diff --git a/CHANGELOG.md b/CHANGELOG.md
index 8cc50e0..3f46902 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -96,6 +96,11 @@ adheres to [Semantic Versioning](https://semver.org/).
 - The missing-pyarrow error names the feature that needs it (for example Arrow
   output), and Parquet metadata reads no longer fail with `AttributeError` in a
   fresh process (#215).
+- Docs: the streaming per-batch field table described `rolling_trust_score` as
+  row-weighted. It is an unweighted mean of the per-batch scores in the window,
+  so a 1-row batch counts as much as a 999-row one; only
+  `cumulative_trust_score` is row-weighted. Documentation fix only — the score
+  is unchanged (#349).
 
 ## [2.0.0] - 2026-07-20
 
diff --git a/docs/streaming.md b/docs/streaming.md
index 8db5cc2..56b74a9 100644
--- a/docs/streaming.md
+++ b/docs/streaming.md
@@ -97,7 +97,7 @@ in `report.to_dict()`) carries:
 | `batch_id` | 1-based index of this batch |
 | `rows_in_batch` / `rows_seen_total` | rows in this batch / cumulative across the stream |
 | `batch_trust_score` | trust score of this cleaned batch (0-100) |
-| `rolling_trust_score` | row-weighted trust over the recent window |
+| `rolling_trust_score` | unweighted mean of the last `rolling_trust_window` batch scores |
 | `cumulative_trust_score` | row-weighted trust over the whole stream |
 | `schema_drift_detected` | whether drift was flagged for this batch |
 | `warmup_phase` | whether this batch was still in warmup |
diff --git a/tests/test_streaming_state.py b/tests/test_streaming_state.py
index 8279792..602d099 100644
--- a/tests/test_streaming_state.py
+++ b/tests/test_streaming_state.py
@@ -2,7 +2,9 @@
 
 import numpy as np
 import pandas as pd
+import pytest
 
+from freshdata.streaming import StreamingCleaner
 from freshdata.streaming._state import ColumnState, StreamingState
 from freshdata.streaming._stats import BoundedCounter, ReservoirSampler, Welford
 
@@ -90,6 +92,53 @@ def test_streaming_state_tracks_rows_and_trust():
     assert abs(cumulative - (80 * 3 + 90 * 1) / 4) < 1e-9  # rows-weighted
 
 
+# --- documented behaviour: the rolling score is NOT row-weighted -------------
+# docs/streaming.md called rolling_trust_score "row-weighted trust over the
+# recent window", but it is an unweighted mean of the per-batch scores in the
+# window — a 1-row batch counts exactly as much as a 999-row one. Only
+# cumulative_trust_score is row-weighted. The docs were corrected to match;
+# these tests pin the arithmetic so the two cannot drift apart again.
+
+def test_rolling_trust_is_unweighted_while_cumulative_is_row_weighted():
+    state = StreamingState(rolling_trust_window=8)
+    state.record_trust(0.0, rows=1)               # tiny, terrible batch
+    rolling, cumulative = state.record_trust(100.0, rows=999)  # huge, clean batch
+
+    assert rolling == 50.0                        # (0 + 100) / 2 — rows ignored
+    assert abs(cumulative - (0 * 1 + 100 * 999) / 1000) < 1e-9  # 99.9, rows honoured
+    assert rolling != pytest.approx(cumulative)   # the two must not be conflated
+
+
+def test_rolling_trust_ignores_batch_size_end_to_end():
+    """The issue's reproduction: a 1-row all-None batch, then 999 clean rows.
+
+    ``drop_empty_rows=False`` matters here: the streaming representation pass
+    clears ``drop_empty_columns`` but keeps the default row dropping, so the
+    all-None row would otherwise be removed before the batch is scored. That
+    leaves a 0-row batch, and a 0-vs-999 split cannot demonstrate anything
+    about weighting — the tiny batch would carry no weight under either rule.
+    """
+    cleaner = StreamingCleaner(verbose=False, drop_empty_rows=False)
+
+    _, tiny_rep = cleaner.clean_batch(pd.DataFrame({"a": [None], "b": [None]}))
+    _, big_rep = cleaner.clean_batch(pd.DataFrame({
+        "a": list(range(999)), "b": [float(i) for i in range(999)],
+    }))
+
+    tiny, big = tiny_rep.streaming, big_rep.streaming
+    scores = [tiny["batch_trust_score"], big["batch_trust_score"]]
+    rows = [tiny["rows_in_batch"], big["rows_in_batch"]]
+    assert rows == [1, 999]          # the 1-vs-999 split the issue describes
+    assert scores[0] < scores[1]     # and the batches really do score differently
+
+    # Derived from the observed scores, so an unrelated change to how a batch is
+    # scored will not break this — only a change to the *weighting* will.
+    assert big["rolling_trust_score"] == pytest.approx(sum(scores) / 2)
+    assert big["rolling_trust_score"] != pytest.approx(
+        sum(s * r for s, r in zip(scores, rows)) / sum(rows)
+    )
+
+
 def test_state_to_dict_is_json_friendly():
     state = StreamingState()
     state.observe_batch(pd.DataFrame({"n": [1.0, 2.0], "c": ["x", "y"]}),