verifiers 0.2.2.dev25__py3-none-any.whl → 0.2.2.dev27__py3-none-any.whl

verifiers/v1/__init__.py

@@ -107,6 +107,7 @@ from verifiers.v1.trace import (
     EvalRunInfo,
     GenerationSpan,
     ModelCall,
+    Reward,
     RunInfo,
     TimeSpan,
     TimeSplit,
@@ -171,6 +172,7 @@ __all__ = [
     "Trace",
     "TraceTask",
     "WireTrace",
+    "Reward",
     "Episode",
     "WireEpisode",
     "TRACE_VERSION",

verifiers/v1/cli/dashboard/eval.py

@@ -341,13 +341,19 @@ def _score_segments(traces: list[Trace], source: str) -> str | None:
         return None
     segments = []
     for name in names:
-        mean = format_mean(
-            traces, lambda t, n=name, s=source: getattr(t, s).get(n, 0.0)
-        )
+        mean = format_mean(traces, lambda t, n=name, s=source: _score(t, s, n))
         segments.append(f"{name} {mean}")
     return "  ·  ".join(segments)
 
 
+def _score(trace: Trace, source: str, name: str) -> float:
+    """Rewards carry raw score + weight; the breakdown shows the raw score."""
+    if source == "rewards":
+        reward = trace.rewards.get(name)
+        return reward.score if reward is not None else 0.0
+    return trace.metrics.get(name, 0.0)
+
+
 def _breakdown(scored: list[Trace], done: list[Trace]) -> Table | None:
     """Score rows read the policy view (`scored` — trainable traces); with several
     roles in play they split per role, each role averaging over its OWN traces (no

verifiers/v1/envs/agentic_judge/__init__.py

@@ -1,3 +1,15 @@
-from verifiers.v1.envs.agentic_judge.env import AgenticJudgeEnv, AgenticJudgeEnvConfig
+from verifiers.v1.envs.agentic_judge.env import (
+    AgenticJudgeEnv,
+    AgenticJudgeEnvConfig,
+    Criterion,
+    JudgeTaskConfig,
+    ScoreConfig,
+)
 
-__all__ = ["AgenticJudgeEnv", "AgenticJudgeEnvConfig"]
+__all__ = [
+    "AgenticJudgeEnv",
+    "AgenticJudgeEnvConfig",
+    "Criterion",
+    "JudgeTaskConfig",
+    "ScoreConfig",
+]

verifiers/v1/envs/agentic_judge/env.py

@@ -1,100 +1,166 @@
-"""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.
+"""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.harness import Harness
+from verifiers.v1.types import StrictBaseModel
 
-TRANSCRIPT_MD = "/tmp/transcript.md"
-TRANSCRIPT_JSON = "/tmp/transcript.json"
 VERDICT_FILE = "/tmp/verdict.json"
+TRACE_FILE = "/tmp/trace.json"
 
-GRADE_PROMPT = f"""\
+GRADE_PROMPT = """\
 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.
+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}}
+{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}`:
 
-    {{{{"score": <integer 0-10>, "reasoning": "<one paragraph>"}}}}
+    {{"verdicts": [{{"name": "<criterion name>", "reason": "<one sentence citing \
+what you verified>", "verdict": "<answer>"}}, ...]}}
 
-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."""
+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 _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)."
-    )
+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, transcript uploaded (and any stale verdict removed) before the judge
-    starts, verdict scraped off the live box after it exits. `NEEDS_CONTAINER`
+    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
+        self.files = files
 
     @classmethod
-    def from_trace(cls, task: vf.Task, solution: vf.Trace) -> "JudgeTask":
+    def from_trace(cls, solution: vf.Trace, config: "JudgeTaskConfig") -> "JudgeTask":
         """Mint the judge's task from the solver's finished trace."""
-        prompt = GRADE_PROMPT.format(prompt=task.data.prompt_text)
+        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=task.data.idx,
-                prompt=prompt + _sandbox_note(task.data),
-                image=task.data.image,
-                workdir=task.data.workdir,
-                resources=task.data.resources,
+                idx=solved.idx,
+                prompt=prompt,
+                image=solved.image,
+                workdir=solved.workdir,
+                resources=solved.resources,
             ),
-            files={
-                TRANSCRIPT_MD: solution.transcript.encode(),
-                TRANSCRIPT_JSON: json.dumps(solution.to_record()).encode(),
-            },
+            files=files,
         )
 
     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():
+        # 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:
@@ -106,39 +172,125 @@ class JudgeTask(vf.Task):
         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'
+                '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. Its runtime must be a container:
-    `--env.judge.runtime.type docker|prime`."""
+    """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_judge(self._harnesses["judge"], config.judge)
+        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_judge(self, harness: Harness, judge: vf.AgentConfig) -> None:
+    def _check_agents(self) -> 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:
+        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 {harness.config.id!r} is a tool-less chat loop — a verdict "
+                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(judge.runtime, vf.SubprocessConfig):
+        if isinstance(self.config.solver.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."
+                "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:
@@ -146,29 +298,50 @@ class AgenticJudgeEnv(vf.Env[AgenticJudgeEnvConfig]):
         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 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:
-        """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):
+        if not isinstance(data, dict) or not isinstance(data.get("verdicts"), list):
             raise ValueError(
-                f"no verdict on the judge's trace (expected {VERDICT_FILE})"
+                f"no verdicts on the judge's trace (expected {VERDICT_FILE} with a "
+                '"verdicts" list)'
             )
-        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
-        ):
+        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"verdict score {score!r} is not on the 0-10 scale; refusing to "
-                "clamp or coerce it"
+                f"judge verdicts name {sorted(answers)}, expected the rubric's "
+                f"{sorted(by_criterion)}"
             )
-        solution.record_reward("judge", float(score) / 10.0)
+        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 reward in solution.rewards.values():
+                reward.weight *= 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/v1/legacy.py

@@ -37,6 +37,7 @@ from verifiers.v1.trace import (
     Error,
     GenerationSpan,
     ModelCall,
+    Reward,
     TimeSpan,
     TimeSplit,
     Timing,
@@ -270,7 +271,7 @@ def rollout_output_to_trace(out: dict, task_idx: int) -> Trace:
             data=_to_wire_task(task_idx, out.get("prompt"), out.get("answer")),
         ),
         tools=_to_v1_tools(out.get("tool_defs")),
-        rewards={"reward": float(out.get("reward") or 0.0)},
+        rewards={"reward": Reward(score=float(out.get("reward") or 0.0))},
         metrics={k: float(v) for k, v in (out.get("metrics") or {}).items()},
         info=dict(out.get("info") or {}),
         is_completed=bool(out.get("is_completed", True)),

verifiers/v1/push.py

@@ -90,9 +90,10 @@ def trace_to_sample(
         else None,
         "info": dict(trace.info) or None,
     }
-    # Flatten sub-rewards to top-level keys the way v0 does; env metrics stay nested.
-    for name, value in trace.rewards.items():
-        sample.setdefault(name, value)
+    # Flatten sub-rewards to top-level keys the way v0 does (raw scores, as v0's
+    # per-function outputs were); env metrics stay nested.
+    for name, reward in trace.rewards.items():
+        sample.setdefault(name, reward.score)
     return sample
 
 
@@ -127,7 +128,8 @@ def _run_metrics(episodes: list[Episode], traces: list[Trace]) -> dict[str, Any]
     sums: dict[str, float] = {}
     counts: dict[str, int] = {}
     for trace in scored:
-        for name, value in {**trace.rewards, **trace.metrics}.items():
+        scores = {name: reward.score for name, reward in trace.rewards.items()}
+        for name, value in {**scores, **trace.metrics}.items():
             sums[name] = sums.get(name, 0.0) + value
             counts[name] = counts.get(name, 0) + 1
     n = len(scored)

verifiers/v1/trace.py

@@ -266,7 +266,7 @@ _NODE_DUMP_EXCLUDE: dict = {
 """Raw tensor fields kept on the msgpack wire but excluded from JSON records."""
 
 
-TRACE_VERSION = 3
+TRACE_VERSION = 4
 """Version of the trace record schema (see `Trace.model_json_schema()`). Bumped on
 breaking shape changes; optional-with-default fields are additive and don't bump it."""
 
@@ -343,6 +343,21 @@ class TraceTask(StrictBaseModel, Generic[DataT]):
     """The (immutable) row being solved."""
 
 
+class Reward(StrictBaseModel):
+    """One named reward as recorded on the trace: the raw score next to its weight,
+    so records keep both readable and the weighted sum stays a derived view."""
+
+    score: float
+    """The raw value the reward function returned, unweighted."""
+    weight: float = 1.0
+    """The multiplier `score` carries in the trace-level `reward` sum."""
+
+    @property
+    def value(self) -> float:
+        """This reward's weighted contribution to the trace-level `reward`."""
+        return self.score * self.weight
+
+
 class Trace(StrictBaseModel, Generic[DataT, StateT, AgentConfigT]):
     id: str = Field(default_factory=lambda: uuid.uuid4().hex)
     """Unique id for this rollout, auto-generated per trace."""
@@ -369,8 +384,9 @@ class Trace(StrictBaseModel, Generic[DataT, StateT, AgentConfigT]):
     """Every provider exchange behind the sampled turns, in order: raw wire request/response
     plus per-call timing and errors, linked into `nodes` via `ModelCall.node`."""
 
-    rewards: dict[str, float] = Field(default_factory=dict)
-    """Weighted contributions from task rewards, judges, and the env's `score()`."""
+    rewards: dict[str, Reward] = Field(default_factory=dict)
+    """Named rewards from tasks, judges, and the env's `score()` — each keeps its
+    raw `score` and `weight`; the trace-level `reward` is their weighted sum."""
     metrics: dict[str, float] = Field(default_factory=dict)
     """Unweighted metrics from tasks, harnesses, and judges."""
     info: dict[str, Any] = Field(default_factory=dict)
@@ -400,7 +416,7 @@ class Trace(StrictBaseModel, Generic[DataT, StateT, AgentConfigT]):
 
     @property
     def reward(self) -> float:
-        return sum(self.rewards.values())
+        return sum(r.value for r in self.rewards.values())
 
     @property
     def error(self) -> Error | None:
@@ -559,12 +575,12 @@ class Trace(StrictBaseModel, Generic[DataT, StateT, AgentConfigT]):
             self.extra_usage.append(response.usage)
 
     def record_reward(self, name: str, value: float, weight: float = 1.0) -> None:
-        contribution = float(value) * float(weight)
+        reward = Reward(score=float(value), weight=float(weight))
         if name in self.rewards:
             logger.warning(
-                "reward %r overridden: %s -> %s", name, self.rewards[name], contribution
+                "reward %r overridden: %s -> %s", name, self.rewards[name], reward
             )
-        self.rewards[name] = contribution
+        self.rewards[name] = reward
 
     def stamp(self, run: RunInfo | None = None, **info: Any) -> None:
         """Stamp identity only the consumer knows (the eval CLI / a trainer) onto the

{verifiers-0.2.2.dev25.dist-info → verifiers-0.2.2.dev27.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev25
+Version: 0.2.2.dev27
 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.dist-info → verifiers-0.2.2.dev27.dist-info}/RECORD

@@ -167,7 +167,7 @@ verifiers/utils/threaded_sandbox_client.py,sha256=Pbr8MA4FDPEitL6z88S8T1qLJOmtXN
 verifiers/utils/tool_utils.py,sha256=gInWZODWQUZN4TyEMuuITyOUm9qlHJxtcvr1zZl_zJs,1020
 verifiers/utils/usage_utils.py,sha256=GPLC0xGY_Obrr8X7huWY-2iODZ7tgX-MtO8WSDN4rXI,3904
 verifiers/utils/version_utils.py,sha256=am3hZLnlaUWTFdllGZR1VMN91hEhiiE8FDmPf1cOG1k,2642
-verifiers/v1/__init__.py,sha256=29oq_Ng33Shgo2x7OGciV7SQrN0ndB8eBxsbebCggqw,6833
+verifiers/v1/__init__.py,sha256=z0yUgI1iVf8znsAtBG2fpRKuwipaQ5yS4PvGhWJbKtE,6859
 verifiers/v1/agent.py,sha256=FP38GsZXuEJe3tJ-U2THpwBf9NWQeo847hWRZoeggNw,31956
 verifiers/v1/decorators.py,sha256=XRMkUQSyvXCYP5fOwzBYV5qEOlxLg6rzYhnhrHHHTIQ,3838
 verifiers/v1/env.py,sha256=w6sHWdWijLRZzhwVyJjadnGtziNqJAPCeMUoNdzHoLg,17712
@@ -176,9 +176,9 @@ verifiers/v1/errors.py,sha256=Pj5Om8x1fP3TDPQQn6nUCoSPoZRsy2JeBz8pXhpPrDY,6883
 verifiers/v1/graph.py,sha256=Mivq5jICUcyK-fhomFTwP4fQn688MXg6-hmdf03aC4Y,27703
 verifiers/v1/harness.py,sha256=Ukzhk7wxSUSVnN4tRAUMfniN9R9MBSNzs2SiT5YcRzQ,10711
 verifiers/v1/judge.py,sha256=ZWvr6uCniSwqzzjKrlyh-lzfCh5q-VyB16W42x0aWlg,9449
-verifiers/v1/legacy.py,sha256=b53EDDOImu8rSaZR_ZE7jNaQTJ6aIw5T8hxlXU4fTh0,22280
+verifiers/v1/legacy.py,sha256=8eVGhutQEgJG4qabhph4Xs1VzTWvmMxMn3-Zk-6ViWM,22306
 verifiers/v1/loaders.py,sha256=FdICekH_c9WYkUe0XJCI9AWxL1sJjjzzJPJ7ZJdeyRk,9962
-verifiers/v1/push.py,sha256=N8UNz0eHKVH_vgh6ZuN-efqPOgYtscQBzn8zz-JYmYw,11138
+verifiers/v1/push.py,sha256=VfLESKAlL8WXzfXx6CrqwMJ-CvPTkP-tUf6Rin12V8w,11273
 verifiers/v1/retries.py,sha256=ZQxY6R_FoXooERmhIMZnhY3b2koJOVjaIvkSO9gw53E,5379
 verifiers/v1/rollout.py,sha256=8_LC938Ws9uWFRPz2DDpivttqDALb3Yir7J5wjnsoIE,20332
 verifiers/v1/scoring.py,sha256=I_mtqhTZ195hj2-CB5jzr3BY5GFKgNEvTf331f7Is4k,5740
@@ -186,7 +186,7 @@ verifiers/v1/session.py,sha256=J6ZPuvsg6vMTJAs9dMUMtfdjCuvp6Hom877DVdKFKlY,6912
 verifiers/v1/state.py,sha256=R8tyQv8nsFV2pztrquDqCOa1Mk1fAp3w2GjqUugZwE0,689
 verifiers/v1/task.py,sha256=qKAAF154ykgvrms2Y2Or6kgGVPvBMsFCD6KqX2QOQaU,10912
 verifiers/v1/taskset.py,sha256=PX1-skAVhSpGF07qjxcG0Ctmq1LiMftMxxok-uc8qRY,3939
-verifiers/v1/trace.py,sha256=J_g3DgruE8B747fjkmhiI9UE8KIesf9mqAANMNo6Z1E,25976
+verifiers/v1/trace.py,sha256=OE7vW6sYGjA18JdRJuHM9UiRaq9XkGS0Yv2m7nuvEYs,26619
 verifiers/v1/types.py,sha256=5tZyG4r4bLJ9a14oybHy7bSCpt5X5A17eRwmIszuRo8,9118
 verifiers/v1/acp/__init__.py,sha256=9dwH6fLopNndRmcpRSYC8ghzzMquaLfgh645QM-Dwic,2189
 verifiers/v1/acp/_runner.py,sha256=FqEnHR0Q_YrCt_EuaJqiqTpd2_kFa41FqcEunwOhvTI,6856
@@ -201,7 +201,7 @@ verifiers/v1/cli/serve.py,sha256=VHzcr2bM8R3XoGEu6WvY1NSQoiBZsk49cVW5AB9N4Kg,266
 verifiers/v1/cli/validate.py,sha256=r3ByzcLc_DFkvlCB7NanYRnBSOIi5EU9cUANy74l3T0,9547
 verifiers/v1/cli/dashboard/__init__.py,sha256=v-baMxQuWxOCsbU7-p_jj2Q9BUnTN-TWi1q2hK6rU2s,198
 verifiers/v1/cli/dashboard/base.py,sha256=kUP93zJSIVLptSbnWX6MOy8kGg63fpeAzPqakCpPyDc,3547
-verifiers/v1/cli/dashboard/eval.py,sha256=F--kqnwjeihY1IrQ55liMwpQJSIkcp1RZmZzhKLhXVY,34081
+verifiers/v1/cli/dashboard/eval.py,sha256=nmaWIq-4ormHZWjwlWm1QSFo383Sc4VFEkw9rYnpi00,34357
 verifiers/v1/cli/dashboard/replay.py,sha256=CvRVaf0dUum5v0IzPQbFPPikBQmTrkMeaunGfVMTWHc,2755
 verifiers/v1/cli/dashboard/validate.py,sha256=rLsQ_31DjTIZpC_VO7zSG41ncUEBxqCx8mL9nTnGiMc,3650
 verifiers/v1/cli/eval/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -235,8 +235,8 @@ verifiers/v1/dialects/base.py,sha256=YZQDnasQseDXK-GHLfJLZgyHIIygS80YhD_sTjmUbfA
 verifiers/v1/dialects/chat.py,sha256=DFvjmIW86jZUMWuz-hOqzxC6tiiFibiAzNEJ3JkpvjU,13998
 verifiers/v1/dialects/responses.py,sha256=vVgeT7-aWjtfED2IKlDwTATKn_AQZruHNL6mkFTTZTU,13548
 verifiers/v1/envs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-verifiers/v1/envs/agentic_judge/__init__.py,sha256=p43dmdpO484Euu9urdUQ1kNaX_hJOTC2YDVcjPpjmQs,143
-verifiers/v1/envs/agentic_judge/env.py,sha256=jzoKZQzX8yFKomQabbx02YmA0et_1GybOHJIYnAhS6o,7523
+verifiers/v1/envs/agentic_judge/__init__.py,sha256=X7vQbbbfmJ_j-mJEfBkuB4a8QUPkUYkpG7yp0xYOEmg,279
+verifiers/v1/envs/agentic_judge/env.py,sha256=_zEKG9i3LsOVIQWULXYAHpbmrQNPWBRayJx2yjvtPbQ,15307
 verifiers/v1/envs/best_of_n/__init__.py,sha256=mubASiwXMNIhnQFDccNAc7yZ0OFsOVK8hH8WIx9b2A4,119
 verifiers/v1/envs/best_of_n/env.py,sha256=KEUJBCVqblnXP_lEwm3mjGdaikBMkecz8UjOFPBhxc0,1987
 verifiers/v1/envs/single_agent/__init__.py,sha256=r0edAc_6gtHBhPzeBGVcdHZZqSdCPZnLMje6eVgyz5U,138
@@ -327,8 +327,8 @@ verifiers/v1/utils/logging.py,sha256=OcMHA6NsYux3oIzjPuI95rDWmFBHNcHDjZeNIhXTX-Y
 verifiers/v1/utils/memory.py,sha256=ZkIvGk6uITAH5sKon65LifKPbvZr8mJ__PVc23FZpOQ,1835
 verifiers/v1/utils/sampling.py,sha256=JczGzBn6s3wsIrSY2Hy3hE8m6NreNgsNzhSjDikQqX0,1037
 verifiers/v1/utils/version.py,sha256=75ZtI2NHBmlb52KpcKLUpiSXp8q4bASr7uKeXoCKlT8,1582
-verifiers-0.2.2.dev25.dist-info/METADATA,sha256=VBIg7f-noyBD7tjlgGM2pvAlM2SKGcvJRycLWfPXarA,4540
-verifiers-0.2.2.dev25.dist-info/WHEEL,sha256=lCkmxWfQsSc9CfIClYeavTdQeEX2toPqufh9gI35EQA,87
-verifiers-0.2.2.dev25.dist-info/entry_points.txt,sha256=dF82JUYEFslR1AOW8LZfuBWeZeQoyX4wCRrduklg8-I,551
-verifiers-0.2.2.dev25.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.2.2.dev25.dist-info/RECORD,,
+verifiers-0.2.2.dev27.dist-info/METADATA,sha256=JnRcvyGkm0iogzBO_JnbUyqr6DO1_EUfDXi-TCV6ezg,4540
+verifiers-0.2.2.dev27.dist-info/WHEEL,sha256=lCkmxWfQsSc9CfIClYeavTdQeEX2toPqufh9gI35EQA,87
+verifiers-0.2.2.dev27.dist-info/entry_points.txt,sha256=dF82JUYEFslR1AOW8LZfuBWeZeQoyX4wCRrduklg8-I,551
+verifiers-0.2.2.dev27.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.2.2.dev27.dist-info/RECORD,,