list[dict[str, object]] = []
+    weights = result.position_weights
+    cash_weights = result.cash_weights
+    for date_value, position in zip(
+        result.valuation_prices.index,
+        result.execution.positions,
+        strict=True,
+    ):
+        session_date = pd.Timestamp(date_value).date()
+        for asset, quantity in position.holdings.items():
+            mark_price = float(result.valuation_prices.at[date_value, asset])
+            rows.append(
+                {
+                    "run_id": run_id,
+                    "trade_date": session_date,
+                    "asset_id": asset,
+                    "asset_type": "security",
+                    "quantity": quantity,
+                    "mark_price": mark_price,
+                    "market_value": quantity * mark_price,
+                    "weight": float(weights.at[date_value, asset]),
+                }
+            )
+        rows.append(
+            {
+                "run_id": run_id,
+                "trade_date": session_date,
+                "asset_id": "CASH",
+                "asset_type": "cash",
+                "quantity": position.cash,
+                "mark_price": 1.0,
+                "market_value": position.cash,
+                "weight": float(cash_weights.at[date_value]),
+            }
+        )
+    return pd.DataFrame(rows, columns=columns)
+
+
+def _build_attribution(
+    result: FactorBacktestResult,
+    run_id: str,
+) -> tuple[pd.DataFrame, pd.DataFrame]:
+    contribution = result.return_attribution()
+    rows: list[dict[str, object]] = []
+    for date_value in contribution.overnight.index:
+        for asset in contribution.overnight.columns:
+            overnight = float(contribution.overnight.at[date_value, asset])
+            intraday = float(contribution.intraday.at[date_value, asset])
+            rows.append(
+                {
+                    "run_id": run_id,
+                    "trade_date": pd.Timestamp(date_value).date(),
+                    "asset_id": asset,
+                    "overnight": overnight,
+                    "intraday": intraday,
+                    "asset_total": overnight + intraday,
+                }
+            )
+    daily = pd.DataFrame(
+        {
+            "run_id": run_id,
+            "trade_date": contribution.total_return.index.date,
+            "transaction_cost": contribution.transaction_cost.to_numpy(),
+            "explained_return": contribution.explained_return.to_numpy(),
+            "residual": contribution.residual.to_numpy(),
+            "total_return": contribution.total_return.to_numpy(),
+        }
+    )
+    return pd.DataFrame(rows), daily
+
+
+def _risk_trade_date(value: object) -> date:
+    try:
+        timestamp = pd.Timestamp(value)
+    except (TypeError, ValueError) as error:
+        raise ValueError("risk snapshot keys must be valid trade dates") from error
+    if pd.isna(timestamp):
+        raise ValueError("risk snapshot keys must be valid trade dates")
+    return date(int(timestamp.year), int(timestamp.month), int(timestamp.day))
+
+
+def _build_risk(
+    result: FactorBacktestResult,
+    run_id: str,
+    risk_snapshots: Mapping[object, CovarianceSnapshot] | None,
+) -> pd.DataFrame:
+    if risk_snapshots is None:
+        return pd.DataFrame(columns=RISK_COLUMNS)
+    if not isinstance(risk_snapshots, Mapping):
+        raise TypeError("risk_snapshots must be a mapping")
+
+    session_by_date = {
+        pd.Timestamp(session).date(): session for session in result.position_weights.index
+    }
+    normalized: dict[date, CovarianceSnapshot] = {}
+    for raw_trade_date, snapshot in risk_snapshots.items():
+        trade_date = _risk_trade_date(raw_trade_date)
+        if trade_date in normalized:
+            raise ValueError(f"duplicate risk snapshot trade date: {trade_date}")
+        if trade_date not in session_by_date:
+            raise ValueError(f"risk snapshot trade date {trade_date} must be a result session")
+        if not isinstance(snapshot, CovarianceSnapshot):
+            raise TypeError("risk snapshot values must be CovarianceSnapshot instances")
+        if snapshot.as_of_date > trade_date:
+            raise ValueError(
+                f"covariance as_of_date {snapshot.as_of_date} must not be after trade date "
+                f"{trade_date}"
+            )
+        normalized[trade_date] = snapshot
+
+    weights_by_date = result.position_weights
+    rows: list[dict[str, object]] = []
+    for trade_date in sorted(normalized):
+        snapshot = normalized[trade_date]
+        session = session_by_date[trade_date]
+        weights = weights_by_date.loc[session].astype(float, copy=True)
+        annualized_covariance = snapshot.covariance * snapshot.periods_per_year
+        decomposition = labeled_component_risk(weights, annualized_covariance)
+        for asset_id in weights.index:
+            rows.append(
+                {
+                    "run_id": run_id,
+                    "trade_date": trade_date,
+                    "asset_id": asset_id,
+                    "weight": float(weights.loc[asset_id]),
+                    "marginal_risk": float(decomposition.marginal.loc[asset_id]),
+                    "component_risk": float(decomposition.component.loc[asset_id]),
+                    "risk_contribution": float(decomposition.percentage.loc[asset_id]),
+                    "covariance_snapshot_id": snapshot.snapshot_id,
+                    "covariance_as_of_date": snapshot.as_of_date,
+                    "risk_measure": "annualized_volatility",
+                    "return_frequency": snapshot.return_frequency,
+                    "periods_per_year": snapshot.periods_per_year,
+                }
+            )
+    return pd.DataFrame(rows, columns=RISK_COLUMNS)
+
+
+def _build_performance(
+    result: FactorBacktestResult,
+    run_id: str,
+    benchmark_returns: pd.Series | None,
+) -> pd.DataFrame:
+    stats = result.stats()
+    relative = (
+        result.benchmark_stats(benchmark_returns)
+        if benchmark_returns is not None
+        else {
+            "tracking_error": float("nan"),
+            "information_ratio": float("nan"),
+            "alpha": float("nan"),
+            "beta": float("nan"),
+        }
+    )
+    total_return = float(result.nav.iloc[-1] - 1.0)
+    return pd.DataFrame(
+        [
+            {
+                "run_id": run_id,
+                "total_ret": total_return,
+                "ann_ret": stats["ann_return"],
+                "ann_volatility": stats["ann_volatility"],
+                "sharpe": stats["sharpe"],
+                "sortino": stats["sortino"],
+                "max_dd": stats["max_drawdown"],
+                "calmar": stats["calmar"],
+                "win_rate": stats["win_rate"],
+                "tracking_error": relative["tracking_error"],
+                "ir": relative["information_ratio"],
+                "alpha": relative["alpha"],
+                "beta": relative["beta"],
+                "n_trades": len(result.execution.trades_frame),
+                "n_days": len(result.returns),
+            }
+        ]
+    )
+
+
+def build_research_run_artifact(
+    result: FactorBacktestResult,
+    *,
+    run_id: str,
+    strategy_id: str,
+    strategy_name: str,
+    strategy_version: str,
+    engine_version: str,
+    code_revision: str,
+    data_snapshot_id: str,
+    calendar: str,
+    timezone: str,
+    started_at: str | pd.Timestamp,
+    finished_at: str | pd.Timestamp,
+    parameters: Mapping[str, object],
+    benchmark_id: str | None = None,
+    benchmark_returns: pd.Series | None = None,
+    risk_snapshots: Mapping[object, CovarianceSnapshot] | None = None,
+) -> ResearchRunArtifact:
+    """Snapshot one successful factor backtest into schema-versioned fact tables."""
+    if not isinstance(result, FactorBacktestResult):
+        raise TypeError("result must be a FactorBacktestResult")
+    if result.nav.empty:
+        raise ValueError("result must contain at least one research session")
+    normalized_run_id = _required_text(run_id, "run_id", max_length=64)
+    normalized_strategy_id = _required_text(strategy_id, "strategy_id")
+    normalized_strategy_name = _required_text(strategy_name, "strategy_name")
+    normalized_strategy_version = _required_text(strategy_version, "strategy_version")
+    normalized_engine_version = _required_text(engine_version, "engine_version")
+    normalized_code_revision = _required_text(code_revision, "code_revision")
+    normalized_snapshot = _required_text(data_snapshot_id, "data_snapshot_id")
+    normalized_calendar = _required_text(calendar, "calendar")
+    normalized_timezone = _required_text(timezone, "timezone")
+    if not isinstance(parameters, Mapping):
+        raise TypeError("parameters must be a mapping")
+
+    started = _aware_timestamp(started_at, "started_at")
+    finished = _aware_timestamp(finished_at, "finished_at")
+    if finished < started:
+        raise ValueError("finished_at must not precede started_at")
+    if (benchmark_id is None) != (benchmark_returns is None):
+        raise ValueError("benchmark_id and benchmark_returns must be provided together")
+    normalized_benchmark = ""
+    if benchmark_id is not None:
+        normalized_benchmark = _required_text(benchmark_id, "benchmark_id")
+        result.benchmark_stats(benchmark_returns)
+
+    params_json = _canonical_mapping_json(parameters)
+    config_hash = hashlib.sha256(params_json.encode("utf-8")).hexdigest()
+    run = pd.DataFrame(
+        [
+            {
+                "schema_version": RESEARCH_ARTIFACT_SCHEMA_VERSION,
+                "run_id": normalized_run_id,
+                "strategy_id": normalized_strategy_id,
+                "strategy_name": normalized_strategy_name,
+                "strategy_version": normalized_strategy_version,
+                "engine_version": normalized_engine_version,
+                "code_revision": normalized_code_revision,
+                "config_hash": config_hash,
+                "data_snapshot_id": normalized_snapshot,
+                "benchmark_id": normalized_benchmark,
+                "benchmark_alignment_policy": (
+                    "exact_session_index" if benchmark_returns is not None else "none"
+                ),
+                "frequency": "1d",
+                "calendar": normalized_calendar,
+                "timezone": normalized_timezone,
+                "initial_capital": result.execution.initial_cash,
+                "start_date": result.nav.index[0].date(),
+                "end_date": result.nav.index[-1].date(),
+                "status": "success",
+                "started_at": started,
+                "finished_at": finished,
+                "params_json": params_json,
+            }
+        ]
+    )
+    attribution, attribution_daily = _build_attribution(result, normalized_run_id)
+    return ResearchRunArtifact(
+        schema_version=RESEARCH_ARTIFACT_SCHEMA_VERSION,
+        _run=run,
+        _signals=_build_signals(result, normalized_run_id),
+        _nav=_build_nav(result, normalized_run_id, benchmark_returns),
+        _trades=_build_trades(result, normalized_run_id),
+        _positions=_build_positions(result, normalized_run_id),
+        _attribution=attribution,
+        _attribution_daily=attribution_daily,
+        _risk=_build_risk(result, normalized_run_id, risk_snapshots),
+        _performance=_build_performance(result, normalized_run_id, benchmark_returns),
+    )

src/quant_engine/metrics.py

@@ -65,6 +65,20 @@ def sharpe_ratio(r: pd.Series, rf: float = 0.0) -> float:
     return (annualized_return(r) - rf) / vol
 
 
+def sortino_ratio(r: pd.Series, rf: float = 0.0) -> float:
+    """Sortino = (年化收益 - rf) / 年化下行偏差。"""
+    r = _clean(r)
+    if len(r) < 2:
+        return 0.0
+    downside = np.minimum(r.to_numpy(dtype=float), 0.0)
+    downside_deviation = float(
+        np.sqrt(np.mean(np.square(downside))) * np.sqrt(TRADING_DAYS_PER_YEAR)
+    )
+    if downside_deviation == 0:
+        return 0.0
+    return (annualized_return(r) - rf) / downside_deviation
+
+
 def max_drawdown(r: pd.Series) -> float:
     """最大回撤（负数）。例如 -0.2 表示最大亏 20%。"""
     r = _clean(r)
@@ -120,6 +134,7 @@ def summary(r: pd.Series, rf: float = 0.0) -> Mapping[str, float]:
         "ann_return": ann_ret,
         "ann_volatility": ann_vol,
         "sharpe": sharpe_ratio(r, rf),
+        "sortino": sortino_ratio(r, rf),
         "max_drawdown": mdd,
         "calmar": calmar_ratio(r),
         "win_rate": win_rate(r),

src/quant_engine/risk.py

@@ -5,7 +5,10 @@
 
 from __future__ import annotations
 
+import hashlib
+import json
 from dataclasses import dataclass
+from datetime import date
 from typing import Any
 
 import numpy as np
@@ -14,13 +17,260 @@ from numpy.typing import NDArray
 
 __all__ = [
     "ComponentRiskResult",
+    "CovarianceSnapshot",
     "component_var",
+    "estimate_covariance_snapshot",
     "labeled_component_risk",
     "marginal_risk_contribution",
     "risk_contribution",
 ]
 
 
+@dataclass(frozen=True, slots=True, init=False, eq=False)
+class CovarianceSnapshot:
+    """Immutable-by-interface covariance input with explicit time semantics."""
+
+    snapshot_id: str
+    as_of_date: date
+    _covariance: pd.DataFrame
+    return_frequency: str
+    periods_per_year: int
+    method: str
+    window_start_date: date | None
+    window_end_date: date | None
+    observations: int | None
+    lookback_sessions: int | None
+    missing_policy: str
+    data_snapshot_id: str
+    input_sha256: str
+
+    def __init__(
+        self,
+        *,
+        snapshot_id: str,
+        as_of_date: str | date | pd.Timestamp,
+        covariance: pd.DataFrame,
+        return_frequency: str,
+        periods_per_year: int,
+        method: str = "provided",
+        window_start_date: str | date | pd.Timestamp | None = None,
+        window_end_date: str | date | pd.Timestamp | None = None,
+        observations: int | None = None,
+        lookback_sessions: int | None = None,
+        missing_policy: str = "provided",
+        data_snapshot_id: str = "",
+        input_sha256: str = "",
+    ) -> None:
+        if not isinstance(snapshot_id, str) or not snapshot_id.strip():
+            raise ValueError("snapshot_id must be non-empty")
+        if not isinstance(return_frequency, str) or not return_frequency.strip():
+            raise ValueError("return_frequency must be non-empty")
+        if isinstance(periods_per_year, bool) or not isinstance(periods_per_year, int):
+            raise TypeError("periods_per_year must be an integer")
+        if periods_per_year <= 0:
+            raise ValueError("periods_per_year must be positive")
+        if not isinstance(covariance, pd.DataFrame):
+            raise TypeError("covariance must be a pandas DataFrame")
+        if covariance.empty:
+            raise ValueError("covariance must contain at least one asset")
+        if not isinstance(method, str) or not method.strip():
+            raise ValueError("method must be non-empty")
+        if not isinstance(missing_policy, str) or not missing_policy.strip():
+            raise ValueError("missing_policy must be non-empty")
+        for value, name in (
+            (observations, "observations"),
+            (lookback_sessions, "lookback_sessions"),
+        ):
+            if value is not None and (
+                isinstance(value, bool) or not isinstance(value, int) or value <= 0
+            ):
+                raise ValueError(f"{name} must be a positive integer when provided")
+        if input_sha256 and (
+            len(input_sha256) != 64
+            or any(character not in "0123456789abcdef" for character in input_sha256)
+        ):
+            raise ValueError("input_sha256 must be a lowercase SHA-256 digest")
+
+        normalized_as_of = _normalized_date(as_of_date, "as_of_date")
+        normalized_window_start = (
+            None
+            if window_start_date is None
+            else _normalized_date(window_start_date, "window_start_date")
+        )
+        normalized_window_end = (
+            None
+            if window_end_date is None
+            else _normalized_date(window_end_date, "window_end_date")
+        )
+        if (normalized_window_start is None) != (normalized_window_end is None):
+            raise ValueError("window_start_date and window_end_date must be provided together")
+        if (
+            normalized_window_start is not None
+            and normalized_window_end is not None
+            and normalized_window_start > normalized_window_end
+        ):
+            raise ValueError("window_start_date must not be after window_end_date")
+        if normalized_window_end is not None and normalized_window_end > normalized_as_of:
+            raise ValueError("window_end_date must not be after as_of_date")
+
+        object.__setattr__(self, "snapshot_id", snapshot_id.strip())
+        object.__setattr__(self, "as_of_date", normalized_as_of)
+        object.__setattr__(self, "_covariance", covariance.copy(deep=True))
+        object.__setattr__(self, "return_frequency", return_frequency.strip())
+        object.__setattr__(self, "periods_per_year", periods_per_year)
+        object.__setattr__(self, "method", method.strip())
+        object.__setattr__(self, "window_start_date", normalized_window_start)
+        object.__setattr__(self, "window_end_date", normalized_window_end)
+        object.__setattr__(self, "observations", observations)
+        object.__setattr__(self, "lookback_sessions", lookback_sessions)
+        object.__setattr__(self, "missing_policy", missing_policy.strip())
+        object.__setattr__(self, "data_snapshot_id", data_snapshot_id.strip())
+        object.__setattr__(self, "input_sha256", input_sha256)
+
+    @property
+    def covariance(self) -> pd.DataFrame:
+        """Return an isolated copy so callers cannot mutate the snapshot."""
+        return self._covariance.copy(deep=True)
+
+
+def _normalized_date(value: object, name: str) -> date:
+    try:
+        timestamp = pd.Timestamp(value)
+    except (TypeError, ValueError) as error:
+        raise ValueError(f"{name} must be a valid date") from error
+    if pd.isna(timestamp):
+        raise ValueError(f"{name} must be a valid date")
+    return date(int(timestamp.year), int(timestamp.month), int(timestamp.day))
+
+
+def _positive_integer(value: int, name: str, *, minimum: int = 1) -> int:
+    if isinstance(value, bool) or not isinstance(value, int) or value < minimum:
+        raise ValueError(f"{name} must be an integer of at least {minimum}")
+    return value
+
+
+def _input_fingerprint(window: pd.DataFrame, session_dates: list[date]) -> str:
+    values = window.to_numpy(dtype=float, copy=True)
+    missing = np.isnan(values)
+    normalized = np.where(missing, 0.0, values).astype("<f8", copy=False)
+    metadata = {
+        "assets": [str(asset) for asset in window.columns],
+        "sessions": [session.isoformat() for session in session_dates],
+        "shape": list(values.shape),
+    }
+    digest = hashlib.sha256(
+        json.dumps(metadata, sort_keys=True, separators=(",", ":")).encode("utf-8")
+    )
+    digest.update(missing.astype(np.uint8, copy=False).tobytes(order="C"))
+    digest.update(normalized.tobytes(order="C"))
+    return digest.hexdigest()
+
+
+def estimate_covariance_snapshot(
+    asset_returns: pd.DataFrame,
+    *,
+    as_of_date: str | date | pd.Timestamp,
+    lookback_sessions: int,
+    min_observations: int,
+    data_snapshot_id: str,
+    return_frequency: str = "1d",
+    periods_per_year: int = 252,
+) -> CovarianceSnapshot:
+    """Estimate a deterministic per-period sample covariance without look-ahead.
+
+    The selected lookback window is truncated at ``as_of_date`` before any
+    calculation. Rows containing a missing asset return are removed as complete
+    cases, preventing pairwise sample sets from producing an ambiguous matrix.
+    """
+    if not isinstance(asset_returns, pd.DataFrame):
+        raise TypeError("asset_returns must be a pandas DataFrame")
+    if asset_returns.empty or asset_returns.shape[1] == 0:
+        raise ValueError("asset_returns must contain observations and assets")
+    if not isinstance(asset_returns.index, pd.DatetimeIndex):
+        raise TypeError("asset_returns index must be a DatetimeIndex")
+    if not asset_returns.index.is_unique or not asset_returns.index.is_monotonic_increasing:
+        raise ValueError("asset_returns index must be unique and strictly increasing")
+    if not asset_returns.columns.is_unique:
+        raise ValueError("asset_returns must contain unique asset labels")
+    if any(not isinstance(asset, str) or not asset.strip() for asset in asset_returns.columns):
+        raise ValueError("asset_returns asset labels must be non-empty strings")
+
+    lookback = _positive_integer(lookback_sessions, "lookback_sessions")
+    minimum = _positive_integer(min_observations, "min_observations", minimum=2)
+    if minimum > lookback:
+        raise ValueError("min_observations must not exceed lookback_sessions")
+    normalized_data_snapshot_id = data_snapshot_id.strip()
+    if not normalized_data_snapshot_id:
+        raise ValueError("data_snapshot_id must be non-empty")
+    normalized_as_of = _normalized_date(as_of_date, "as_of_date")
+
+    returns = asset_returns.astype(float, copy=True)
+    values = returns.to_numpy()
+    if np.isinf(values).any():
+        raise ValueError("asset_returns must not contain infinite values")
+    session_dates = [
+        _normalized_date(index_value, "asset_returns index") for index_value in returns.index
+    ]
+    if len(set(session_dates)) != len(session_dates):
+        raise ValueError("asset_returns must contain at most one observation per session date")
+    historical_mask = [session <= normalized_as_of for session in session_dates]
+    window = returns.loc[historical_mask].tail(lookback)
+    if window.empty:
+        raise ValueError("asset_returns contain no observations on or before as_of_date")
+    window_dates = [
+        _normalized_date(index_value, "asset_returns index") for index_value in window.index
+    ]
+    complete = window.dropna(axis=0, how="any")
+    if len(complete) < minimum:
+        raise ValueError(
+            f"complete observations must be at least {minimum}; received {len(complete)}"
+        )
+
+    covariance = complete.cov(ddof=1)
+    covariance_values = covariance.to_numpy()
+    if not np.isfinite(covariance_values).all():
+        raise ValueError("sample covariance must be finite")
+    input_sha256 = _input_fingerprint(window, window_dates)
+    identity = {
+        "as_of_date": normalized_as_of.isoformat(),
+        "assets": list(returns.columns),
+        "data_snapshot_id": normalized_data_snapshot_id,
+        "estimator": "sample-cov-v1",
+        "input_sha256": input_sha256,
+        "lookback_sessions": lookback,
+        "min_observations": minimum,
+        "missing_policy": "complete_case",
+        "observations": len(complete),
+        "periods_per_year": periods_per_year,
+        "return_frequency": return_frequency,
+        "window_end_date": window_dates[-1].isoformat(),
+        "window_start_date": window_dates[0].isoformat(),
+    }
+    identity_bytes = json.dumps(
+        identity,
+        sort_keys=True,
+        separators=(",", ":"),
+    ).encode("utf-8")
+    digest = hashlib.sha256(identity_bytes)
+    digest.update(covariance_values.astype("<f8", copy=False).tobytes(order="C"))
+    snapshot_id = f"sample-cov-v1:{digest.hexdigest()}"
+    return CovarianceSnapshot(
+        snapshot_id=snapshot_id,
+        as_of_date=normalized_as_of,
+        covariance=covariance,
+        return_frequency=return_frequency,
+        periods_per_year=periods_per_year,
+        method="sample",
+        window_start_date=window_dates[0],
+        window_end_date=window_dates[-1],
+        observations=len(complete),
+        lookback_sessions=lookback,
+        missing_policy="complete_case",
+        data_snapshot_id=normalized_data_snapshot_id,
+        input_sha256=input_sha256,
+    )
+
+
 @dataclass(frozen=True, slots=True, eq=False)
 class ComponentRiskResult:
     """Label-preserving Euler decomposition of portfolio volatility."""

tests/test_artifact.py

@@ -0,0 +1,281 @@
+"""Stable research-run artifact contracts for downstream persistence."""
+
+from __future__ import annotations
+
+import json
+from datetime import date
+
+import pandas as pd
+import pytest
+
+from quant_engine.artifact import (
+    RESEARCH_ARTIFACT_SCHEMA_VERSION,
+    ResearchRunArtifact,
+    build_research_run_artifact,
+)
+from quant_engine.execution import ExecutionConfig
+from quant_engine.research_pipeline import FactorBacktestResult, run_factor_backtest_research
+from quant_engine.risk import CovarianceSnapshot
+
+
+def _backtest_result() -> FactorBacktestResult:
+    dates = pd.date_range("2026-01-05", periods=4, freq="B")
+    scores = pd.DataFrame(
+        {"A": [2.0, 0.0], "B": [1.0, 3.0]},
+        index=dates[:2],
+    )
+    opens = pd.DataFrame(
+        {"A": [10.0, 10.0, 15.0, 15.0], "B": [20.0, 20.0, 20.0, 21.0]},
+        index=dates,
+    )
+    closes = pd.DataFrame(
+        {"A": [10.0, 12.0, 15.0, 15.0], "B": [20.0, 20.0, 18.0, 21.0]},
+        index=dates,
+    )
+    return run_factor_backtest_research(
+        scores,
+        opens,
+        closes,
+        top_k=1,
+        execution_price_field="open",
+        valuation_price_field="close",
+        initial_cash=1_000.0,
+        config=ExecutionConfig(
+            commission_bps=0,
+            stamp_tax_bps=0,
+            slippage_bps=0,
+            min_trade_amount=0,
+        ),
+    )
+
+
+def _build(
+    result: FactorBacktestResult,
+    *,
+    parameters: dict[str, object] | None = None,
+    risk_snapshots: dict[date, CovarianceSnapshot] | None = None,
+) -> ResearchRunArtifact:
+    benchmark = pd.Series(
+        [0.0, 0.01, -0.01, 0.02],
+        index=result.returns.index,
+        name="benchmark_return",
+    )
+    return build_research_run_artifact(
+        result,
+        run_id="run-20260105-a",
+        strategy_id="alpha-top1",
+        strategy_name="Alpha Top 1",
+        strategy_version="1.0.0",
+        engine_version="1.2.0",
+        code_revision="3b1ad07",
+        data_snapshot_id="qtdb-pro-20260108-v1",
+        calendar="CN-A",
+        timezone="Asia/Shanghai",
+        started_at="2026-01-08T10:00:00+08:00",
+        finished_at="2026-01-08T10:01:00+08:00",
+        parameters=parameters or {"top_k": 1, "lag_sessions": 1},
+        benchmark_id="000300.SH",
+        benchmark_returns=benchmark,
+        risk_snapshots=risk_snapshots,
+    )
+
+
+def test_research_artifact_projects_versioned_queryable_fact_tables() -> None:
+    result = _backtest_result()
+
+    artifact = _build(result)
+
+    assert artifact.schema_version == RESEARCH_ARTIFACT_SCHEMA_VERSION
+    assert artifact.run.loc[0, "run_id"] == "run-20260105-a"
+    assert artifact.run.loc[0, "benchmark_alignment_policy"] == "exact_session_index"
+    assert artifact.nav["run_id"].unique().tolist() == ["run-20260105-a"]
+    assert artifact.nav["pnl_pct"].tolist() == pytest.approx(result.returns.tolist())
+    assert artifact.nav["benchmark_return"].tolist() == pytest.approx(
+        [0.0, 0.01, -0.01, 0.02]
+    )
+    assert artifact.signals.columns.tolist() == [
+        "run_id",
+        "signal_date",
+        "execution_date",
+        "asset_id",
+        "factor_score",
+        "target_weight",
+    ]
+    first_signal = artifact.signals[
+        artifact.signals["signal_date"] == result.factor_scores.index[0].date()
+    ]
+    assert first_signal.set_index("asset_id").loc["A", "factor_score"] == 2.0
+    assert first_signal.set_index("asset_id").loc["A", "target_weight"] == 1.0
+    assert first_signal["execution_date"].unique().tolist() == [
+        result.schedule.signal_to_execution.iloc[0].date()
+    ]
+    assert set(artifact.trades["side"]) == {"buy", "sell"}
+    assert artifact.trades["trade_id"].is_unique
+    assert artifact.trades["trade_id"].str.startswith("run-20260105-a:").all()
+    assert artifact.trades["signal_id"].str.startswith("run-20260105-a:signal:").all()
+    assert {"security", "cash"}.issubset(set(artifact.positions["asset_type"]))
+    assert artifact.positions.groupby("trade_date")["weight"].sum().tolist() == pytest.approx(
+        [1.0, 1.0, 1.0, 1.0]
+    )
+    assert set(artifact.attribution.columns) == {
+        "run_id",
+        "trade_date",
+        "asset_id",
+        "overnight",
+        "intraday",
+        "asset_total",
+    }
+    assert artifact.attribution_daily["residual"].abs().max() < 1e-12
+    assert artifact.risk.empty
+    assert artifact.risk.columns.tolist() == [
+        "run_id",
+        "trade_date",
+        "asset_id",
+        "weight",
+        "marginal_risk",
+        "component_risk",
+        "risk_contribution",
+        "covariance_snapshot_id",
+        "covariance_as_of_date",
+        "risk_measure",
+        "return_frequency",
+        "periods_per_year",
+    ]
+    assert artifact.performance.loc[0, "n_trades"] == len(artifact.trades)
+    assert artifact.performance.loc[0, "ir"] == pytest.approx(
+        result.benchmark_stats(pd.Series([0.0, 0.01, -0.01, 0.02], index=result.returns.index))[
+            "information_ratio"
+        ]
+    )
+    assert "sortino" in artifact.performance.columns
+
+
+def test_research_artifact_projects_annualized_risk_from_actual_positions() -> None:
+    result = _backtest_result()
+    trade_date = result.position_weights.index[-1].date()
+    covariance = pd.DataFrame(
+        [[0.0001, 0.00002], [0.00002, 0.0004]],
+        index=["A", "B"],
+        columns=["A", "B"],
+    )
+    snapshot = CovarianceSnapshot(
+        snapshot_id="cov-20260107-v1",
+        as_of_date="2026-01-07",
+        covariance=covariance,
+        return_frequency="1d",
+        periods_per_year=252,
+    )
+
+    artifact = _build(result, risk_snapshots={trade_date: snapshot})
+
+    risk = artifact.risk.set_index("asset_id")
+    expected_weights = result.position_weights.loc[pd.Timestamp(trade_date)]
+    assert artifact.schema_version == "1.1.0"
+    assert risk.index.tolist() == ["A", "B"]
+    assert risk["weight"].tolist() == pytest.approx(expected_weights.tolist())
+    assert risk["covariance_snapshot_id"].unique().tolist() == ["cov-20260107-v1"]
+    assert risk["covariance_as_of_date"].unique().tolist() == [date(2026, 1, 7)]
+    assert risk["risk_measure"].unique().tolist() == ["annualized_volatility"]
+    assert risk["return_frequency"].unique().tolist() == ["1d"]
+    assert risk["periods_per_year"].unique().tolist() == [252]
+    assert risk["component_risk"].sum() == pytest.approx((0.0004 * 252) ** 0.5)
+    assert risk["risk_contribution"].sum() == pytest.approx(1.0)
+
+
+def test_research_artifact_rejects_future_or_misaligned_risk_snapshots() -> None:
+    result = _backtest_result()
+    trade_date = result.position_weights.index[-1].date()
+    covariance = pd.DataFrame(
+        [[0.0001, 0.0], [0.0, 0.0004]],
+        index=["A", "B"],
+        columns=["A", "B"],
+    )
+
+    with pytest.raises(ValueError, match="must not be after trade date"):
+        _build(
+            result,
+            risk_snapshots={
+                trade_date: CovarianceSnapshot(
+                    snapshot_id="future-covariance",
+                    as_of_date="2026-01-09",
+                    covariance=covariance,
+                    return_frequency="1d",
+                    periods_per_year=252,
+                )
+            },
+        )
+
+    with pytest.raises(ValueError, match="same asset labels"):
+        _build(
+            result,
+            risk_snapshots={
+                trade_date: CovarianceSnapshot(
+                    snapshot_id="incomplete-universe",
+                    as_of_date="2026-01-07",
+                    covariance=covariance.loc[["B"], ["B"]],
+                    return_frequency="1d",
+                    periods_per_year=252,
+                )
+            },
+        )
+
+
+def test_research_artifact_serialization_and_hashes_are_deterministic() -> None:
+    result = _backtest_result()
+    first = _build(result, parameters={"top_k": 1, "lag_sessions": 1})
+    second = _build(result, parameters={"lag_sessions": 1, "top_k": 1})
+
+    assert first.run.loc[0, "config_hash"] == second.run.loc[0, "config_hash"]
+    assert first.content_sha256 == second.content_sha256
+    assert first.manifest() == second.manifest()
+    decoded = json.loads(first.canonical_json())
+    assert decoded["schema_version"] == RESEARCH_ARTIFACT_SCHEMA_VERSION
+    assert decoded["tables"]["nav"][0]["trade_date"] == "2026-01-05"
+
+    leaked_copy = first.nav
+    leaked_copy.loc[0, "nav"] = -999.0
+    assert first.nav.loc[0, "nav"] != -999.0
+    assert first.content_sha256 == second.content_sha256
+
+
+def test_research_artifact_requires_complete_reproducibility_identity() -> None:
+    result = _backtest_result()
+
+    with pytest.raises(ValueError, match="code_revision"):
+        build_research_run_artifact(
+            result,
+            run_id="run-1",
+            strategy_id="alpha-top1",
+            strategy_name="Alpha Top 1",
+            strategy_version="1.0.0",
+            engine_version="1.2.0",
+            code_revision="",
+            data_snapshot_id="snapshot-1",
+            calendar="CN-A",
+            timezone="Asia/Shanghai",
+            started_at="2026-01-08T10:00:00+08:00",
+            finished_at="2026-01-08T10:01:00+08:00",
+            parameters={},
+        )
+
+
+def test_research_artifact_requires_benchmark_identity_and_returns_together() -> None:
+    result = _backtest_result()
+
+    with pytest.raises(ValueError, match="benchmark_id and benchmark_returns"):
+        build_research_run_artifact(
+            result,
+            run_id="run-1",
+            strategy_id="alpha-top1",
+            strategy_name="Alpha Top 1",
+            strategy_version="1.0.0",
+            engine_version="1.2.0",
+            code_revision="3b1ad07",
+            data_snapshot_id="snapshot-1",
+            calendar="CN-A",
+            timezone="Asia/Shanghai",
+            started_at="2026-01-08T10:00:00+08:00",
+            finished_at="2026-01-08T10:01:00+08:00",
+            parameters={},
+            benchmark_id="000300.SH",
+        )

tests/test_metrics.py

@@ -14,6 +14,7 @@ from quant_engine.metrics import (
     calmar_ratio,
     max_drawdown,
     sharpe_ratio,
+    sortino_ratio,
     summary,
     win_rate,
 )
@@ -48,9 +49,22 @@ def test_zero_volatility_metrics_return_zero() -> None:
     returns = pd.Series([0.0, 0.0, 0.0])
 
     assert sharpe_ratio(returns) == 0.0
+    assert sortino_ratio(returns) == 0.0
     assert calmar_ratio(returns) == 0.0
 
 
+def test_sortino_ratio_uses_all_sessions_for_downside_deviation() -> None:
+    returns = pd.Series([0.02, -0.01, 0.0, -0.03])
+    downside = np.minimum(returns.to_numpy(), 0.0)
+    downside_deviation = np.sqrt(np.mean(np.square(downside))) * np.sqrt(
+        TRADING_DAYS_PER_YEAR
+    )
+
+    assert sortino_ratio(returns) == pytest.approx(
+        annualized_return(returns) / downside_deviation
+    )
+
+
 def test_max_drawdown_includes_loss_from_initial_capital() -> None:
     returns = pd.Series([-0.20, 0.0])
 
@@ -80,7 +94,15 @@ def test_summary_aliases_match_canonical_fields() -> None:
 
 @pytest.mark.parametrize(
     "metric",
-    [annualized_return, annualized_volatility, sharpe_ratio, max_drawdown, calmar_ratio, win_rate],
+    [
+        annualized_return,
+        annualized_volatility,
+        sharpe_ratio,
+        sortino_ratio,
+        max_drawdown,
+        calmar_ratio,
+        win_rate,
+    ],
 )
 def test_metrics_reject_non_series_input(metric) -> None:
     with pytest.raises(TypeError, match=r"expected pd\.Series"):

tests/test_risk.py

@@ -8,13 +8,164 @@ import pytest
 
 from quant_engine.risk import (
     ComponentRiskResult,
+    CovarianceSnapshot,
     component_var,
+    estimate_covariance_snapshot,
     labeled_component_risk,
     marginal_risk_contribution,
     risk_contribution,
 )
 
 
+def test_estimate_covariance_snapshot_is_complete_case_and_reproducible() -> None:
+    dates = pd.date_range("2026-01-05", periods=6, freq="B")
+    returns = pd.DataFrame(
+        {
+            "A": [0.01, 0.02, 0.03, 0.04, 0.05, 99.0],
+            "B": [0.02, 0.01, np.nan, 0.03, 0.04, -99.0],
+        },
+        index=dates,
+    )
+    as_of = dates[4]
+
+    snapshot = estimate_covariance_snapshot(
+        returns,
+        as_of_date=as_of,
+        lookback_sessions=4,
+        min_observations=3,
+        data_snapshot_id="market-returns-20260109-v1",
+        return_frequency="1d",
+        periods_per_year=252,
+    )
+
+    expected_window = returns.loc[:as_of].tail(4)
+    expected = expected_window.dropna(how="any").cov()
+    pd.testing.assert_frame_equal(snapshot.covariance, expected)
+    assert snapshot.snapshot_id.startswith("sample-cov-v1:")
+    assert snapshot.as_of_date == as_of.date()
+    assert snapshot.method == "sample"
+    assert snapshot.window_start_date == expected_window.index[0].date()
+    assert snapshot.window_end_date == as_of.date()
+    assert snapshot.observations == 3
+    assert snapshot.lookback_sessions == 4
+    assert snapshot.missing_policy == "complete_case"
+    assert snapshot.data_snapshot_id == "market-returns-20260109-v1"
+    assert len(snapshot.input_sha256) == 64
+
+    future_changed = returns.copy()
+    future_changed.loc[dates[-1], :] = [1_000_000.0, -1_000_000.0]
+    repeated = estimate_covariance_snapshot(
+        future_changed,
+        as_of_date=as_of,
+        lookback_sessions=4,
+        min_observations=3,
+        data_snapshot_id="market-returns-20260109-v1",
+        return_frequency="1d",
+        periods_per_year=252,
+    )
+    assert repeated.snapshot_id == snapshot.snapshot_id
+    pd.testing.assert_frame_equal(repeated.covariance, snapshot.covariance)
+
+
+def test_covariance_snapshot_identity_captures_data_and_estimator_contract() -> None:
+    dates = pd.date_range("2026-01-05", periods=4, freq="B")
+    returns = pd.DataFrame(
+        {"A": [0.01, 0.02, -0.01, 0.03], "B": [0.02, -0.01, 0.01, 0.04]},
+        index=dates,
+    )
+    base = estimate_covariance_snapshot(
+        returns,
+        as_of_date=dates[-1],
+        lookback_sessions=4,
+        min_observations=3,
+        data_snapshot_id="snapshot-a",
+    )
+    different_source = estimate_covariance_snapshot(
+        returns,
+        as_of_date=dates[-1],
+        lookback_sessions=4,
+        min_observations=3,
+        data_snapshot_id="snapshot-b",
+    )
+
+    assert base.snapshot_id != different_source.snapshot_id
+    assert base.covariance.equals(different_source.covariance)
+
+
+def test_estimate_covariance_snapshot_rejects_ambiguous_or_insufficient_history() -> None:
+    dates = pd.date_range("2026-01-05", periods=4, freq="B")
+    returns = pd.DataFrame(
+        {"A": [0.01, np.nan, 0.03, 0.04], "B": [0.02, 0.01, np.nan, 0.03]},
+        index=dates,
+    )
+
+    with pytest.raises(ValueError, match="complete observations"):
+        estimate_covariance_snapshot(
+            returns,
+            as_of_date=dates[-1],
+            lookback_sessions=4,
+            min_observations=3,
+            data_snapshot_id="snapshot-a",
+        )
+
+    with pytest.raises(ValueError, match="strictly increasing"):
+        estimate_covariance_snapshot(
+            returns.iloc[::-1],
+            as_of_date=dates[-1],
+            lookback_sessions=4,
+            min_observations=2,
+            data_snapshot_id="snapshot-a",
+        )
+
+
+def test_covariance_snapshot_is_validated_and_immutable_by_interface() -> None:
+    covariance = pd.DataFrame(
+        [[0.04, 0.01], [0.01, 0.09]],
+        index=["A", "B"],
+        columns=["A", "B"],
+    )
+    snapshot = CovarianceSnapshot(
+        snapshot_id="cov-20260107-v1",
+        as_of_date="2026-01-07",
+        covariance=covariance,
+        return_frequency="1d",
+        periods_per_year=252,
+    )
+
+    covariance.loc["A", "A"] = 999.0
+    leaked_copy = snapshot.covariance
+    leaked_copy.loc["B", "B"] = 999.0
+
+    assert snapshot.as_of_date == pd.Timestamp("2026-01-07").date()
+    assert snapshot.covariance.loc["A", "A"] == pytest.approx(0.04)
+    assert snapshot.covariance.loc["B", "B"] == pytest.approx(0.09)
+
+
+@pytest.mark.parametrize(
+    ("kwargs", "message"),
+    [
+        ({"snapshot_id": ""}, "snapshot_id"),
+        ({"return_frequency": ""}, "return_frequency"),
+        ({"periods_per_year": 0}, "periods_per_year"),
+    ],
+)
+def test_covariance_snapshot_rejects_incomplete_identity(
+    kwargs: dict[str, object],
+    message: str,
+) -> None:
+    values: dict[str, object] = {
+        "snapshot_id": "cov-20260107-v1",
+        "as_of_date": "2026-01-07",
+        "covariance": pd.DataFrame([[0.04]], index=["A"], columns=["A"]),
+        "return_frequency": "1d",
+        "periods_per_year": 252,
+    }
+    values.update(kwargs)
+
+    with pytest.raises((TypeError, ValueError), match=message):
+        CovarianceSnapshot(**values)
+
+
 def test_risk_contribution_sums_to_one_for_positive_portfolio_variance() -> None:
     weights = np.array([0.5, 0.5])
     covariance = np.diag([1.0, 4.0])