Establish public agent orientation policy

.gitignore

@@ -17,3 +17,5 @@ ENV/
 .hypothesis/
 CLAUDE.md
 .claude/
+.codex/
+.workspace/

AGENTS.md

@@ -0,0 +1,52 @@
+# ml4t-models
+
+Finance-specific models for latent-factor estimation, factor-premium forecasting, stochastic
+discount factors, direct asset prediction, and end-to-end portfolio learning. The package separates
+model estimation from forecasting and mapping so outputs can move explicitly into ML4T Backtest and
+Diagnostic workflows.
+
+## Public entry points
+
+```python
+from ml4t.models import CrossSectionBatch, IPCAModel, IPCAConfig
+from ml4t.models import LatentFactorForecastPipeline, ExpandingMeanFactorForecaster
+```
+
+Use exports from `ml4t.models`. Torch-backed models load lazily and require the `deep` extra;
+Polars and `ml4t-specs` integrations require the `integration` extra.
+
+## Source map
+
+| Path | Responsibility |
+|---|---|
+| `src/ml4t/models/api.py` | Protocols shared across model families |
+| `src/ml4t/models/types.py` | Finance-native batches, fitted state, forecasts, and weights |
+| `src/ml4t/models/latent_factors/` | PCA, risk-premium PCA, IPCA, and conditional autoencoders |
+| `src/ml4t/models/forecasters/` | Ex ante factor-premium forecasting |
+| `src/ml4t/models/mappers/` | Mapping structural outputs to asset forecasts |
+| `src/ml4t/models/stochastic_discount_factor/` | Weight-native SDF estimation |
+| `src/ml4t/models/asset_prediction/` | Direct asset prediction models |
+| `src/ml4t/models/portfolio/` | Linear, recurrent, and deep portfolio models |
+| `src/ml4t/models/integration/` | Data, backtest, diagnostic, and artifact boundaries |
+| `scripts/ci/` | Release, coverage, performance, and hardware qualification |
+
+## Modeling constraints
+
+- Preserve temporal and cross-sectional identities in every batch and output object.
+- Keep structural estimation, forecasting, asset mapping, and allocation as explicit stages.
+- Make stochastic training reproducible and checkpoint semantics observable.
+- Keep optional heavy dependencies out of the core import path.
+- Record model-family assumptions in the relevant reference page and test financial invariants, not
+  only array shapes.
+
+## Quality commands
+
+```bash
+uv run ruff check src/ tests/ examples/ scripts/
+uv run ruff format --check src/ tests/ examples/ scripts/
+uv run ty check
+uv run pytest tests/ -q --cov-report=json:coverage.json
+uv run python scripts/ci/check_coverage.py coverage.json
+uv run mkdocs build --strict
+uv build
+```