Release 2.0.0

.github/workflows/run_pytest.yml

@@ -33,37 +33,53 @@ jobs:
 
     - name: Install dependencies
       run: |
-        # --all-extras is REQUIRED, not tidiness. `views-frames` and `pandas` are
-        # optional extras, and tests/test_metric_frame.py and
-        # tests/test_evaluation_report.py both guard with a MODULE-LEVEL
-        # pytest.importorskip. Without the extras those files are skipped whole and
+        # --all-extras is REQUIRED, not tidiness. `views-frames` is an optional
+        # extra, and tests/test_metric_frame.py guards with a MODULE-LEVEL
+        # pytest.importorskip. Without the extra that file is skipped whole and
         # CI reports green without ever touching the MetricFrame cross-repo contract
         # (ADR-022 §1). Guarded by tests/test_falsification_ci_coverage_gap.py.
         poetry install --all-extras
 
     # Nothing else ever exercises "works without the dev group": the local env and the
-    # steps below always carry scikit-learn. `poetry sync --only main` removes the dev
-    # group (scikit-learn among it) and keeps the extras; the probe then asserts that
-    # scikit-learn is genuinely ABSENT — not merely unloaded, which would be vacuous —
-    # and that importing the package does not load pandas, which the extras install
-    # (ADR-011; C-05). The reinstall restores the dev group; the verify step below
-    # proves the extras are present, and the pytest step needs the dev group or it
-    # cannot collect.
+    # steps below always carry scikit-learn and pandas. `poetry sync --only main`
+    # removes the dev group (both among it) and keeps the extras; the probe then
+    # asserts that scikit-learn and pandas are genuinely ABSENT — not merely unloaded,
+    # which would be vacuous — and then evaluates a frame and emits a MetricFrame,
+    # because the emit path imports everything it needs at call time and an
+    # import-only probe would not execute a line of it (ADR-011; C-05, C-40). The
+    # "does not LOAD pandas when it is present" half is the purity guard in
+    # tests/test_metric_calculators.py, which fails rather than skips if pandas is
+    # missing from the dev group. The reinstall restores the dev group; the verify
+    # step below proves the extra is present, and the pytest step needs the dev group
+    # or it cannot collect.
     - name: Prove the runtime-only install works
       run: |
         set -e
         poetry sync --only main --all-extras
-        poetry run python -c "import sys, importlib.util as u, views_evaluation, views_evaluation.evaluation.metric_frame; assert u.find_spec('sklearn') is None, 'scikit-learn still installed'; assert 'pandas' not in sys.modules, 'import loaded pandas'"
+        poetry run python - <<'PY'
+        import importlib.util as u
+        import numpy as np
+        from views_evaluation import EvaluationFrame, NativeEvaluator
+        assert u.find_spec('sklearn') is None, 'scikit-learn still installed'
+        assert u.find_spec('pandas') is None, 'pandas still installed'
+        ids = {k: np.array(v) for k, v in {'time': [1, 1], 'unit': [1, 2], 'origin': [0, 0], 'step': [1, 1]}.items()}
+        cfg = {'steps': [1], 'classification_targets': ['t'], 'classification_point_metrics': ['AP', 'Brier_cls_point']}
+        report = NativeEvaluator(cfg).evaluate(EvaluationFrame(np.array([0., 1.]), np.array([[0.2], [0.8]]), ids, metadata={'target': 't'}))
+        mf = report.to_metric_frame(model_id='ci', run_type='test', timestamp=1758153600, evaluation_timestamp='2026-09-18T00:00:00')
+        assert isinstance(mf.metadata.provenance.timestamp, int), 'timestamp is int in the cross-repo contract (C-47)'
+        assert mf.n_rows == len(cfg['classification_point_metrics']) * 2 * 3, mf.n_rows  # metrics x (group + mean) x schemas
+        print('runtime-only install: evaluate + to_metric_frame OK, rows =', mf.n_rows)
+        PY
         poetry install --all-extras
 
-    # Proves the extras ARRIVED, not merely that --all-extras was requested.
+    # Proves the extra ARRIVED, not merely that --all-extras was requested.
     # `poetry install` exits 0 even if it resolves without an optional package, and
-    # every downstream guard is a module-level pytest.importorskip that turns absence
+    # the downstream guard is a module-level pytest.importorskip that turns absence
     # into silence. Without this step a partial install silently shrinks the suite
     # and CI still reports green. Guarded by
     # tests/test_falsification_extras_actually_installed.py.
     - name: Verify optional extras are installed
-      run: poetry run python -c "import views_frames, pandas"
+      run: poetry run python -c "import views_frames"
 
     - name: Run tests
       run: |

CHANGELOG.md

@@ -21,6 +21,97 @@ _Nothing yet._
 
 ---
 
+## [2.0.0] — 2026-09-18
+
+**MAJOR** (ADR-022 §5): this release removes the surface 1.1.0 deprecated, one release
+cycle later as §2 requires. It is the end of the pandas-free epic (#66): `import
+views_evaluation` loads neither pandas nor scikit-learn, no public method returns a
+pandas object, and the runtime dependencies are numpy, scipy and the optional `frames`
+extra. `MetricFrame` (on disk and in memory), `to_dict()` and `get_schema_results()` are
+unchanged from 1.1.0: every test on them passes unmodified. Stories S7 (#63) and S8
+(#72); tracking #73.
+
+### Removed
+
+- **Removed: `EvaluationReport.to_dataframe()`** and, with it, `BaseEvaluationMetrics.evaluation_dict_to_dataframe()`,
+  the three `make_*_evaluation_dict` factories (unused since 0.4.0 removed `EvaluationManager`,
+  their only caller; never documented, never warned on — they go as private cleanup), and
+  the **`dataframe` extra** (`pip install views-evaluation[dataframe]` is no longer a thing
+  to install). Breaking. The `DeprecationWarning` for the removed `to_dataframe()` and
+  `evaluation_dict_to_dataframe()` shipped in 1.1.0 (see that section's *Deprecated*
+  entry). views-pipeline-core, the one production caller: its `development` branch stopped
+  calling it (their #513, observed 2026-09-16), but as of 2026-09-18 that is in no release
+  of theirs — their released 3.2.0 still calls the removed `to_dataframe()`, and their
+  `^1.0.0` pin excludes this release until they move (their #515, after views-reporting).
+  **Migration:**
+  build the table in the caller —
+  `pd.DataFrame.from_dict(report.to_dict()['schemas'][schema], orient='index')`. Same
+  values; not byte-identical, as the 1.1.0 entry spells out: the old method dropped any
+  column that was `nan` in every group (register **C-40**, closed by this removal —
+  `to_dict()` and `to_metric_frame()` now are the only export paths and agree on the
+  metric set by construction), ordered columns by dataclass field, and kept an empty
+  group as a `nan` row. `schema='raw'` callers use `to_dict()['schemas']`.
+  `get_schema_results()` and the four typed dataclasses' fields are unchanged (the
+  dataclasses lose the four inherited helpers named above and nothing else).
+- `examples/benchmark_probabilistic_scaling.py`, a relic that never imported this package.
+
+### Changed
+
+- **Nothing under `views_evaluation/` imports pandas any more**, lazily or otherwise; the
+  package's runtime dependencies are numpy, scipy and the optional `frames` extra. pandas
+  moves to the dev group so the import-purity guard, which fails rather than skips when
+  pandas is absent, keeps something to prove. The CI runtime-only install step now asserts
+  pandas is *absent* alongside scikit-learn.
+
+### Release checklist (ADR-022 §7)
+
+- [x] **Does this release do anything rule 2 governs — remove an `__all__` symbol, remove
+  a supported config key, narrow an accepted input, or change a raised exception type?
+  If so, did a `DeprecationWarning` ship at least one release ago?** It removes a
+  documented method of an `__all__` symbol (the removed `EvaluationReport.to_dataframe()`)
+  and the helper behind it. Yes: the `DeprecationWarning` naming `to_dict()` as the replacement
+  shipped in **1.1.0** (published 2026-09-18; see its *Deprecated* entry), one full release
+  cycle before this one, and both release notes record it (§2 rule 3). Said plainly: §2
+  counts cycles in releases, not days, and both releases were cut on the same date, so a
+  consumer who never installed 1.1.0 never saw the warning at runtime. The release notes
+  and the pre-tag notifications below are that consumer's notice; the maintainer chose
+  two same-day releases over one so the warning exists in a published artifact and the
+  §2 record is honest rather than skipped. The three
+  `make_*_evaluation_dict` factories were never documented behaviour of any `__all__`
+  symbol (§1) and had no caller since 0.4.0; they go as private cleanup and are listed
+  above for completeness, not because rule 2 reached them. No config key is removed, no
+  accepted input narrowed, no raised exception type changed.
+- [x] **Does this release make previously-accepted input fail? If so, are the release
+  notes explicit, and have known consumers been notified?** Yes — a call to
+  `to_dataframe()` (any schema, including `'raw'`) now raises `AttributeError`, and
+  `pip install views-evaluation[dataframe]` no longer selects an extra. The migration is
+  above. Known consumers, measured 2026-09-18: **views-pipeline-core** — their released
+  3.2.0 still calls the removed `report.to_dataframe()` in `managers/evaluation/stage.py`;
+  their `development` branch does not (their #513) and keeps a deliberate tripwire test
+  that calls it; their pin `^1.0.0` excludes this release until they move (their #515).
+  **views-reporting** — `main` and `development` pin `>=1.0.0,<3.0.0`; their released
+  v0.3.3 still pins `<2.0.0`, so an environment built from that tag excludes this
+  release too; they call nothing that changed. Both were
+  notified **before the tag** on 2026-09-18: views-platform/views-pipeline-core#515
+  (comment 5733334619) and views-platform/views-reporting#289 (comment 5733334868);
+  the comments are the committed artifact.
+- [x] **Does the version bump match the change class (rule 5)?** `1.1.0` → `2.0.0`,
+  MAJOR. Removing a documented method is breaking under §1/§5; nothing here is argued
+  MINOR-eligible.
+- [x] **Do the release notes list every breaking change with its migration?** Yes — the
+  one breaking change is the *Removed* entry, with the `to_dict()` recipe, the three ways
+  it differs from what the old method returned, and the `schema='raw'` replacement. The
+  dependency change (pandas to dev-only) removes nothing any consumer imported through
+  this package: the `dataframe` extra was the only way pandas arrived, and it is the
+  removed surface.
+- [x] **Does `MetricFrame`'s format or axis vocabulary change? If so, has it been agreed
+  with views-reporting and views-pipeline-core?** No. Axes, `MEAN_GROUP_ID`,
+  `SCHEMA_TO_EVAL_TYPE`, dtype and the `save`/`load` layout are unchanged; the CI
+  runtime-only step emits a frame from the release commit with pandas and scikit-learn
+  removed.
+
+---
+
 ## [1.1.0] — 2026-09-18
 
 The first release cut through the gated publish path (S2, #68): `publish_package.yml` now

README.md

@@ -89,7 +89,7 @@ The evaluation ontology has been updated to be more explicit and task-specific.
 
 The **VIEWS Evaluation** repository provides a standardized framework for **assessing time-series forecasting models** used in the **VIEWS conflict prediction pipeline**. It ensures consistent, robust, and interpretable evaluations through **metrics tailored to conflict-related data**, which often exhibit **right-skewness and zero-inflation**.
 
-The library is built on a **three-layer architecture** with a framework-agnostic NumPy core, ensuring that all mathematical evaluation logic is independent of Pandas or any other data-frame library.  
+The library is built as a pure NumPy + SciPy core (ADR-011's Level 0, which includes `EvaluationReport`) plus a thin emit layer — `MetricFrame` and `EvaluationReport.to_metric_frame()`, the emit path of the logging standard §5.1 — with orchestration left to the caller (ADR-011's Level 2, external). All mathematical evaluation logic is independent of Pandas or any other data-frame library — nothing in the package imports one.  
 
 ---
 
@@ -119,8 +119,8 @@ report = evaluator.evaluate(ef)
 # 3. Access results
 report.to_dict()                     # {'target', 'task', 'pred_type', 'schemas': {'step': {...}, ...}}
 report.get_schema_results("month")   # typed metrics dataclass
-# to_dataframe() is deprecated and removed in 2.0.0; build a DataFrame from
-# to_dict()['schemas'][schema] in the caller if you need one.
+# Need a DataFrame? Build it in the caller:
+#   pd.DataFrame.from_dict(report.to_dict()['schemas']['month'], orient='index')
 ```
 
 > For the full walkthrough including input formatting and sample evaluation, see [`documentation/integration_guide.md`](documentation/integration_guide.md).
@@ -257,7 +257,7 @@ pip install views_evaluation
 ---
 ## 🏗 **Architecture**
 
-The library follows a strict three-layer architecture (ADR-011):
+The library follows the strict layering of ADR-011; the first two levels live in this repository:
 
 ```
 Level 0 — Pure Core (NumPy + SciPy; no dataframe libraries)
@@ -268,7 +268,7 @@ Level 0 — Pure Core (NumPy + SciPy; no dataframe libraries)
   Profiles                   Named hyperparameter sets (base, hydranet_ucdp, ...)
 
 Level 1 — Bridge / Emit
-  EvaluationReport      Results container with dict / MetricFrame export (DataFrame export deprecated)
+  EvaluationReport      Results container with dict / MetricFrame export
   MetricFrame           Typed, provenance-stamped evaluation-of-record (views-frames ADR-020)
 
 Level 2 — Orchestration

documentation/ADRs/041_evaluation_output_schema.md

@@ -22,7 +22,7 @@ The 2025 version of this ADR asserted that *"views-evaluation returns a structur
 
 - `to_dict()` — nested plain dict; the pandas-free, extra-free path.
 - `get_schema_results(schema)` — mapped onto the typed metrics dataclasses.
-- `to_dataframe(schema)` — pandas DataFrame; requires the optional `dataframe` extra. **Deprecated; removed in 2.0.0** (ADR-022 §2): build the DataFrame from `to_dict()` in the caller.
+- `to_dataframe(schema)` — *removed in 2.0.0* after a 1.1.0 `DeprecationWarning` (ADR-022 §2). It returned a pandas DataFrame behind the optional `dataframe` extra; a caller builds one from `to_dict()` now.
 
 `views-evaluation` does **not** serialise these to JSON or render HTML. That remains the orchestrator's job, and the original rationale still holds for them: formatting and run context belong where the run is orchestrated.