fix: render unestablished effective context

src/shiftedx_bench/model_card.py

@@ -141,8 +141,14 @@ def render_benchmark_block(
             f"{100 * float(card['accuracy']):.1f}% | {_format_float(card['mean_wall_s'])} s | "
             f"{_format_float(card['mean_decode_tokens_per_second'])} tok/s | "
             f"{_format_memory(card['max_active_memory_bytes'])} |"
-        )
+    )
     lengths = ", ".join(f"{int(value):,}" for value in sorted(map(int, context["accuracy_by_length"])))
+    effective_context = context.get("effective_context_length")
+    effective_context_text = (
+        f"{int(effective_context):,} tokens"
+        if effective_context is not None
+        else "not established at the benchmark's 90% threshold"
+    )
     total_wall = sum(float(row.get("telemetry", {}).get("wall_s") or 0) for row in rows)
     lines.extend(
         [
@@ -150,7 +156,7 @@ def render_benchmark_block(
             f"- Tested model revision: [`{model_revision}`](https://huggingface.co/{model_repo}/tree/{model_revision})",
             f"- Benchmark: [Shiftedx Bench v{benchmark_version}](https://huggingface.co/datasets/Shiftedx/shiftedx-bench/tree/{benchmark_revision})",
             f"- Context lengths represented: {lengths} prompt tokens; effective tested context: "
-            f"{int(context['effective_context_length']):,} tokens",
+            f"{effective_context_text}",
             f"- Runtime contract: {runtime}; KV cache `{kv_cache}`; MTP depth `{mtp_depth}`",
             f"- Host: {host}",
             f"- Total measured request wall time: {total_wall:.2f} seconds",

tests/test_model_card.py

@@ -117,3 +117,36 @@ def test_model_card_export_rejects_incomplete_or_wrong_revision(tmp_path):
             model_revision="c" * 40,
             benchmark_revision=BENCHMARK_REVISION,
         )
+
+
+def test_model_card_renders_null_effective_context_as_not_established(tmp_path):
+    _write_run(tmp_path)
+    context_path = tmp_path / "context.jsonl"
+    rows = [json.loads(line) for line in context_path.read_text().splitlines()]
+    for row in rows:
+        row["passed"] = False
+    context_path.write_text(
+        "".join(json.dumps(row) + "\n" for row in rows), encoding="utf-8"
+    )
+    validated, version = validate_post_publish_run(
+        tmp_path,
+        variant=VARIANT,
+        profile="vision",
+        model_repo=MODEL_REPO,
+        model_revision=MODEL_REVISION,
+        benchmark_revision=BENCHMARK_REVISION,
+    )
+    block = render_benchmark_block(
+        validated,
+        variant=VARIANT,
+        model_repo=MODEL_REPO,
+        model_revision=MODEL_REVISION,
+        benchmark_revision=BENCHMARK_REVISION,
+        benchmark_version=version,
+        host="Apple Silicon, 64 GiB",
+        runtime="MTPLX 2.7.1",
+        kv_cache="off",
+        mtp_depth="2",
+        max_window_status="not-run",
+    )
+    assert "effective tested context: not established at the benchmark's 90% threshold" in block