verifiers 0.2.2.dev25__tar.gz → 0.2.2.dev26__tar.gz

{verifiers-0.2.2.dev25 → verifiers-0.2.2.dev26}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev25
+Version: 0.2.2.dev26
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers

{verifiers-0.2.2.dev25 → verifiers-0.2.2.dev26}/tests/v1/test_e2e.py

@@ -392,23 +392,32 @@ async def test_env_id_best_of_n(run_v1, tmp_path):
 
 @pytest.mark.e2e
 async def test_env_id_agentic_judge(run_v1, tmp_path):
-    """The agentic judge over the echo taskset (needs docker): the judge lands in
-    its own box with the graded transcript uploaded, investigates with real
-    execution, and its parsed verdict lands on the solver's trace under the spec's
-    reward key. Wiring, not taste: the judge followed the verdict-file contract's
-    output contract — the grade itself is the model's call."""
+    """The agentic judge over the echo taskset (needs docker): the box is
+    provisioned once from the solver's runtime policy, the solver plays in it,
+    the judge lands in the SAME box with the graded trace uploaded,
+    investigates with real execution, and its parsed verdict
+    lands on the solver's trace under the spec's reward key. Wiring, not taste:
+    the judge followed the verdict-file contract — the grade itself is the
+    model's call. Exercises the config surface too: a policy-only prompt
+    override (the verdict contract is appended regardless) and
+    reward-composition weights."""
     traces = await run_v1(
         "echo-v1",
         harness=None,  # seats pin their own harness; there is no run-level one
         env={
             "id": "agentic-judge",
-            "solver": {"harness": {"id": "null"}},
-            # The judge reads the transcript and reasons before it writes the
+            # The solver owns the shared box, so the container is pinned here.
+            "solver": {"harness": {"id": "bash"}, "runtime": {"type": "docker"}},
+            # The judge reads the trace and reasons before it writes the
             # verdict file; the shared 2048-token run cap truncates it mid-audit.
             "judge": {
-                "runtime": {"type": "docker"},
+                "harness": {"id": "bash"},
                 "max_output_tokens": 8192,
             },
+            "task": {
+                "prompt": "Check EMPIRICALLY that the agent echoed the word back.",
+            },
+            "score": {"task_weight": 0.5},
         },
         output_dir=tmp_path,
         max_turns=10,
@@ -419,7 +428,7 @@ async def test_env_id_agentic_judge(run_v1, tmp_path):
     (judge,) = [t for t in traces if t.agent_name == "judge"]
     assert solver.ok and judge.ok
     assert judge.trainable is False
-    assert solver.rewards["echoed"] == 1.0  # the task's own reward still runs
+    assert solver.rewards["echoed"] == 0.5  # the task's own reward, rescaled
     assert isinstance(judge.info.get("verdict"), dict)  # scraped off the box
     assert 0.0 <= solver.rewards["judge"] <= 1.0
 

verifiers-0.2.2.dev26/verifiers/v1/envs/agentic_judge/__init__.py

@@ -0,0 +1,15 @@
+from verifiers.v1.envs.agentic_judge.env import (
+    AgenticJudgeEnv,
+    AgenticJudgeEnvConfig,
+    Criterion,
+    JudgeTaskConfig,
+    ScoreConfig,
+)
+
+__all__ = [
+    "AgenticJudgeEnv",
+    "AgenticJudgeEnvConfig",
+    "Criterion",
+    "JudgeTaskConfig",
+    "ScoreConfig",
+]

verifiers-0.2.2.dev26/verifiers/v1/envs/agentic_judge/env.py

@@ -0,0 +1,347 @@
+"""agentic-judge: a solver plays the task, a judge verifies it in the same box.
+
+A reusable env (`--env.id agentic-judge` over any taskset): the box is
+provisioned from the solver's runtime policy, the solver plays the task in it,
+and a code-executing judge then inspects the work as the agent left it, with
+the solver's full trace record uploaded at `/tmp/trace.json`. The judge grades
+rubric criteria (`[env.task]`: policy prompt, criteria file) and writes its
+verdicts to `/tmp/verdict.json`; `finalize()` validates them strictly onto the
+solver's trace — `judge/<name>` metrics plus a weighted-mean `judge` reward,
+composed with the taskset's own rewards via `[env.score]` (judge-only by
+default).
+"""
+
+import json
+import math
+import re
+import tomllib
+from pathlib import Path
+
+from pydantic import field_validator
+
+import verifiers.v1 as vf
+from verifiers.v1.types import StrictBaseModel
+
+VERDICT_FILE = "/tmp/verdict.json"
+TRACE_FILE = "/tmp/trace.json"
+
+GRADE_PROMPT = """\
+You are grading another agent's attempt at a task. Verify the work EMPIRICALLY:
+reconstruct what the agent did from its trace and test it with real execution
+in your sandbox — never take the trace's word for an outcome you can check."""
+
+TASK_SECTION = """\
+## The task the agent was given
+
+{prompt}"""
+
+
+class Criterion(StrictBaseModel):
+    """One rubric criterion — the plugged rubric judge's format, mirrored so the
+    same `criteria` files grade both judges."""
+
+    name: str
+    """Key for the criterion's metric (`judge/<name>`)."""
+    text: str
+    weight: float = 1.0
+    """The criterion's share of the reward."""
+    choices: list[str] = ["no", "yes"]
+    """Allowed answers, ordered **worst → best**: the first scores 0.0, the last 1.0, the rest
+    evenly spaced by rank. Default `["no", "yes"]` is a binary check. Needs >= 2, no duplicates."""
+
+    @field_validator("choices")
+    @classmethod
+    def _check_choices(cls, v: list[str]) -> list[str]:
+        if len(v) < 2:
+            raise ValueError(f"`choices` needs at least two options, got {v}")
+        if len(set(v)) != len(v):
+            raise ValueError(f"`choices` has duplicate options: {v}")
+        return v
+
+
+SOLVED = Criterion(
+    name="solved",
+    text="The task is fully solved: what the task asked for is achieved, and "
+    "you verified it with real execution.",
+)
+
+
+def _verdict_section(criteria: list[Criterion]) -> str:
+    listing = "\n".join(
+        f"- {c.name}: {c.text} (answer one of, worst to best: {', '.join(c.choices)})"
+        for c in criteria
+    )
+    return f"""\
+## Your verdict
+
+Grade the attempt on these criteria:
+
+{listing}
+
+When you are done verifying, write your verdict as JSON to `{VERDICT_FILE}`:
+
+    {{"verdicts": [{{"name": "<criterion name>", "reason": "<one sentence citing \
+what you verified>", "verdict": "<answer>"}}, ...]}}
+
+with one entry per criterion, using each criterion's exact name. For each, first
+write the one-sentence reason grounded in what you actually verified, then set
+verdict to exactly one of the options listed in parentheses after that
+criterion."""
+
+
+def _render(template: str, **fields: str) -> str:
+    """Substitute documented placeholders in one pass over the original template —
+    str.format would crash on any literal brace in a custom prompt, and sequential
+    replaces would re-scan substituted values. An unknown placeholder stays as written."""
+    pattern = re.compile(r"\{(" + "|".join(map(re.escape, fields)) + r")\}")
+    return pattern.sub(lambda m: fields[m.group(1)], template)
+
+
+SANDBOX_NOTE = f"""\
+## Your workspace
+
+Your sandbox is the SAME box the graded agent worked in, in the state the agent
+left it — its edits (and any scoring side effects) are applied. The agent's raw
+trace record (JSON: messages, tool calls, and its `info` artifacts) is uploaded
+at `{TRACE_FILE}`. The record can be very large — never dump it whole; peek
+selectively (list its keys, then slice out specific fields with python or jq)
+and pull only what you need. It is complete — it may also carry the task's own
+scores/metrics and reference material (a gold answer, a reference solution,
+held-out tests). Those are context, not your standard: recorded scores can be
+wrong and references can be narrower than the task; do not over-index on how a
+reference solves it. Your verdict is what YOU verified by execution."""
+
+HINT_SECTION = """\
+## Hints
+
+{hint}"""
+
+
+class JudgeTask(vf.Task):
+    """The judge's verdict task: the solver task's world mirrored onto the minted
+    row, the trace record written (and any stale verdict removed) before the
+    judge starts, verdict scraped off the live box after it exits. `NEEDS_CONTAINER`
+    keeps `Agent.run`'s per-task backstop aligned with the judge's declared need."""
+
+    NEEDS_CONTAINER = True
+
+    def __init__(self, data: vf.TaskData, files: dict[str, bytes]) -> None:
+        super().__init__(data)
+        self.files = files
+
+    @classmethod
+    def from_trace(cls, solution: vf.Trace, config: "JudgeTaskConfig") -> "JudgeTask":
+        """Mint the judge's task from the solver's finished trace."""
+        solved = solution.task.data
+        files = {TRACE_FILE: json.dumps(solution.to_record()).encode()}
+        template = config.build_prompt()
+        body = _render(template, prompt=solved.prompt_text)
+        if "{prompt}" not in template:
+            # A policy that doesn't place the task statement itself still needs it.
+            body += "\n\n" + _render(TASK_SECTION, prompt=solved.prompt_text)
+        sections = [body, _verdict_section(config.criteria()), SANDBOX_NOTE]
+        if (hint := config.build_hint()) is not None:
+            sections.insert(1, _render(HINT_SECTION, hint=hint))
+        prompt = "\n\n".join(sections)
+        return cls(
+            vf.TaskData(
+                idx=solved.idx,
+                prompt=prompt,
+                image=solved.image,
+                workdir=solved.workdir,
+                resources=solved.resources,
+            ),
+            files=files,
+        )
+
+    async def setup(self, trace: vf.Trace, runtime: vf.Runtime) -> None:
+        # The solver had this box first: a pre-seeded verdict must never read as
+        # the judge's own, and a file (or planted symlink) at an upload path must
+        # never survive it — a symlinked TRACE_FILE would redirect the write onto
+        # any file the solver chose.
+        await runtime.run(["rm", "-f", VERDICT_FILE, *self.files], env={})
+        for path, content in self.files.items():
+            await runtime.write(path, content)
+
+    async def finalize(self, trace: vf.Trace, runtime: vf.Runtime) -> None:
+        """Scrape the verdict off the box while it's alive. A judge that wrote no
+        file (or garbage) fails HERE — on the judge's own trace, the retryable
+        unit — never silently."""
+        try:
+            raw = await runtime.read(VERDICT_FILE)
+        except Exception as e:
+            raise ValueError(
+                f"the judge wrote no verdict to {VERDICT_FILE}; its final act must "
+                'be writing {"verdicts": [{"name", "reason", "verdict"}, ...]} there'
+            ) from e
+        trace.info["verdict"] = json.loads(raw)
+
+
+class JudgeTaskConfig(vf.BaseConfig):
+    """The judge's minted task: the grading policy and what lands in its box."""
+
+    prompt: Path | str | None = None
+    """Grading-policy override: inline text, or a policy file (a value ending in
+    `.md`/`.txt` is read from disk). Replaces only the policy body — the verdict
+    contract and workspace note are always appended, so a custom policy cannot
+    break verdict scraping. May reference `{prompt}` (the solver task's prompt);
+    if it doesn't, the task statement is appended after the policy."""
+    hint: Path | str | None = None
+    """Optional hints injected as their own section (inline text or a `.md`/
+    `.txt` file): task-family pointers into the trace or box — e.g. for math,
+    where the reference answer lives in the record; for SWE, to diff the repo
+    or read `info.patch`."""
+    rubric: Path | None = None
+    """Criteria the judge grades against: a `.toml`/`.json` file with a
+    `criteria` list — the plugged rubric judge's format, so the same rubric
+    files work for both. None grades the single built-in `solved` criterion."""
+
+    @staticmethod
+    def _resolve(value: Path | str) -> str:
+        path = Path(value)
+        if isinstance(value, Path) or path.suffix in (".md", ".txt"):
+            return path.read_text(encoding="utf-8")
+        return str(value)
+
+    def build_prompt(self) -> str:
+        if self.prompt is None:
+            return GRADE_PROMPT + "\n\n" + TASK_SECTION
+        return self._resolve(self.prompt)
+
+    def build_hint(self) -> str | None:
+        return self._resolve(self.hint) if self.hint is not None else None
+
+    def criteria(self) -> list[Criterion]:
+        if self.rubric is None:
+            return [SOLVED]
+        text = self.rubric.read_text(encoding="utf-8")
+        data = (
+            tomllib.loads(text)
+            if self.rubric.suffix.lower() == ".toml"
+            else json.loads(text)
+        )
+        items = data.get("criteria", []) if isinstance(data, dict) else data
+        criteria = [Criterion.model_validate(item) for item in items]
+        if not criteria:
+            raise ValueError(f"rubric file '{self.rubric}' lists no criteria")
+        names = [criterion.name for criterion in criteria]
+        if len(set(names)) != len(names):
+            raise ValueError(
+                f"rubric file '{self.rubric}' has duplicate criterion names"
+            )
+        if bad := [c.name for c in criteria if not 0 <= c.weight < math.inf]:
+            raise ValueError(
+                f"rubric '{self.rubric}' has negative or non-finite criterion "
+                f"weights: {bad}"
+            )
+        if sum(criterion.weight for criterion in criteria) <= 0:
+            raise ValueError(f"rubric '{self.rubric}' has no positive criterion weight")
+        return criteria
+
+
+class ScoreConfig(vf.BaseConfig):
+    """How the judge's verdict composes with the taskset's own rewards on the
+    solver's trace. Judge-only by default."""
+
+    task_weight: float = 0.0
+    """Scale applied to the taskset's own rewards; 1 keeps them next to the verdict."""
+    judge_weight: float = 1.0
+    """Weight of the judge's verdict in the solver's reward."""
+
+
+class AgenticJudgeEnvConfig(vf.EnvConfig):
+    solver: vf.AgentConfig = vf.AgentConfig()
+    """The solver agent. It owns the shared box, so its runtime must be a
+    container: `--env.solver.runtime.type docker|prime`."""
+    judge: vf.AgentConfig = vf.AgentConfig()
+    """The judge agent. It plays in the solver's box; its own runtime policy is
+    ignored (overwritten with the solver's)."""
+    task: JudgeTaskConfig = JudgeTaskConfig()
+    score: ScoreConfig = ScoreConfig()
+
+
+class AgenticJudgeEnv(vf.Env[AgenticJudgeEnvConfig]):
+    def __init__(self, config: AgenticJudgeEnvConfig) -> None:
+        # The judge plays in the solver's box, so its effective runtime IS the
+        # solver's policy — aligning the config keeps the base env's subprocess
+        # warning and the runtime stamped on the judge's trace truthful.
+        config.judge = config.judge.model_copy(
+            update={"runtime": config.solver.runtime}
+        )
+        super().__init__(config)
+        self._check_agents()
+        # A missing policy file or a malformed rubric fails here, not mid-episode.
+        config.task.build_prompt()
+        config.task.build_hint()
+        config.task.criteria()
+
+    def _check_agents(self) -> None:
+        """The judge executes real code, never on the host — refuse an impossible
+        pairing at construction, not after burning a full solver run."""
+        judge = self._harnesses["judge"]
+        if not judge.EXECUTES_CODE:
+            raise ValueError(
+                "agentic-judge plays a code-executing judge in its own sandbox, but "
+                f"harness {judge.config.id!r} is a tool-less chat loop — a verdict "
+                "that needs no execution is a plugged judge "
+                "(--env.taskset.task.judges), not an agent."
+            )
+        if isinstance(self.config.solver.runtime, vf.SubprocessConfig):
+            raise ValueError(
+                "agentic-judge plays its judge in the solver's box, but the solver "
+                "(which provisions it) resolves to the subprocess runtime; use "
+                "--env.solver.runtime.type docker or prime"
+            )
+
+    async def setup(self, agents: vf.Agents) -> None:
+        # The judge grades the policy; its tokens are never training data.
+        agents.judge.trainable = False
+
+    async def run(self, task: vf.Task, agents: vf.Agents) -> None:
+        async with agents.solver.provision(task) as box:
+            solution = await agents.solver.run(task, runtime=box)
+            judge_task = JudgeTask.from_trace(solution, self.config.task)
+            await agents.judge.run(judge_task, runtime=box)
+
+    async def finalize(self, task: vf.Task, episode: vf.Episode) -> None:
+        by_agent = {t.agent_name: t for t in episode.traces}
+        solution, verdict = by_agent["solver"], by_agent["judge"]
+        data = verdict.info.get("verdict")
+        if not isinstance(data, dict) or not isinstance(data.get("verdicts"), list):
+            raise ValueError(
+                f"no verdicts on the judge's trace (expected {VERDICT_FILE} with a "
+                '"verdicts" list)'
+            )
+        criteria = self.config.task.criteria()
+        by_criterion = {c.name: c for c in criteria}
+        answers: dict[str, str] = {}
+        for entry in data["verdicts"]:
+            if not isinstance(entry, dict):
+                raise ValueError(f"verdict entry {entry!r} is not an object")
+            name = str(entry.get("name"))
+            if name in answers:
+                # Contradictory duplicates must not collapse to whichever came last.
+                raise ValueError(f"judge answered criterion {name!r} more than once")
+            answers[name] = str(entry.get("verdict"))
+        if sorted(answers) != sorted(by_criterion):
+            raise ValueError(
+                f"judge verdicts name {sorted(answers)}, expected the rubric's "
+                f"{sorted(by_criterion)}"
+            )
+        scores: dict[str, float] = {}
+        for name, answer in answers.items():
+            choices = by_criterion[name].choices
+            # An off-menu answer is a judge failure, not a zero score.
+            if answer not in choices:
+                raise ValueError(
+                    f"judge answered {answer!r} for '{name}', expected one of {choices}"
+                )
+            scores[name] = choices.index(answer) / (len(choices) - 1)
+        for criterion in criteria:
+            solution.record_metric(f"judge/{criterion.name}", scores[criterion.name])
+        if self.config.score.task_weight != 1.0:
+            for name in solution.rewards:
+                solution.rewards[name] *= self.config.score.task_weight
+        total = sum(criterion.weight for criterion in criteria)
+        reward = sum(c.weight * scores[c.name] for c in criteria) / total
+        solution.record_reward("judge", reward, weight=self.config.score.judge_weight)

verifiers-0.2.2.dev25/verifiers/v1/envs/agentic_judge/__init__.py

@@ -1,3 +0,0 @@
-from verifiers.v1.envs.agentic_judge.env import AgenticJudgeEnv, AgenticJudgeEnvConfig
-
-__all__ = ["AgenticJudgeEnv", "AgenticJudgeEnvConfig"]

verifiers-0.2.2.dev25/verifiers/v1/envs/agentic_judge/env.py

@@ -1,174 +0,0 @@
-"""agentic-judge: a solver plays the task, a code-executing judge verifies it in a sandbox.
-
-Agent-as-judge as a reusable env (`--env.id agentic-judge` over any taskset). The
-judge's task mirrors the solver task's world — same image/workdir/resources, in a
-FRESH box in its original state (the solver's runtime is gone by judge time) —
-with the graded transcript uploaded, so the judge reconstructs and tests the work
-empirically, always in its own sandbox, never on the host.
-
-The verdict channel is a file, not the chat: the judge writes
-`{"score": 0-10, "reasoning": ...}` to `/tmp/verdict.json` in its box (a file
-survives a chatty final reply), `JudgeTask.finalize` scrapes it off the live
-runtime onto the judge's trace, and the env's `finalize()` validates it strictly
-onto the solver's trace — a missing, malformed, or off-scale verdict fails loudly instead
-of clamping to full marks.
-"""
-
-import json
-import math
-
-import verifiers.v1 as vf
-from verifiers.v1.harness import Harness
-
-TRANSCRIPT_MD = "/tmp/transcript.md"
-TRANSCRIPT_JSON = "/tmp/transcript.json"
-VERDICT_FILE = "/tmp/verdict.json"
-
-GRADE_PROMPT = f"""\
-You are grading another agent's attempt at a task. Verify the work EMPIRICALLY:
-reconstruct what the agent did from its transcript and test it with real
-execution in your sandbox — never take the transcript's word for an outcome you
-can check.
-
-## The task the agent was given
-
-{{prompt}}
-
-## Your verdict
-
-When you are done verifying, write your verdict as JSON to `{VERDICT_FILE}`:
-
-    {{{{"score": <integer 0-10>, "reasoning": "<one paragraph>"}}}}
-
-10 = the task is fully solved (you verified it); 0 = no progress. The score MUST
-be an integer between 0 and 10 — nothing else is accepted."""
-
-
-def _sandbox_note(solver: vf.TaskData) -> str:
-    """What an agentic judge must know about its box before it starts verifying."""
-    world = (
-        f"a fresh instance of the same environment the graded agent worked in "
-        f"(image {solver.image}), in its ORIGINAL state — the agent's edits are "
-        "NOT applied; reconstruct them from the transcript to verify"
-        if solver.image is not None
-        else "your own — the graded agent worked elsewhere"
-    )
-    return (
-        f"\n\n## Your workspace\nYour sandbox is {world}. The agent's full transcript "
-        f"is uploaded at {TRANSCRIPT_MD} (rendered) and {TRANSCRIPT_JSON} (the raw "
-        "trace record)."
-    )
-
-
-class JudgeTask(vf.Task):
-    """The judge's verdict task: the solver task's world mirrored onto the minted
-    row, transcript uploaded (and any stale verdict removed) before the judge
-    starts, verdict scraped off the live box after it exits. `NEEDS_CONTAINER`
-    keeps `Agent.run`'s per-task backstop aligned with the judge's declared need."""
-
-    NEEDS_CONTAINER = True
-
-    def __init__(self, data: vf.TaskData, files: dict[str, bytes]) -> None:
-        super().__init__(data)
-        self._files = files
-
-    @classmethod
-    def from_trace(cls, task: vf.Task, solution: vf.Trace) -> "JudgeTask":
-        """Mint the judge's task from the solver's finished trace."""
-        prompt = GRADE_PROMPT.format(prompt=task.data.prompt_text)
-        return cls(
-            vf.TaskData(
-                idx=task.data.idx,
-                prompt=prompt + _sandbox_note(task.data),
-                image=task.data.image,
-                workdir=task.data.workdir,
-                resources=task.data.resources,
-            ),
-            files={
-                TRANSCRIPT_MD: solution.transcript.encode(),
-                TRANSCRIPT_JSON: json.dumps(solution.to_record()).encode(),
-            },
-        )
-
-    async def setup(self, trace: vf.Trace, runtime: vf.Runtime) -> None:
-        # A pre-seeded verdict (baked into the image) must never read as the
-        # judge's own; remove it before the judge starts.
-        await runtime.run(["rm", "-f", VERDICT_FILE], env={})
-        for path, content in self._files.items():
-            await runtime.write(path, content)
-
-    async def finalize(self, trace: vf.Trace, runtime: vf.Runtime) -> None:
-        """Scrape the verdict off the box while it's alive. A judge that wrote no
-        file (or garbage) fails HERE — on the judge's own trace, the retryable
-        unit — never silently."""
-        try:
-            raw = await runtime.read(VERDICT_FILE)
-        except Exception as e:
-            raise ValueError(
-                f"the judge wrote no verdict to {VERDICT_FILE}; its final act must "
-                'be writing {"score": <0-10>, "reasoning": ...} there'
-            ) from e
-        trace.info["verdict"] = json.loads(raw)
-
-
-class AgenticJudgeEnvConfig(vf.EnvConfig):
-    solver: vf.AgentConfig = vf.AgentConfig()
-    judge: vf.AgentConfig = vf.AgentConfig()
-    """The judge agent. Its runtime must be a container:
-    `--env.judge.runtime.type docker|prime`."""
-
-
-class AgenticJudgeEnv(vf.Env[AgenticJudgeEnvConfig]):
-    def __init__(self, config: AgenticJudgeEnvConfig) -> None:
-        super().__init__(config)
-        self._check_judge(self._harnesses["judge"], config.judge)
-
-    def _check_judge(self, harness: Harness, judge: vf.AgentConfig) -> None:
-        """The judge executes real code, never on the host — refuse an impossible
-        judge at construction, not after burning a full solver run."""
-        if not harness.EXECUTES_CODE:
-            raise ValueError(
-                "agentic-judge plays a code-executing judge in its own sandbox, but "
-                f"harness {harness.config.id!r} is a tool-less chat loop — a verdict "
-                "that needs no execution is a plugged judge "
-                "(--env.taskset.task.judges), not an agent."
-            )
-        if isinstance(judge.runtime, vf.SubprocessConfig):
-            raise ValueError(
-                "agentic-judge plays its judge in a container (JudgeTask mirrors "
-                "the solver task's image), but the judge resolves to the "
-                "subprocess runtime; use --env.judge.runtime.type docker "
-                "or prime."
-            )
-
-    async def setup(self, agents: vf.Agents) -> None:
-        # The judge grades the policy; its tokens are never training data.
-        agents.judge.trainable = False
-
-    async def run(self, task: vf.Task, agents: vf.Agents) -> None:
-        solution = await agents.solver.run(task)
-        await agents.judge.run(JudgeTask.from_trace(task, solution))
-
-    async def finalize(self, task: vf.Task, episode: vf.Episode) -> None:
-        """Record the scraped verdict on the SOLVER's trace. Strict on scale: an
-        off-scale score raises (a judge answering `95` must not clamp to full
-        marks), failing the episode rather than scoring the solver wrong."""
-        by_agent = {t.agent_name: t for t in episode.traces}
-        solution, verdict = by_agent["solver"], by_agent["judge"]
-        data = verdict.info.get("verdict")
-        if not isinstance(data, dict):
-            raise ValueError(
-                f"no verdict on the judge's trace (expected {VERDICT_FILE})"
-            )
-        score = data.get("score")
-        if (
-            not isinstance(score, (int, float))
-            or isinstance(score, bool)
-            or math.isnan(float(score))
-            or not 0 <= float(score) <= 10
-        ):
-            raise ValueError(
-                f"verdict score {score!r} is not on the 0-10 scale; refusing to "
-                "clamp or coerce it"
-            )
-        solution.record_reward("judge", float(score) / 10.0)