fix(report): write the report card as UTF-8, not the platform locale encoding

src/steerbench/cli.py

@@ -56,7 +56,7 @@ def _ensure_side_csv(side_csv: Path | None, out_dir: Path) -> Path:
         return side_csv
     out_dir.mkdir(parents=True, exist_ok=True)
     stub = out_dir / "side_effects.csv"
-    stub.write_text("benchmark,unsteered_acc,steered_acc\n")
+    stub.write_text("benchmark,unsteered_acc,steered_acc\n", encoding="utf-8")
     return stub
 
 

src/steerbench/metrics.py

@@ -577,7 +577,7 @@ def load_mmlu_slice(
         raise ValueError(f"n must be >= 1, got {n}")
     path = _cache_path(cache_dir, f"mmlu_{n}_{seed}.json")
     if path.exists():
-        rows = json.loads(path.read_text())[:n]
+        rows = json.loads(path.read_text(encoding="utf-8"))[:n]
         return [
             MMLUExample(question=r["question"], choices=list(r["choices"]), answer=int(r["answer"]))
             for r in rows
@@ -598,7 +598,8 @@ def load_mmlu_slice(
     path.write_text(
         json.dumps(
             [{"question": e.question, "choices": e.choices, "answer": e.answer} for e in examples]
-        )
+        ),
+        encoding="utf-8",
     )
     return examples
 
@@ -617,7 +618,7 @@ def load_gsm8k_slice(
         raise ValueError(f"n must be >= 1, got {n}")
     path = _cache_path(cache_dir, f"gsm8k_{n}_{seed}.json")
     if path.exists():
-        rows = json.loads(path.read_text())[:n]
+        rows = json.loads(path.read_text(encoding="utf-8"))[:n]
         return [GSM8KExample(question=r["question"], answer=str(r["answer"])) for r in rows]
 
     from datasets import load_dataset  # local import: gated optional dependency
@@ -628,7 +629,10 @@ def load_gsm8k_slice(
     for row in ds:
         gold = extract_gsm8k_answer(str(row["answer"]))
         examples.append(GSM8KExample(question=str(row["question"]), answer=gold if gold else ""))
-    path.write_text(json.dumps([{"question": e.question, "answer": e.answer} for e in examples]))
+    path.write_text(
+        json.dumps([{"question": e.question, "answer": e.answer} for e in examples]),
+        encoding="utf-8",
+    )
     return examples
 
 

src/steerbench/report.py

@@ -208,7 +208,7 @@ class _RawRow:
 def _read_sweep_rows(path: Path, x_column: str) -> list[_RawRow]:
     """Parse a raw sweep CSV keyed on ``x_column`` (``coeff`` for dose)."""
     rows: list[_RawRow] = []
-    with path.open(newline="") as fh:
+    with path.open(newline="", encoding="utf-8") as fh:
         reader = csv.DictReader(fh)
         missing = {x_column, *_SWEEP_COLUMNS} - set(reader.fieldnames or [])
         if missing:
@@ -273,7 +273,7 @@ def load_layer_curve(path: Path, x_column: str = "layer") -> list[SweepPoint]:
 def load_side_effects(path: Path) -> list[SideEffect]:
     """Parse the side-effect CSV (benchmark, unsteered_acc, steered_acc)."""
     effects: list[SideEffect] = []
-    with path.open(newline="") as fh:
+    with path.open(newline="", encoding="utf-8") as fh:
         reader = csv.DictReader(fh)
         required = {"benchmark", "unsteered_acc", "steered_acc"}
         missing = required - set(reader.fieldnames or [])
@@ -796,8 +796,15 @@ def build_report(
 
     md_path = out_dir / f"{stem}.md"
     html_path = out_dir / f"{stem}.html"
-    md_path.write_text(render_markdown(data, dose_png_path.name, layer_png_path.name))
-    html_path.write_text(render_html(data, dose_png, layer_png))
+    # UTF-8 explicitly: the card carries non-ASCII (the Δ column header, the ⚠️
+    # trap warnings, ± in prose), and Path.write_text otherwise uses the
+    # platform's locale encoding — cp1252 on a default Windows install, which
+    # cannot encode them. The HTML also declares charset=utf-8, so writing it
+    # in anything else would be self-contradictory.
+    md_path.write_text(
+        render_markdown(data, dose_png_path.name, layer_png_path.name), encoding="utf-8"
+    )
+    html_path.write_text(render_html(data, dose_png, layer_png), encoding="utf-8")
 
     return {
         "markdown": md_path,

tests/test_report.py

@@ -259,6 +259,77 @@ def test_render_html_self_contained(data: ReportData) -> None:
     assert "Degenerate trap at layer 1" in html
 
 
+def test_report_files_are_written_as_utf8(csvs: tuple[Path, Path, Path]) -> None:
+    # The card is not ASCII: the side-effects table header carries Δ and the
+    # trap callouts carry ⚠️. Every text read/write in the package therefore has
+    # to name its encoding — Path.write_text/Path.open otherwise fall back to
+    # the platform locale encoding, which is cp1252 on a stock Windows install
+    # and cannot encode either character (build_report raises UnicodeEncodeError
+    # there).
+    #
+    # Asserting on the written bytes alone would not catch a regression on a
+    # UTF-8 platform, so a fresh interpreter runs the whole path under
+    # -X warn_default_encoding, where io.text_encoding() reports every
+    # unencoded call site as an EncodingWarning attributed to its caller. Only
+    # warnings raised from inside steerbench count; optional third-party
+    # dependencies are not this repo's problem.
+    import subprocess
+    import sys
+    import textwrap
+
+    pytest.importorskip("matplotlib")
+    dose_csv, layer_csv, side_csv = csvs
+    out_dir = dose_csv.parent / "card"
+
+    code = textwrap.dedent(
+        """
+        import sys, warnings
+        from pathlib import Path
+        from steerbench import report as r
+
+        package_dir = Path(r.__file__).resolve().parent
+
+        with warnings.catch_warnings(record=True) as caught:
+            warnings.simplefilter("always")
+            written = r.build_report(
+                dose_csv=Path(sys.argv[1]),
+                layer_csv=Path(sys.argv[2]),
+                side_csv=Path(sys.argv[3]),
+                out_dir=Path(sys.argv[4]),
+            )
+
+        offenders = sorted(
+            f"{Path(w.filename).name}:{w.lineno}"
+            for w in caught
+            if issubclass(w.category, EncodingWarning)
+            and package_dir in Path(w.filename).resolve().parents
+        )
+        assert not offenders, "text I/O without an explicit encoding: " + ", ".join(offenders)
+
+        markdown = written["markdown"].read_text(encoding="utf-8")
+        assert "\\u0394" in markdown, "side-effects delta header missing"
+        assert "\\u26a0" in markdown, "degenerate-trap warning missing"
+        html = written["html"].read_text(encoding="utf-8")
+        assert 'charset="utf-8"' in html, "html no longer declares utf-8"
+        assert "\\u26a0" in html, "degenerate-trap callout missing from html"
+        """
+    )
+    subprocess.run(
+        [
+            sys.executable,
+            "-X",
+            "warn_default_encoding",
+            "-c",
+            code,
+            str(dose_csv),
+            str(layer_csv),
+            str(side_csv),
+            str(out_dir),
+        ],
+        check=True,
+    )
+
+
 def test_no_matplotlib_imported() -> None:
     # The parse/analysis/markdown+html path must stay matplotlib-free. Checked in
     # a fresh interpreter: a global sys.modules check is unreliable in a suite