benchflow 0.6.7.dev1839__tar.gz → 0.6.7.dev1840__tar.gz

{benchflow-0.6.7.dev1839 → benchflow-0.6.7.dev1840}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: benchflow
-Version: 0.6.7.dev1839
+Version: 0.6.7.dev1840
 Summary: Multi-turn agent benchmarking with ACP — run any agent, any model, any provider.
 Project-URL: Homepage, https://github.com/benchflow-ai/benchflow
 Project-URL: Repository, https://github.com/benchflow-ai/benchflow

{benchflow-0.6.7.dev1839 → benchflow-0.6.7.dev1840}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "benchflow"
-version = "0.6.7.dev1839"
+version = "0.6.7.dev1840"
 description = "Multi-turn agent benchmarking with ACP — run any agent, any model, any provider."
 readme = "README.md"
 requires-python = ">=3.12"

benchflow-0.6.7.dev1840/src/benchflow/cli/_failure_evidence.py

@@ -0,0 +1,167 @@
+"""Verifier-artifact mining for the eval report's failure lines.
+
+The one file-touching corner of the CLI's final report block: when a failed
+task's reason would otherwise be the bare ``reward X`` fallback, mine the
+rollout's verifier artifacts for a one-line explanation. Lives in its own
+module so ``cli/_shared.py`` stays the side-effect-free display helpers it
+advertises.
+
+Contract (the engine stays file-free — these reads are CLI-side, report-time
+only):
+
+- The rollout dir is resolved exactly, from the ``rollout_name`` the engine
+  records on every result — never guessed from globs.
+- Evidence sources are tried in order (CTRF report, then test-stdout tail);
+  the first that yields a line wins.
+- Every read is bounded (``_ARTIFACT_READ_BYTES`` per file) and nothing here
+  raises — any surprise degrades to ``None``, i.e. the bare reward reason.
+"""
+
+from __future__ import annotations
+
+import json
+import re
+from typing import TYPE_CHECKING, NamedTuple
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+    from pathlib import Path
+
+# Never read more than this many bytes per file, and evidence is only mined
+# for the <= _MAX_FAILURE_LINES tasks the report block displays.
+_ARTIFACT_READ_BYTES = 64 * 1024
+# Pytest final summary, decoration stripped: "1 failed, 2 passed in 40.86s".
+_PYTEST_SUMMARY_RE = re.compile(r"\d+ failed\b")
+
+
+class FailureLine(NamedTuple):
+    """A failure one-liner in two parts with distinct truncation contracts.
+
+    ``body`` is the free-text evidence (test name plus assertion) and competes
+    for the display line's char budget; ``suffix`` is the compact multi-failure
+    roll-up count (empty when the evidence covers everything) and must survive
+    display truncation. Renderers give ``body`` the full line budget and append
+    ``suffix`` whole past it — never truncating the concatenation, or a long
+    assertion would silently eat the "there is more broken than this" signal.
+    """
+
+    body: str
+    suffix: str = ""
+
+
+def _display_test_name(raw_name: str) -> str:
+    """Test name for display: node-id path segments dropped, param id kept.
+
+    Full pytest node ids ("tests/test_x.py::test_build[param]") waste the
+    100-char line budget; keep the function name plus any ``[param]`` id.
+    The split must not touch the bracket part — a param id may itself contain
+    ``::`` — so only the text before the first ``[`` is segment-split. (When
+    the report was written by pytest-json-ctrf — as of 0.5.x — the param id
+    is already gone — the plugin stores ``nodeid.split('[')[0]`` as ``name``
+    — but other CTRF producers keep the full name, and we must not re-trim
+    it.)
+    """
+    head, bracket, param = raw_name.partition("[")
+    return head.rsplit("::", 1)[-1] + bracket + param
+
+
+def _ctrf_failure_line(ctrf_path: Path) -> FailureLine | None:
+    """``<test> failed[: <assertion>]`` from the first failed CTRF test.
+
+    The verifier's pytest run writes a CTRF report (``pytest --ctrf``, see the
+    standard path in ``task_authoring/structural_checks.py``). Per test the
+    useful failure evidence is the ``E …`` assertion line inside ``trace``;
+    ``message`` is only a generic phase description, so it is the fallback.
+    When the report holds more than one failed test, the first failure's line
+    carries a count suffix (``(+N more failures; P/T checks passed)``) so the
+    console never under-reports how much is broken.
+    """
+    if ctrf_path.stat().st_size > _ARTIFACT_READ_BYTES:
+        # JSON only parses whole — an oversized report is skipped (not
+        # truncated) and the next evidence source gets its turn.
+        return None
+    data = json.loads(ctrf_path.read_text(encoding="utf-8", errors="replace"))
+    raw_tests = (data.get("results") or {}).get("tests") or []
+    tests = [test for test in raw_tests if isinstance(test, dict)]
+    failed = [test for test in tests if test.get("status") == "failed"]
+    if not failed:
+        return None
+    suffix = ""
+    if len(failed) > 1:
+        # Counts come from the same rolled-up test list the first-failure pick
+        # uses, so the suffix can never disagree with the shown evidence.
+        # T is all entries — skips count against the denominator, not as passes.
+        extra = len(failed) - 1
+        passed = sum(1 for test in tests if test.get("status") == "passed")
+        plural = "" if extra == 1 else "s"
+        suffix = (
+            f" (+{extra} more failure{plural}; {passed}/{len(tests)} checks passed)"
+        )
+    test = failed[0]
+    name = _display_test_name(str(test.get("name") or "test"))
+    trace = test.get("trace")
+    if isinstance(trace, str):
+        for line in trace.splitlines():
+            stripped = line.strip()
+            if stripped.startswith("E "):
+                assertion = stripped[2:].strip()
+                # "<test> failed:" already says it's an assertion — drop
+                # the prefix to reclaim line budget for the message.
+                assertion = assertion.removeprefix("AssertionError: ")
+                return FailureLine(f"{name} failed: {assertion}", suffix)
+    message = test.get("message")
+    if isinstance(message, str) and message.strip():
+        return FailureLine(f"{name} failed: {' '.join(message.split())}", suffix)
+    return FailureLine(f"{name} failed", suffix)
+
+
+def _stdout_tail_failure_line(stdout_path: Path) -> FailureLine | None:
+    """Last pytest summary ("N failed, M passed …") — else last ``FAILED …``
+    line — from a bounded tail of the verifier's test-stdout capture. No count
+    suffix: the summary line already carries the full counts itself."""
+    with stdout_path.open("rb") as fh:
+        fh.seek(0, 2)  # SEEK_END
+        fh.seek(max(0, fh.tell() - _ARTIFACT_READ_BYTES))
+        tail = fh.read(_ARTIFACT_READ_BYTES).decode("utf-8", errors="replace")
+    last_failed: str | None = None
+    for line in reversed(tail.splitlines()):
+        bare = line.strip().strip("=").strip()
+        if _PYTEST_SUMMARY_RE.match(bare):
+            return FailureLine(bare)
+        if last_failed is None and line.strip().startswith("FAILED "):
+            last_failed = line.strip()
+    return FailureLine(last_failed) if last_failed is not None else None
+
+
+def artifact_failure_evidence(
+    job_dir: Path, rollout_name: str
+) -> tuple[FailureLine, Path] | None:
+    """``FailureLine`` mined from the rollout's verifier artifacts, or ``None``.
+
+    Returns the structured one-liner plus the verifier dir it came from (for
+    the report block's single ``(details: …)`` pointer line). Never raises.
+    """
+    # Local import: RolloutPaths pulls in the task package, which the CLI
+    # shouldn't pay for until a failure actually needs artifact evidence.
+    from benchflow.task.paths import RolloutPaths
+
+    rollout_paths = RolloutPaths(rollout_dir=job_dir / rollout_name)
+    verifier_dir = rollout_paths.verifier_dir
+    attempts: list[tuple[Path, Callable[[Path], FailureLine | None]]] = [
+        # ctrf.json has no RolloutPaths property — the verifier recovers it by
+        # literal name (verifier_core.py, `_recover_main_verifier_outputs`).
+        (verifier_dir / "ctrf.json", _ctrf_failure_line),
+        (rollout_paths.test_stdout_path, _stdout_tail_failure_line),
+    ]
+    for path, extract in attempts:
+        try:
+            if not path.is_file():
+                continue
+            detail = extract(path)
+        except Exception:
+            # Unreadable file, bad JSON, permissions … try the next source;
+            # the report must never crash over evidence mining.
+            continue
+        if detail is not None:
+            return detail, verifier_dir
+    return None

{benchflow-0.6.7.dev1839 → benchflow-0.6.7.dev1840}/src/benchflow/cli/_shared.py

@@ -21,7 +21,7 @@ from rich.console import Console
 from rich.markup import escape
 
 from benchflow._utils.text import truncate_end
-from benchflow.cli._failure_evidence import artifact_failure_evidence
+from benchflow.cli._failure_evidence import FailureLine, artifact_failure_evidence
 
 if TYPE_CHECKING:
     from pathlib import Path
@@ -106,7 +106,8 @@ def _exit_if_evaluation_had_errors(result: object) -> None:
 
 
 # Final-block failure lines: keep the block skimmable on big jobs and each
-# line inside a typical terminal width.
+# line inside a typical terminal width — except lines carrying a multi-failure
+# count suffix, which deliberately run ~40 chars past the budget.
 _MAX_FAILURE_LINES = 5
 _FAILURE_LINE_LIMIT = 100
 _FAILURE_REASON_METRICS = 3
@@ -114,7 +115,7 @@ _FAILURE_REASON_METRICS = 3
 
 def _failure_reason(
     failure: TaskFailure, job_dir: Path | None = None
-) -> tuple[str, Path | None]:
+) -> tuple[FailureLine, Path | None]:
     """One cheap line explaining why a FAILED (scored, reward != 1) task
     failed, plus the verifier dir when artifacts supplied the evidence.
 
@@ -123,13 +124,15 @@ def _failure_reason(
     first — they explain the miss); else the reward plus a one-liner mined
     from the rollout's verifier artifacts (bounded CLI-side reads resolved via
     the ``rollout_name`` the engine records — the engine stays file-free);
-    else just the reward. The returned path is non-``None`` only for the
-    artifact tier, so the caller can print one ``(details: …)`` pointer per
-    report block.
+    else just the reward. The reason is a ``FailureLine`` so the artifact
+    tier's multi-failure count suffix stays separable from the truncatable
+    body (only the artifact tier ever sets it). The returned path is
+    non-``None`` only for the artifact tier, so the caller can print one
+    ``(details: …)`` pointer per report block.
     """
     if failure.verifier_error:
         # Collapse whitespace: verifier errors are routinely multi-line.
-        return " ".join(failure.verifier_error.split()), None
+        return FailureLine(" ".join(failure.verifier_error.split())), None
     rewards = failure.rewards or {}
     reward = rewards.get("reward")
     metrics = [
@@ -144,15 +147,18 @@ def _failure_reason(
         shown = ", ".join(
             f"{name} {value}" for name, value in metrics[:_FAILURE_REASON_METRICS]
         )
-        return f"reward {reward} — {shown}", None
+        return FailureLine(f"reward {reward} — {shown}"), None
     # No rollout_name (pre-#957-follow-up result.json, sharded aggregation):
     # nothing to resolve — the bare reward is as honest as it gets.
     if job_dir is not None and failure.rollout_name:
         evidence = artifact_failure_evidence(job_dir, failure.rollout_name)
         if evidence is not None:
             detail, verifier_dir = evidence
-            return f"reward {reward} — {detail}", verifier_dir
-    return f"reward {reward}", None
+            return (
+                FailureLine(f"reward {reward} — {detail.body}", detail.suffix),
+                verifier_dir,
+            )
+    return FailureLine(f"reward {reward}"), None
 
 
 def _report_eval_result(result: EvaluationResult, job_dir: Path | None = None) -> None:
@@ -207,11 +213,16 @@ def _report_eval_result(result: EvaluationResult, job_dir: Path | None = None) -
         reason, verifier_dir = _failure_reason(failure, job_dir)
         if artifact_pointer is None and verifier_dir is not None:
             artifact_pointer = verifier_dir
-        line = truncate_end(
-            f"  ✗ {failure.task_name}: {reason}",
+        # The char budget governs only the free-text body; the multi-failure
+        # count suffix is appended AFTER truncation so a long assertion can
+        # never swallow the "more than this is broken" signal. Worst case the
+        # line runs ~40 chars past the budget — completeness beats strict
+        # width for this one signal.
+        body = truncate_end(
+            f"  ✗ {failure.task_name}: {reason.body}",
             _FAILURE_LINE_LIMIT,
         )
-        console.print(f"[dim]{escape(line)}[/dim]")
+        console.print(f"[dim]{escape(body + reason.suffix)}[/dim]")
     extra = len(failures) - _MAX_FAILURE_LINES
     if extra > 0:
         console.print(f"[dim]  … and {extra} more[/dim]")

{benchflow-0.6.7.dev1839 → benchflow-0.6.7.dev1840}/tests/test_cli_live_progress.py

@@ -476,20 +476,11 @@ def _bare_failure(task_name: str, rollout_name: str | None = None):
     )
 
 
-def _write_ctrf(verifier_dir, *, name: str | None, trace: str | None = None) -> None:
-    """Write a minimal CTRF report (pytest-json-ctrf shape).
-
-    ``name`` is the failed test's node id; ``None`` writes an all-passed report.
-    """
+def _write_ctrf_tests(verifier_dir, tests: list[dict]) -> None:
+    """Write a CTRF report (pytest-json-ctrf shape) with the given tests."""
     import json
 
     verifier_dir.mkdir(parents=True, exist_ok=True)
-    tests = [{"name": "tests/test_x.py::test_ok", "status": "passed"}]
-    if name is not None:
-        test: dict = {"name": name, "status": "failed"}
-        if trace is not None:
-            test["trace"] = trace
-        tests.append(test)
     (verifier_dir / "ctrf.json").write_text(
         json.dumps(
             {
@@ -500,6 +491,20 @@ def _write_ctrf(verifier_dir, *, name: str | None, trace: str | None = None) ->
     )
 
 
+def _write_ctrf(verifier_dir, *, name: str | None, trace: str | None = None) -> None:
+    """Write a minimal CTRF report.
+
+    ``name`` is the failed test's node id; ``None`` writes an all-passed report.
+    """
+    tests = [{"name": "tests/test_x.py::test_ok", "status": "passed"}]
+    if name is not None:
+        test: dict = {"name": name, "status": "failed"}
+        if trace is not None:
+            test["trace"] = trace
+        tests.append(test)
+    _write_ctrf_tests(verifier_dir, tests)
+
+
 def _write_stdout(verifier_dir, text: str) -> None:
     verifier_dir.mkdir(parents=True, exist_ok=True)
     (verifier_dir / "test-stdout.txt").write_text(text)
@@ -527,6 +532,121 @@ def test_failure_reason_artifact_tier_reads_ctrf(tmp_path):
         "Build failed for py312" in out
     )
     assert f"(details: {tmp_path / 'fix-build__ab12cd34' / 'verifier'})" in out
+    # A single failed test needs no roll-up count.
+    assert "(+" not in out
+
+
+def test_ctrf_multi_failure_appends_count_suffix(tmp_path):
+    # Dogfood follow-up (skillsbench dialogue-parser, reward 0.667): the CTRF
+    # held 2 failed / 4 passed but the console named only the first failure —
+    # a reader stopping at the console under-counted what was broken. With >1
+    # failed test the line gains a count suffix, and a param id carried by the
+    # report's test name is preserved end-to-end. (pytest-json-ctrf, as of
+    # 0.5.x, strips the param id at generation time — nodeid.split('[')[0] —
+    # so this fixture uses the spec-conformant full name other producers emit.)
+    _write_ctrf_tests(
+        tmp_path / "dialogue-parser__ab12cd34" / "verifier",
+        [
+            {"name": "test_outputs.py::test_system_basics", "status": "passed"},
+            {"name": "test_outputs.py::test_narrative_content", "status": "passed"},
+            {
+                "name": "test_outputs.py::test_graph_logic[reachability]",
+                "status": "failed",
+                "trace": (
+                    ">           assert not unreachable\n"
+                    "E           AssertionError: Unreachable nodes found: "
+                    "['End']...\n"
+                ),
+            },
+            {
+                "name": "test_outputs.py::test_visualization_validity",
+                "status": "failed",
+            },
+            {"name": "test_outputs.py::test_content_integrity", "status": "passed"},
+            {"name": "test_outputs.py::test_structural_integrity", "status": "passed"},
+        ],
+    )
+    from benchflow.evaluation import TaskFailure
+
+    out = _reported(
+        _failed_result(
+            [
+                TaskFailure(
+                    task_name="dialogue-parser",
+                    rewards={"reward": 0.667},
+                    verifier_error=None,
+                    rollout_name="dialogue-parser__ab12cd34",
+                )
+            ]
+        ),
+        tmp_path,
+    )
+    # The body hits the 100-char budget and truncates; the suffix rides after
+    # it whole. Counts are report-wide: 1 extra failure, 4 of 6 checks passed.
+    assert (
+        "✗ dialogue-parser: reward 0.667 — test_graph_logic[reachability] "
+        "failed: Unreachable nodes found:… (+1 more failure; 4/6 checks passed)" in out
+    )
+
+
+def test_ctrf_count_suffix_plural_and_param_id_with_colons(tmp_path):
+    # 3 failed tests pluralize the suffix, and a param id containing "::" must
+    # not be mangled by the node-id segment split (`test_foo[a::b]`, not `b]`).
+    _write_ctrf_tests(
+        tmp_path / "multi__ab12cd34" / "verifier",
+        [
+            {"name": "tests/test_x.py::test_foo[a::b]", "status": "failed"},
+            {"name": "tests/test_x.py::test_bar", "status": "failed"},
+            {"name": "tests/test_x.py::test_baz", "status": "failed"},
+            {"name": "tests/test_x.py::test_ok", "status": "passed"},
+        ],
+    )
+    out = _reported(
+        _failed_result([_bare_failure("multi", "multi__ab12cd34")]), tmp_path
+    )
+    assert (
+        "✗ multi: reward 0.0 — test_foo[a::b] failed "
+        "(+2 more failures; 1/4 checks passed)" in out
+    )
+
+
+def test_ctrf_count_suffix_survives_truncation(tmp_path):
+    # A long assertion must not push the count suffix off the line: the body
+    # alone is truncated to the 100-char budget, then the suffix is appended
+    # whole — so the "more than this is broken" signal always survives.
+    _write_ctrf_tests(
+        tmp_path / "long__ab12cd34" / "verifier",
+        [
+            {
+                "name": "tests/test_x.py::test_long",
+                "status": "failed",
+                "trace": "E   AssertionError: " + "verbose diagnosis " * 20 + "\n",
+            },
+            {"name": "tests/test_x.py::test_other", "status": "failed"},
+            {"name": "tests/test_x.py::test_ok", "status": "passed"},
+        ],
+    )
+    out = _reported(_failed_result([_bare_failure("long", "long__ab12cd34")]), tmp_path)
+    suffix = " (+1 more failure; 1/3 checks passed)"
+    (line,) = [ln for ln in out.splitlines() if "✗ long" in ln]
+    assert line.rstrip().endswith(f"…{suffix}")
+    assert len(line.rstrip()) <= 100 + len(suffix)
+
+
+def test_stdout_tail_tier_never_gets_count_suffix(tmp_path):
+    # The suffix is a CTRF-tier concept: the stdout tail's pytest summary line
+    # already carries the full counts, so nothing is appended there.
+    _write_stdout(
+        tmp_path / "fix-build__ab12cd34" / "verifier",
+        "FAILED tests/test_build.py::test_a - AssertionError: boom\n"
+        "FAILED tests/test_build.py::test_b - AssertionError: boom\n"
+        "=========== 2 failed, 1 passed in 4.20s ===========\n",
+    )
+    out = _reported(
+        _failed_result([_bare_failure("fix-build", "fix-build__ab12cd34")]), tmp_path
+    )
+    assert "✗ fix-build: reward 0.0 — 2 failed, 1 passed in 4.20s" in out
+    assert "(+" not in out
 
 
 def test_failure_reason_artifact_tier_requires_rollout_name(tmp_path):

benchflow-0.6.7.dev1839/src/benchflow/cli/_failure_evidence.py

@@ -1,122 +0,0 @@
-"""Verifier-artifact mining for the eval report's failure lines.
-
-The one file-touching corner of the CLI's final report block: when a failed
-task's reason would otherwise be the bare ``reward X`` fallback, mine the
-rollout's verifier artifacts for a one-line explanation. Lives in its own
-module so ``cli/_shared.py`` stays the side-effect-free display helpers it
-advertises.
-
-Contract (the engine stays file-free — these reads are CLI-side, report-time
-only):
-
-- The rollout dir is resolved exactly, from the ``rollout_name`` the engine
-  records on every result — never guessed from globs.
-- Evidence sources are tried in order (CTRF report, then test-stdout tail);
-  the first that yields a line wins.
-- Every read is bounded (``_ARTIFACT_READ_BYTES`` per file) and nothing here
-  raises — any surprise degrades to ``None``, i.e. the bare reward reason.
-"""
-
-from __future__ import annotations
-
-import json
-import re
-from typing import TYPE_CHECKING
-
-if TYPE_CHECKING:
-    from collections.abc import Callable
-    from pathlib import Path
-
-# Never read more than this many bytes per file, and evidence is only mined
-# for the <= _MAX_FAILURE_LINES tasks the report block displays.
-_ARTIFACT_READ_BYTES = 64 * 1024
-# Pytest final summary, decoration stripped: "1 failed, 2 passed in 40.86s".
-_PYTEST_SUMMARY_RE = re.compile(r"\d+ failed\b")
-
-
-def _ctrf_failure_line(ctrf_path: Path) -> str | None:
-    """``<test> failed[: <assertion>]`` from the first failed CTRF test.
-
-    The verifier's pytest run writes a CTRF report (``pytest --ctrf``, see the
-    standard path in ``task_authoring/structural_checks.py``). Per test the
-    useful failure evidence is the ``E …`` assertion line inside ``trace``;
-    ``message`` is only a generic phase description, so it is the fallback.
-    """
-    if ctrf_path.stat().st_size > _ARTIFACT_READ_BYTES:
-        # JSON only parses whole — an oversized report is skipped (not
-        # truncated) and the next evidence source gets its turn.
-        return None
-    data = json.loads(ctrf_path.read_text(encoding="utf-8", errors="replace"))
-    tests = (data.get("results") or {}).get("tests") or []
-    for test in tests:
-        if not isinstance(test, dict) or test.get("status") != "failed":
-            continue
-        # Full pytest node ids ("tests/test_x.py::test_build") waste the
-        # 100-char line budget; keep the test function name.
-        name = str(test.get("name") or "test").rsplit("::", 1)[-1]
-        trace = test.get("trace")
-        if isinstance(trace, str):
-            for line in trace.splitlines():
-                stripped = line.strip()
-                if stripped.startswith("E "):
-                    assertion = stripped[2:].strip()
-                    # "<test> failed:" already says it's an assertion — drop
-                    # the prefix to reclaim line budget for the message.
-                    assertion = assertion.removeprefix("AssertionError: ")
-                    return f"{name} failed: {assertion}"
-        message = test.get("message")
-        if isinstance(message, str) and message.strip():
-            return f"{name} failed: {' '.join(message.split())}"
-        return f"{name} failed"
-    return None
-
-
-def _stdout_tail_failure_line(stdout_path: Path) -> str | None:
-    """Last pytest summary ("N failed, M passed …") — else last ``FAILED …``
-    line — from a bounded tail of the verifier's test-stdout capture."""
-    with stdout_path.open("rb") as fh:
-        fh.seek(0, 2)  # SEEK_END
-        fh.seek(max(0, fh.tell() - _ARTIFACT_READ_BYTES))
-        tail = fh.read(_ARTIFACT_READ_BYTES).decode("utf-8", errors="replace")
-    last_failed: str | None = None
-    for line in reversed(tail.splitlines()):
-        bare = line.strip().strip("=").strip()
-        if _PYTEST_SUMMARY_RE.match(bare):
-            return bare
-        if last_failed is None and line.strip().startswith("FAILED "):
-            last_failed = line.strip()
-    return last_failed
-
-
-def artifact_failure_evidence(
-    job_dir: Path, rollout_name: str
-) -> tuple[str, Path] | None:
-    """One-liner mined from the rollout's verifier artifacts, or ``None``.
-
-    Returns the one-liner plus the verifier dir it came from (for the report
-    block's single ``(details: …)`` pointer line). Never raises.
-    """
-    # Local import: RolloutPaths pulls in the task package, which the CLI
-    # shouldn't pay for until a failure actually needs artifact evidence.
-    from benchflow.task.paths import RolloutPaths
-
-    rollout_paths = RolloutPaths(rollout_dir=job_dir / rollout_name)
-    verifier_dir = rollout_paths.verifier_dir
-    attempts: list[tuple[Path, Callable[[Path], str | None]]] = [
-        # ctrf.json has no RolloutPaths property — the verifier recovers it by
-        # literal name (verifier_core.py, `_recover_main_verifier_outputs`).
-        (verifier_dir / "ctrf.json", _ctrf_failure_line),
-        (rollout_paths.test_stdout_path, _stdout_tail_failure_line),
-    ]
-    for path, extract in attempts:
-        try:
-            if not path.is_file():
-                continue
-            detail = extract(path)
-        except Exception:
-            # Unreadable file, bad JSON, permissions … try the next source;
-            # the report must never crash over evidence mining.
-            continue
-        if detail is not None:
-            return detail, verifier_dir
-    return None