Lazy experiment lifecycle: configure → optional prior checks → fit()

ARCHITECTURE.md

@@ -66,7 +66,15 @@ The optional third backend, `PyMCForecastModel` (`causalpy/pymc_forecast_models.
 
 ## Experiment Lifecycle
 
-Instantiation fits eagerly in `__init__`: `_build_design_matrices()` → `_prepare_data()` → `algorithm()`. There is no separate `.fit()` on the experiment. Each subclass's public `plot(*, ...)` delegates to `_render_plot()`, which calls the subclass's backend-agnostic `_plot()`. Uncertainty rendering keys on data properties of the canonical prediction container (`has_posterior_draws()`), not on backend identity. `effect_summary()` returns `EffectSummary(table, text)` using helpers from `causalpy.reporting`.
+Construction is lazy: `__init__` runs validation and deterministic preprocessing only (`_build_design_matrices()` → `_prepare_data()`); nothing is sampled. The lifecycle is `configure → optional prior checks → fit()`:
+
+- **`build()`** — public, idempotent, auto-called by both samplers. Merges data-driven priors (defaults → data-derived → user) and constructs the PyMC graph so the spec is inspectable (`pm.model_to_graphviz(exp.model)`, `exp.model.basic_RVs`) before any compute is spent.
+- **`sample_prior_predictive(**kwargs)`** — optional prior phase; populates `exp.prior_result`. Raises `PriorPredictiveNotSupportedException` on backends without a prior phase (sklearn, pymc-forecast, IV, state space).
+- **`fit(**kwargs)`** — posterior phase; runs NUTS + posterior predictive first (forward-sampling machinery mutates shared data nodes, so posterior must sample on freshly armed data), then fills the absent prior phase so `idata` matches the historical eager output. Returns `Self`, so call sites migrate with one appended token: `cp.InterruptedTimeSeries(...).fit()`.
+
+Draw-derived state lives in per-experiment bundles (`causalpy/experiments/_results.py`) exposed through two raising properties: `exp.result` (posterior group) and `exp.prior_result` (prior group). Nothing derived from draws is assigned to the experiment itself; `is_fitted` / `has_prior_predictive` are predicates over those two slots. Read methods take keyword-only `group: Literal["prior", "posterior"]`; the guard and group→bundle resolution live once in `_render_plot()` / `_resolve_group()`. Missing groups raise `GroupNotSampleedException` naming the call to make; both exceptions are re-exported from `causalpy`. Re-sampling overwrites its own group only (`fit()` again replaces posterior and warns; prior state survives). Assigning a new model — also the documented prior-revision path instead of a `set_priors()` — resets everything, because graph identity *is* the model instance.
+
+Each subclass's public `plot(*, ...)` delegates to `_render_plot()`, which calls the subclass's backend-agnostic `_plot(bundle=..., group=...)`. Uncertainty rendering keys on data properties of the canonical prediction container (`has_posterior_draws()`), not on backend identity. `plot(group="prior")` renders a reduced panel set (counterfactual vs observed only). `effect_summary(group=...)` returns `EffectSummary(table, text)` using helpers from `causalpy.reporting`; prior-group prose reads as a plausibility check, not a causal claim.
 
 ## Experiment Inventory
 
@@ -98,21 +106,19 @@ Instantiation fits eagerly in `__init__`: `_build_design_matrices()` → `_prepa
 
 | Topic | Detail |
 |-------|--------|
+| **Lazy fitting** | `__init__` never samples. Explicit `fit()` / `sample_prior_predictive()` run the phases; see Experiment Lifecycle. |
 | **Formulas** | Patsy `dmatrices()` for design matrices; `build_design_matrices()` for counterfactual prediction. Bare datetime predictors are encoded as continuous elapsed days from the fitted origin; use `C(date)` for date fixed effects. `PiecewiseITS` uses `step()`/`ramp()` stateful transforms. |
 | **obs_ind** | All experiments set `data.index.name = "obs_ind"`. Canonical xarray/PyMC dimension name. |
 | **treated_units always 2D** | Even single-unit experiments use `treated_units=["unit_0"]`. Never pass 1D y to PyMC. |
 | **Impact uses mu, not y_hat** | The adapter's `predict()` extracts posterior `mu` (conditional expected outcome in observed units), not `y_hat` (with observation noise); impact is `y - predict(X)`. For GLMs, `mu` must be inverse-linked before impact; see `docs/source/knowledgebase/prediction-contract.md`. |
 | **Intercept handling** | Patsy includes intercept by default. sklearn models must use `fit_intercept=False`. |
-| **Eager fitting** | MCMC runs during `__init__`. No lazy `.fit()` on the experiment. |
-| **HDI_PROB** | Project default is 0.94 (ArviZ default), not 0.95. |
 | **create_causalpy_compatible_class** | Applied during `make_model_adapter()` for sklearn backends; clones the user instance before patching. |
 
 ## Adding New Code
 
 Copy the closest existing experiment or model and follow the `BaseExperiment` contract:
 
-- Declare `supports_ols` / `supports_bayes` (and `supports_pymc_forecast` to opt into the optional pymc-forecast backend); implement a single backend-agnostic `_plot()` (and an explicit `get_plot_data(*, ...)` only where that view is supported) that consumes the canonical prediction container, keying uncertainty rendering on `has_posterior_draws()` rather than backend identity
-- `algorithm()` with the fit/predict/impact flow; every concrete experiment declares its own explicit `effect_summary(...)` contract, using helpers in `causalpy.reporting` where that summary is implemented
+- Implement `_fit_inputs()` returning the `(X, y, coords)` handed to the backend at build time, and `_finalize(group)` computing the experiment's result bundle from the group's draws; every concrete experiment declares its own explicit `effect_summary(*, group=..., ...)` contract, using helpers in `causalpy.reporting` where that summary is implemented
 - Public APIs expose explicit named parameters rather than bare `*args` / `**kwargs`; use keyword-only optional controls for public plotting and plot-data APIs (enforced by `causalpy/tests/test_public_signatures.py` and surveyed by `scripts/audit_public_signatures.py`). A genuine dynamic or third-party forwarder requires an `Other Parameters` contract and a narrow structural-test exemption. For experiments without a unified plot view (e.g. `InversePropensityWeighting`, `InstrumentalVariable`), declare an explicit `plot()` stub that raises `NotImplementedError`. For `hdi_prob` defaults, use ``Defaults to :data:`~causalpy.constants.HDI_PROB` (currently 0.94).`` in the docstring.
 - Raise `FormulaException`, `DataException`, or `BadIndexException` from `causalpy.custom_exceptions` for formula, data, and index errors
 - Avoid backwards-compat shims for APIs introduced in the same PR

causalpy/__init__.py

@@ -22,6 +22,13 @@
 from causalpy.transforms import ramp, step
 from causalpy.version import __version__
 
+from .custom_exceptions import (
+    BadIndexException,
+    DataException,
+    FormulaException,
+    GroupNotSampleedException,
+    PriorPredictiveNotSupportedException,
+)
 from .data import load_data
 from .experiments.diff_in_diff import DifferenceInDifferences
 from .experiments.instrumental_variable import InstrumentalVariable
@@ -49,14 +56,18 @@
 
 __all__ = [
     "__version__",
+    "BadIndexException",
     "checks",
+    "DataException",
     "create_causalpy_compatible_class",
     "DifferenceInDifferences",
     "EstimateEffect",
     "EffectSummary",
     "extract_lift_for_mmm",
+    "FormulaException",
     "GenerateReport",
     "InstrumentalVariable",
+    "GroupNotSampleedException",
     "InterruptedTimeSeries",
     "InversePropensityWeighting",
     "load_data",
@@ -67,6 +78,7 @@
     "PanelRegression",
     "plot_correlations",
     "PrePostNEGD",
+    "PriorPredictiveNotSupportedException",
     "pymc_forecast_models",
     "pymc_models",
     "ramp",

causalpy/checks/bandwidth.py

@@ -109,7 +109,7 @@ def run(
                 kw["model"] = clone_model(kw["model"])
 
             try:
-                alt_experiment = method(context.data, **kw)
+                alt_experiment = method(context.data, **kw).fit()
                 summary = alt_experiment.effect_summary()
                 row: dict[str, Any] = {"bandwidth": bw}
                 if summary.table is not None and not summary.table.empty:

causalpy/checks/leave_one_out.py

@@ -104,7 +104,7 @@ def run(
                 kw["model"] = clone_model(kw["model"])
 
             try:
-                alt_experiment = method(context.data, **kw)
+                alt_experiment = method(context.data, **kw).fit()
                 summary = alt_experiment.effect_summary()
                 row: dict[str, Any] = {"dropped_unit": dropped}
                 if summary.table is not None and not summary.table.empty:

causalpy/checks/outcome_falsification.py

@@ -192,7 +192,7 @@ def _build_experiment(
         if "model" in kwargs and kwargs["model"] is not None:
             kwargs["model"] = clone_model(kwargs["model"])
 
-        return method(context.data, **kwargs)
+        return method(context.data, **kwargs).fit()
 
     @staticmethod
     def _extract_effect_stats(

causalpy/checks/placebo_in_space.py

@@ -120,7 +120,7 @@ def run(
                 kw["model"] = clone_model(kw["model"])
 
             try:
-                alt_experiment = method(context.data, **kw)
+                alt_experiment = method(context.data, **kw).fit()
                 summary = alt_experiment.effect_summary()
                 row: dict[str, Any] = {"placebo_treated": placebo_treated}
                 if summary.table is not None and not summary.table.empty:

causalpy/checks/placebo_in_time.py

@@ -439,7 +439,7 @@ def validate(self, experiment: BaseExperiment) -> None:
                 f"explicit treatment time."
             )
         # Any InferenceData-capable backend (PyMCModel or PyMCForecastModel)
-        # yields the draw-level post_impact this check consumes.
+        # yields the draw-level posterior impact this check consumes.
         backend = getattr(experiment, "_model_backend", None)
         if backend is None or not backend.supports_idata:
             raise TypeError(
@@ -518,7 +518,7 @@ def _factory(
                 kw["model"] = self._clone_model_for_fold(
                     model_template, fold_random_seed
                 )
-            return method(data, **kw)
+            return method(data, **kw).fit()
 
         return _factory
 
@@ -835,7 +835,7 @@ def _extract_cumulative_impact(experiment: BaseExperiment) -> xr.DataArray:
         obtained by summing over ``obs_ind`` and stacking
         ``(chain, draw)``.
         """
-        post_impact = experiment.post_impact  # type: ignore[attr-defined]
+        post_impact = experiment.result.impact_post
 
         if "treated_units" in post_impact.dims:
             post_impact = post_impact.isel(treated_units=0)

causalpy/checks/pre_treatment_placebo.py

@@ -32,7 +32,8 @@ class PreTreatmentPlaceboCheck:
     """Check that pre-treatment event-study estimates are near zero.
 
     Wraps the pre-treatment placebo effects already computed by
-    ``StaggeredDifferenceInDifferences`` in ``att_event_time_``.
+    ``StaggeredDifferenceInDifferences`` in its result bundle
+    (``result.att_event_time``).
 
     Parameters
     ----------
@@ -61,10 +62,10 @@ def validate(self, experiment: BaseExperiment) -> None:
                 "PreTreatmentPlaceboCheck requires a "
                 "StaggeredDifferenceInDifferences experiment."
             )
-        if not hasattr(experiment, "att_event_time_"):
+        if not experiment.is_fitted:
             raise ValueError(
-                "Experiment does not have att_event_time_. "
-                "Ensure the experiment has been fitted."
+                "Experiment has not been fitted. "
+                "Ensure fit() has been called before running this check."
             )
 
     def run(
@@ -82,7 +83,7 @@ def run(
             Pipeline context (unused; required by the check protocol).
         """
         sdid = experiment
-        att_et = sdid.att_event_time_  # type: ignore[attr-defined]
+        att_et = sdid.result.att_event_time
 
         pre_treatment = att_et[att_et["event_time"] < 0].copy()
 

causalpy/checks/prior_sensitivity.py

@@ -139,7 +139,7 @@ def run(
             model = clone_model(alt["model"])
             logger.info("PriorSensitivity: fitting with '%s'", name)
 
-            alt_experiment = method(context.data, model=model, **base_kwargs)
+            alt_experiment = method(context.data, model=model, **base_kwargs).fit()
 
             try:
                 summary = alt_experiment.effect_summary()

causalpy/custom_exceptions.py

@@ -58,3 +58,46 @@ class DataException(Exception):
     def __init__(self, message: str):
         super().__init__(message)
         self.message = message
+
+
+class GroupNotSampleedException(Exception):
+    """Raised when a read method requests a draw group that has not been sampled.
+
+    Under the lazy lifecycle an experiment holds no draws until
+    :meth:`~causalpy.experiments.base.BaseExperiment.fit` (posterior group) or
+    :meth:`~causalpy.experiments.base.BaseExperiment.sample_prior_predictive`
+    (prior group) is called. This exception carries the missing group and the
+    call that would populate it so the fix is actionable.
+
+    Parameters
+    ----------
+    message : str
+        Human-readable description naming the missing group and the call to
+        make.
+    group : str, optional
+        The draw group that was requested but not sampled.
+    """
+
+    def __init__(self, message: str, group: str | None = None):
+        super().__init__(message)
+        self.message = message
+        self.group = group
+
+
+class PriorPredictiveNotSupportedException(Exception):
+    """Raised when prior predictive sampling is requested of a backend that cannot do it.
+
+    Prior-phase support is a property of the model backend. OLS/sklearn
+    models, ``PyMCForecastModel``, and the PyMC state-space and instrumental
+    variable models do not expose a prior predictive phase.
+
+    Parameters
+    ----------
+    message : str
+        Human-readable description naming the model class that lacks the
+        capability.
+    """
+
+    def __init__(self, message: str):
+        super().__init__(message)
+        self.message = message

causalpy/experiments/_results.py

@@ -0,0 +1,136 @@
+#   Copyright 2022 - 2026 The PyMC Labs Developers
+#
+#   Licensed under the Apache License, Version 2.0 (the "License");
+#   you may not use this file except in compliance with the License.
+#   You may obtain a copy of the License at
+#
+#       http://www.apache.org/licenses/LICENSE-2.0
+#
+#   Unless required by applicable law or agreed to in writing, software
+#   distributed under the License is distributed on an "AS IS" BASIS,
+#   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#   See the License for the specific language governing permissions and
+#   limitations under the License.
+"""Per-experiment result bundles for the lazy experiment lifecycle.
+
+Every bundle is *fully populated* whenever it exists: an experiment either has
+no bundle yet (the corresponding lifecycle verb has not run) or a complete one.
+Nothing derived from model draws lives directly on the experiment object; the
+two public slots ``experiment.result`` (posterior group) and
+``experiment.prior_result`` (prior group) hold these bundles and back the
+``is_fitted`` / ``has_prior_predictive`` state predicates.
+
+All prediction/impact fields are typed :class:`xarray.DataArray` on the
+canonical ``("chain", "draw", "obs_ind"[, ...])`` dimensions produced by
+:class:`~causalpy.experiments.model_adapter.ModelAdapter.predict`.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+import pandas as pd
+import xarray as xr
+
+__all__ = [
+    "CausalResult",
+    "CoefficientResult",
+    "DiscontinuityResult",
+    "GroupComparisonScenario",
+    "KinkResult",
+    "StaggeredDifferenceInDifferencesResult",
+    "SyntheticDifferenceInDifferencesResult",
+]
+
+
+@dataclass(frozen=True)
+class CausalResult:
+    """Result bundle for predict-contrast experiments.
+
+    Used by :class:`~causalpy.experiments.interrupted_time_series.InterruptedTimeSeries`,
+    :class:`~causalpy.experiments.synthetic_control.SyntheticControl`, and
+    :class:`~causalpy.experiments.piecewise_its.PiecewiseITS`.
+
+    For ``PiecewiseITS`` ``predictions_pre`` carries the fitted expectation over
+    the full observation window and ``predictions_post`` / ``impact_post`` /
+    ``impact_post_cumulative`` carry the post-first-interruption slices consumed
+    by the reporting helpers.
+    """
+
+    predictions_pre: xr.DataArray
+    predictions_post: xr.DataArray
+    impact_pre: xr.DataArray
+    impact_post: xr.DataArray
+    impact_post_cumulative: xr.DataArray
+    score: pd.Series | None = None
+
+
+@dataclass(frozen=True)
+class SyntheticDifferenceInDifferencesResult(CausalResult):
+    """Result bundle for :class:`~causalpy.experiments.synthetic_difference_in_differences.SyntheticDifferenceInDifferences`.
+
+    The ``CausalResult`` prediction/impact fields are reconstructed from the
+    synthetic-control imputation; ``tau_posterior`` carries the analytic
+    double-difference treatment-effect draws with dimensions
+    ``("chain", "draw")``.
+    """
+
+    tau_posterior: xr.DataArray = field(kw_only=True)
+
+
+@dataclass(frozen=True)
+class GroupComparisonScenario:
+    """One scenario plotted or summarized by a group-comparison experiment."""
+
+    inputs: pd.DataFrame
+    prediction: xr.DataArray
+
+
+@dataclass(frozen=True)
+class CoefficientResult:
+    """Result bundle for coefficient-contrast experiments (DiD, PrePostNEGD).
+
+    ``causal_impact`` holds draws of the treatment-effect coefficient (or its
+    algebraically-equivalent prediction contrast) with canonical coefficient
+    dimensions.
+    """
+
+    causal_impact: xr.DataArray
+    scenario_control: GroupComparisonScenario
+    scenario_treated: GroupComparisonScenario
+    scenario_counterfactual: GroupComparisonScenario | None = None
+    score: pd.Series | None = None
+
+
+@dataclass(frozen=True)
+class DiscontinuityResult:
+    """Result bundle for :class:`~causalpy.experiments.regression_discontinuity.RegressionDiscontinuity`."""
+
+    predictions: xr.DataArray
+    discontinuity_at_threshold: xr.DataArray
+    score: pd.Series | None = None
+
+
+@dataclass(frozen=True)
+class KinkResult:
+    """Result bundle for :class:`~causalpy.experiments.regression_kink.RegressionKink`."""
+
+    predictions: xr.DataArray
+    gradient_change: xr.DataArray
+    score: pd.Series | None = None
+
+
+@dataclass(frozen=True)
+class StaggeredDifferenceInDifferencesResult:
+    """Result bundle for :class:`~causalpy.experiments.staggered_did.StaggeredDifferenceInDifferences`.
+
+    ``att_group_time`` and ``att_event_time`` are aggregated ATT tables;
+    ``y_pred`` retains the raw counterfactual draws so placebos and alternate
+    HDI levels can re-derive effects without resampling.
+    """
+
+    att_group_time: pd.DataFrame
+    att_event_time: pd.DataFrame
+    y_pred: xr.DataArray
+    hdi_prob: float
+    score: pd.Series | None = None