fix: use per-iteration lambda in EM iteration history records

splink/internals/em_training_session.py

@@ -406,7 +406,7 @@ def _iteration_history_records(self) -> list[ModelParameterIterationDetailedReco
             records = [
                 ModelParameterIterationDetailedRecord.from_settings_param_detailed_record(
                     r,
-                    probability_two_random_records_match=self.core_model_settings.probability_two_random_records_match,
+                    probability_two_random_records_match=core_model_settings.probability_two_random_records_match,
                     iteration=iteration,
                 )
                 for r in core_model_settings.parameters_as_detailed_records

tests/test_em_iteration_history_records.py

@@ -0,0 +1,78 @@
+from splink.internals.comparison_level import ComparisonLevelDetailedRecord
+from splink.internals.em_training_session import (
+    EMTrainingSession,
+    ModelParameterDetailedRecord,
+)
+
+
+def _parameter_detailed_record() -> ModelParameterDetailedRecord:
+    cl_rec = ComparisonLevelDetailedRecord(
+        sql_condition="l.first_name = r.first_name",
+        label_for_charts="Exact match on first_name",
+        has_tf_adjustments=False,
+        tf_adjustment_column=None,
+        tf_adjustment_weight=None,
+        is_null_level=False,
+        m_probability=0.9,
+        u_probability=0.1,
+        m_probability_description=None,
+        u_probability_description=None,
+        bayes_factor=9.0,
+        log2_bayes_factor=3.17,
+        bayes_factor_description="",
+        comparison_vector_value=1,
+        max_comparison_vector_value=2,
+        comparison_name="first_name",
+    )
+    return ModelParameterDetailedRecord.from_cl_detailed_record(
+        cl_rec,
+        probability_two_random_records_match=0.5,
+        comparison_sort_order=0,
+    )
+
+
+class _StubCoreModelSettings:
+    def __init__(self, probability_two_random_records_match):
+        self.probability_two_random_records_match = (
+            probability_two_random_records_match
+        )
+        self.parameters_as_detailed_records = [_parameter_detailed_record()]
+
+
+def _session_with_history(lam_history):
+    # EMTrainingSession.__init__ needs a live linker and database, which the
+    # iteration history records do not use; build the instance directly.
+    session = EMTrainingSession.__new__(EMTrainingSession)
+    session._core_model_settings_history = [
+        _StubCoreModelSettings(lam) for lam in lam_history
+    ]
+    # The final (converged) settings, as set by _train after the EM loop.
+    session.core_model_settings = _StubCoreModelSettings(lam_history[-1])
+    return session
+
+
+def test_iteration_history_records_use_each_iterations_probability():
+    # https://github.com/moj-analytical-services/splink/issues/1619
+    # Every iteration's records must carry that iteration's probability, not
+    # the final converged value, otherwise the m/u and match weights history
+    # chart tooltips report the wrong lambda for all but the last iteration.
+    lam_history = [0.001, 0.002, 0.004]
+
+    session = _session_with_history(lam_history)
+
+    records = session._iteration_history_records
+
+    assert [r.iteration for r in records] == [0, 1, 2]
+    assert [r.probability_two_random_records_match for r in records] == lam_history
+
+
+def test_iteration_history_records_single_iteration_matches_final_probability():
+    lam_history = [0.01]
+
+    session = _session_with_history(lam_history)
+
+    records = session._iteration_history_records
+
+    assert len(records) == 1
+    assert records[0].probability_two_random_records_match == 0.01
+    assert records[0].iteration == 0