@@ -42,7 +42,7 @@ construction steps are documented in this repository.
 
 ## License in this repository
 
-`environment/LICENSE` grants recipients permission from `ricefan-tech` to use,
+`environment/LICENSE` grants recipients permission from `EdotEnv` to use,
 copy, run, and modify the task code and bundled data solely for internal
 evaluation, research, and demonstration. It prohibits redistribution, public
 hosting, commercial deployment, and real-money trading without separate written
 
@@ -0,0 +1,256 @@
+# Causal Feature Engineering for Quantitative Trading
+
+## Scientific Domain
+
+**Domain:** Quantitative Finance  
+**Field:** Systematic Trading  
+**Subfield:** Machine-Learned Alpha Research
+
+## Scientific Problem
+
+Quantitative researchers rarely succeed by fitting one model and accepting the
+largest public-period score. Minute-level market data are noisy, predictive
+relationships are weak, regimes change, and apparently strong signals can
+disappear after turnover, execution costs, or exposure normalization. The
+research problem is to discover causal predictive structure that remains useful
+outside the period in which it was designed.
+
+This task compresses that research loop into a fixed, reproducible environment.
+The agent is expected to inspect market data, form feature hypotheses, run
+controlled chronological experiments, compare predictive and economic
+diagnostics, reject unstable candidates, and select one strategy for hidden
+future scoring. Public metrics may conflict: a candidate can have attractive
+information coefficient but poor after-cost performance, or a high Sharpe on
+one interval with weak sign and regime stability.
+
+The final predictive estimator is deliberately fixed to an exact Elastic Net so
+that improvements must come primarily from causal feature construction,
+preprocessing, selection, regularization, and experimental judgment rather than
+from switching to a more powerful model family. Symbol identities, calendar
+coordinates, and target semantics are obscured to reduce historical lookup and
+force evidence-based research on the supplied panel.
+
+## Workflow Details
+
+The agent phase has two cooperating containers. The `main` container contains
+the public Parquet panel, task instructions, and the Python data-science stack.
+The `mcp-server` sidecar owns model fitting, artifact storage, public
+backtesting, research budgets, and final submission promotion. The agent
+interacts with that sidecar through five MCP tools:
+
+- `train_model`
+- `get_train_model_result`
+- `backtest`
+- `get_backtest_result`
+- `submit_strategy`
+
+The research workflow has four stages:
+
+1. **Public research.** Inspect the 2022–2023 panel chronologically and by
+   symbol. Form hypotheses for past-only transformations such as returns, lags,
+   rolling statistics, expanding histories, price-range structure, activity,
+   and order-flow imbalance. Keep every symbol at a timestamp in the same
+   chronological fold and fit learned preprocessing on training rows only.
+2. **Source-bound training.** Send complete Python source defining
+   `train_model(X, y)` to the asynchronous training tool. The returned fitted
+   object must be an exact `ElasticNet` or an exact sklearn `Pipeline` ending in
+   an exact `ElasticNet`. It must use `weight_std_dollar_vol` as fitting weight,
+   remain serializable with Joblib, and expose a stable inference schema.
+3. **Controlled public backtesting.** Backtest accepted models on public windows
+   strictly later than their training windows. Compare candidates on consistent
+   periods using information coefficient, error, prediction dispersion,
+   turnover, after-cost Sharpe, CAGR, and maximum drawdown. Use later slices to
+   challenge apparently strong candidates rather than optimizing one interval.
+4. **Candidate selection.** Submit the `strategy_id` from one successful
+   backtest after all asynchronous operations are idle. The sidecar promotes one
+   immutable bundle containing the selected source, fitted public artifact,
+   inference contract, strategy settings, and integrity hashes.
+
+Training and backtesting share a budget of 100 accepted research attempts. One
+training operation and one backtest may overlap, but two operations of the same
+type may not. This permits deliberate parallel research without allowing
+unbounded search.
+
+After the agent exits, Harbor collects the bundle and starts a separate
+verifier. The verifier checks bundle integrity, imports the submitted source in
+a networkless model worker, and calls `train_model` once on every public
+2022–2023 row. It serializes and reloads that newly fitted estimator, predicts
+on hidden 2024 feature matrices, audits causal invariance by perturbing future
+suffixes, and runs the hidden after-cost backtest. Hidden labels and
+scoring-only columns never enter the model worker.
+
+## Dependencies & System Requirements
+
+### Agent-facing model environment
+
+Submitted model source runs on Python 3.12 and may import only `math`,
+`statistics`, `numpy`, `pandas`, and `sklearn`. The model must fit and predict
+without file I/O, external services, or network access. Source is limited to
+20,000 UTF-8 bytes; each fit and prediction operation has a 1,800-second
+deadline. The returned estimator must serialize with Joblib and predict after
+reload in a fresh process. No GPU is available or required.
+
+The broader agent container includes the pinned Python scientific stack needed
+to inspect the Parquet data and communicate with the MCP service. Core versions
+are Python 3.12, NumPy 2.5.1, pandas 3.0.5, SciPy 1.18.0, scikit-learn 1.9.0,
+PyArrow 25.0.0, Joblib 1.5.3, and FastMCP 3.4.5.
+
+### Task deployment environment
+
+The sample targets Harbor 0.20.0 with Docker Compose and Git LFS. The task
+requests four CPU cores, 8 GiB of memory, 32 GiB of storage, and no GPU. The
+configured limits are 30 minutes for environment construction, four hours for
+the agent, and one hour for verification. The agent environment has public
+network access for general research, while submitted source executes in a
+networkless, read-only model worker with narrowly controlled imports and file
+operations.
+
+## Dataset
+
+The benchmark uses a fixed panel of historical cryptocurrency minute bars for
+four anonymized symbols. The canonical source universe is BTCUSDT, DOGEUSDT,
+ETHUSDT, and SOLUSDT, but mapped to `symbol_01` through `symbol_04`.
+Datetimes are replaced with synthetic UTC coordinates while preserving 
+ordering and elapsed durations.
+
+The chronological split is:
+
+- **Public research:** 2022-01-01 through 2023-12-31, containing 4,204,792 panel
+  rows.
+- **Hidden evaluation:** 2024-01-01 through 2024-12-31, containing 2,108,152
+  panel rows.
+
+The agent-visible public schema contains:
+
+- identity: `datetime`, `symbol`;
+- candidate predictors: `open`, `high`, `low`, `close`, `volume`,
+  `quote_asset_volume`, `number_of_trades`,
+  `taker_buy_base_asset_volume`, and `taker_buy_quote_asset_volume`;
+- fitting metadata: `weight_std_dollar_vol`;
+- supervised response: `target_horizon_1`.
+
+`target_horizon_1` is intentionally anonymized. `weight_std_dollar_vol` is fitting
+metadata rather than a predictor. Two additional columns, `tradable_return` and
+`beta_10d_fwd_1`, are retained only by the trusted backtest and verifier. They
+are never included in model fitting or prediction matrices.
+
+The canonical public and hidden Parquet files are approximately 330 MB and
+176 MB respectively. The agent image receives a compressed projection of the
+public artifact with only identity, predictor, weight, and target columns.
+Hidden data remain exclusively in the verifier environment. The dataset is
+bundled with the task; no external download is required at runtime.
+
+## Evaluation Strategy
+
+The benchmark uses a continuous outcome reward rather than a binary pass
+threshold. The primary score is annualized Sharpe on the fixed hidden 2024
+after-cost strategy returns. The same value is emitted as `primary_score`,
+`reward`, and `sharpe`. The verifier also reports hidden CAGR, maximum drawdown,
+and Pearson information coefficient so that risk-adjusted performance can be
+interpreted alongside predictive quality and downside behavior.
+
+The portfolio backtest applies a one-basis-point linear execution fee and uses
+525,600 periods per year for minute-level Sharpe annualization. The submitted
+`max_gross_exposure` must be finite and strictly positive. Public and hidden
+backtests use the same task-owned portfolio construction and accounting code.
+The absence of a binary cutoff is intentional: the scalar reward supports
+ranking, reinforcement learning, and comparison across research policies.
+
+Before scoring, the verifier enforces the following safeguards:
+
+1. **Submission integrity.** Exactly one publicly accepted bundle must exist.
+   Source and artifact hashes must match the manifest, and bundle paths may not
+   escape the submission directory.
+2. **Estimator family.** `train_model` must return either an exact
+   `sklearn.linear_model.ElasticNet` or an exact sklearn `Pipeline` whose final
+   step is an exact `ElasticNet`. Subclasses and other final estimators are
+   rejected.
+3. **Executable contract.** Submitted source is size- and import-restricted.
+   The estimator must be fitted, expose unique ordered string
+   `feature_names_in_`, serialize with Joblib, reload in a fresh process, and
+   return finite predictions in the original row order.
+4. **Full-public refit.** Hidden scoring uses a new estimator fitted by calling
+   the submitted `train_model` source once on the complete public 2022–2023
+   panel. The estimator fitted during public experimentation is not scored.
+5. **Causal audit.** Three hidden future-suffix probes materially perturb later
+   feature rows. Predictions for earlier rows must remain invariant within
+   strict numerical tolerances.
+6. **Hidden isolation.** The model worker receives feature matrices but never
+   hidden targets, tradable returns, market betas, or task-owned portfolio
+   state.
+
+A missing final submission receives `-100` for every reward field. Contract
+violations are rejected before hidden economic metrics are exposed.
+
+## Complexity
+
+The computational primitives are standard, but the research problem is
+difficult because useful minute-level signals are weak and easy to overfit.
+Every candidate must survive chronological separation, cross-symbol analysis,
+causal batch execution, regularization, portfolio normalization, turnover, and
+costs. Fixing the final estimator to Elastic Net removes model-family search as
+an escape hatch: the agent's advantage must come from better hypotheses,
+features, controls, and candidate selection.
+
+The public panel contains more than 4.2 million rows, and accepted training and
+backtesting calls share a 100-attempt budget. A strong policy must allocate that
+budget among data diagnostics, feature ablations, regularization comparisons,
+chronological confirmation, and exposure experiments while asynchronous jobs
+are running.
+
+Observed task runs demonstrate a broad outcome range. The current reference
+strategy receives hidden Sharpe `0.3723068959` with hidden Pearson IC
+`0.0196662380`. Recorded end-to-end and candidate-validation artifacts include
+missing or invalid submissions at `-100`, valid strategies with negative Sharpe,
+degenerate historical baselines at `0`, and positive candidates between
+approximately `0.15` and `0.39`. These are continuous outcomes rather than
+pass/fail labels.
+
+A representative Harbor run completed in 20 minutes 53 seconds end to end:
+14 seconds for environment setup, 28 seconds for agent setup, 16 minutes
+40 seconds for agent execution, and 3 minutes 11 seconds for verification. The
+configured limits leave room for substantially deeper research than the
+reference policy performs.
+
+## References & Resources
+
+The sample is self-contained. Its implementation contracts and supporting
+evidence are documented in:
+
+- `instruction.md` — agent-facing objective, data split, MCP workflow, and
+  full-public refit disclosure;
+- `environment/agent/context/research.md` — causal research and chronological
+  validation guidance;
+- `environment/agent/context/submission.md` — exact estimator, source,
+  serialization, and prediction contract;
+- `environment/agent/context/backtesting.md` — public candidate-comparison and
+  strategy-selection guidance;
+- `DATA_PROVENANCE.md` — fixed artifact hashes, source universe,
+  anonymization, split construction, and license record;
+- `environment/mcp-server/task_config.yaml` — task-owned data, backtest, cost,
+  execution, reward, and budget configuration;
+- `environment/mcp-server/feature_engineering/evaluation/verifier.py` and
+  `scoring/official.py` — independent verification and hidden scoring flow;
+- `solution/solve.py` — reference MCP-only agent;
+- `environment/mcp-server/tests/` — causal, portfolio, runtime, release,
+  end-to-end, and verifier behavior contracts.
+
+The complete local Python verification command is:
+
+```sh
+uv run --project environment/mcp-server test
+```
+
+## Additional Information
+
+This document describes the OpenAI-facing sample as an outcome-scored research
+environment rather than a binary programming exercise. Its hidden Sharpe reward
+is intended for continuous evaluation, policy comparison, and reinforcement
+learning. The fixed Elastic Net family is an experimental control that focuses
+the task on feature engineering and quantitative-research decisions.
+
+## Author Information
+
+**Contributing organization:** EdotEnv  
+**Contribution:** Task design, implementation, data packaging, verifier,
+reference policy, and calibration.
@@ -1,7 +1,7 @@
 External Evaluation Sample License
 Version 1.0, July 2026
 
-Copyright (c) 2026 ricefan-tech
+Copyright (c) 2026 EdotEnv
 
 Permission is granted, without fee, to recipients of this sample to use, copy,
 run, and modify the task code and bundled evaluation data solely for internal
 
@@ -4,7 +4,7 @@ artifacts = [{ source = "/app/submission", service = "mcp-server" }]
 [task]
 name = "mmcollab-dotcom/feature-engineering"
 description = "Engineer causal features and submit one calibrated quantitative-ML trading strategy on a fixed public/hidden crypto split."
-authors = [{ name = "ricefan-tech", email = "r.wang97@yahoo.de" }]
+authors = [{ name = "EdotEnv", email = "michael.zhang@edotenv.com" }]
 keywords = ["machine-learning-research", "quantitative-research", "quantitative-trading", "feature-engineering", "harbor"]
 
 [metadata]