Harden redaction-pressure reward and provenance

.gitea/workflows/ci.yml

@@ -0,0 +1,31 @@
+name: arena-environments
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+  workflow_dispatch:
+
+jobs:
+  redaction-pressure:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python: ["3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v6
+        with:
+          version: "0.12.5"
+          enable-cache: false
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+      - name: Verify lockfile
+        run: uv lock --project environments/redaction_pressure --check
+      - name: Install locked environment
+        run: uv sync --project environments/redaction_pressure --frozen
+      - name: Run environment tests
+        run: uv run --project environments/redaction_pressure python -m unittest discover -s environments/redaction_pressure/tests -v
+      - name: Verify documented root entrypoint
+        run: uv run --project environments/redaction_pressure eval --help >/dev/null

README.md

@@ -24,18 +24,27 @@ Three rules follow, and every environment in this repository is held to them:
 
 | | |
 |---|---|
-| [`redaction-pressure`](environments/redaction_pressure) | Write redaction rules for a ticket corpus. Graded on the next records the generator would have produced. Every secret class is shape-matched to a decoy, so a rule keyed on shape earns on one and pays on the other. Reward is a scanner — no judge model, no runtime, milliseconds per rollout. |
+| [`redaction-pressure`](environments/redaction_pressure) | Write redaction rules for a ticket corpus. Source-character provenance makes partial edits residual secrets, while the clean gate requires character-for-character preservation of all non-secret source content. Hostile regexes share one episode budget. |
 
 ## Running one
 
 ```bash
-uv run eval @ configs/redaction_pressure.toml --model <model-id>
+uv run --project environments/redaction_pressure eval @ configs/redaction_pressure.toml --model <model-id>
+```
+
+Run its release gate from the repository root:
+
+```bash
+uv lock --project environments/redaction_pressure --check
+uv sync --project environments/redaction_pressure --frozen
+uv run --project environments/redaction_pressure \
+  python -m unittest discover -s environments/redaction_pressure/tests -v
 ```
 
 ## Why these, and not benchmark reimplementations
 
 The Hub is filling with reimplementations, which are worth doing and are not what we have.
-What Lumbridge has is a set of tools that already do verifiable work — Bench's
+What Lumbridge has is a set of tools aimed at verifiable work. Those gates make good
-contamination gate, Forge's signed build chain — and those gates make good rewards
+rewards only after their ceiling, floor, adversarial cases, and replay identity are
-precisely because they were built to be un-arguable about a machine. `redaction-pressure`
+committed beside the environment. `redaction-pressure` is the first small contract for
-is Forge's `redact` stage with the exit code turned into a gradient.
+that path; Bench/Forge integration remains future work rather than an implied dependency.

configs/redaction_pressure.toml

@@ -6,6 +6,7 @@ num_rollouts = 2
 
 [env.taskset]
 id = "redaction-pressure"
+split = "dev"   # public deterministic development split; never use it as a private holdout
 visible = 4    # records the model reads
 held_out = 12  # records it is scored on
 

environments/redaction_pressure/README.md

@@ -0,0 +1,52 @@
+# redaction-pressure v0.2
+
+A one-turn Verifiers v1 environment. The model returns regex rules; the scanner runs
+them against later records from the same deterministic corpus stream.
+
+## Reward contract
+
+- **Recall (0.35):** a labelled secret counts only when every original character in
+  its span is gone. Partial edits earn zero span recall.
+- **Precision (0.35):** source characters belonging to fully removed secret spans,
+  divided by all removed source characters. Partial damage stays in the denominator.
+- **Clean gate (0.30):** every secret fully removed, every non-secret character and
+  decoy preserved, and no validation or timeout errors.
+- **Validity:** malformed JSON, bad entries, more than 24 rules, unsafe replacements,
+  invalid regexes, excessive matches, timeouts, or an exhausted episode budget multiply
+  every reward by zero.
+
+Replacement is intentionally limited to `"[REDACTED]"` or `""`. That prevents a rule
+from matching a complete secret and then leaking it through a capture expansion.
+
+## Splits
+
+`train` and `dev` have separate, deterministic public seed ranges. `private_eval` derives
+seeds with HMAC from `REDACTION_PRESSURE_PRIVATE_SEED`, which must contain at least 32
+characters. The secret is never task data; traces carry its SHA-256 commitment and reveal
+the derived task seed only in the completed trace so the episode can be reproduced by an
+authorized evaluator.
+
+Never present public `dev` scores as private evaluation.
+
+## Trace identity
+
+Each task records the environment, corpus, and reward-schema versions, split commitment,
+and a SHA-256 digest of the behavior-bearing installed Python sources. Scoring repeats
+the identity under `trace.info.environment` and records validation failures under
+`trace.info.redaction_errors`.
+
+## Checks
+
+From the Arena repository root:
+
+```bash
+uv lock --project environments/redaction_pressure --check
+uv sync --project environments/redaction_pressure --frozen
+uv run --project environments/redaction_pressure \
+  python -m unittest discover -s environments/redaction_pressure/tests -v
+```
+
+The committed tests demonstrate oracle 1.0 and inaction 0.0, reject the historical
+first-character exploit across all 64 public development tasks, exercise parser and
+rule validation, prove deterministic/private split behavior, and interrupt hostile
+regexes under the shared episode budget.

environments/redaction_pressure/pyproject.toml

@@ -1,12 +1,16 @@
 [project]
 name = "redaction-pressure"
-version = "0.1.0"
+version = "0.2.0"
-description = "redaction-pressure — write redaction rules graded on records the author never saw."
+description = "Provenance-safe redaction rules graded on records the author never saw."
 requires-python = ">=3.11"
-dependencies = ["verifiers", "regex"]
+dependencies = [
+  "pydantic==2.13.4",
+  "regex==2026.7.19",
+  "verifiers==0.3.0",
+]
 
 [build-system]
-requires = ["hatchling"]
+requires = ["hatchling>=1.27,<2"]
 build-backend = "hatchling.build"
 
 [tool.hatch.build.targets.wheel]

environments/redaction_pressure/redaction_pressure/__init__.py

@@ -1,3 +1,4 @@
+from redaction_pressure.provenance import ENVIRONMENT_VERSION
 from redaction_pressure.taskset import RedactionTaskset
 
-__all__ = ["RedactionTaskset"]
+__all__ = ["ENVIRONMENT_VERSION", "RedactionTaskset"]

environments/redaction_pressure/redaction_pressure/provenance.py

@@ -0,0 +1,37 @@
+"""Stable environment identity recorded with every task and trace."""
+
+from __future__ import annotations
+
+import hashlib
+from functools import lru_cache
+from pathlib import Path
+
+ENVIRONMENT_VERSION = "0.2.0"
+CORPUS_VERSION = "redaction-pressure-corpus-v2"
+REWARD_SCHEMA_VERSION = "redaction-pressure-reward-v2"
+SOURCE_FILES = ("corpus.py", "provenance.py", "scan.py", "taskset.py")
+
+
+@lru_cache(maxsize=1)
+def environment_source_sha256() -> str:
+    """Hash the behavior-bearing installed sources, independent of checkout paths."""
+    root = Path(__file__).resolve().parent
+    digest = hashlib.sha256()
+    for name in SOURCE_FILES:
+        payload = (root / name).read_bytes()
+        digest.update(name.encode("utf-8"))
+        digest.update(b"\0")
+        digest.update(len(payload).to_bytes(8, "big"))
+        digest.update(payload)
+    return digest.hexdigest()
+
+
+def trace_provenance(*, split: str, seed_commitment: str) -> dict[str, str]:
+    return {
+        "environment_version": ENVIRONMENT_VERSION,
+        "corpus_version": CORPUS_VERSION,
+        "reward_schema_version": REWARD_SCHEMA_VERSION,
+        "source_sha256": environment_source_sha256(),
+        "split": split,
+        "seed_commitment": seed_commitment,
+    }

environments/redaction_pressure/redaction_pressure/scan.py

@@ -1,37 +1,33 @@
-"""Applying the agent's rules, and measuring what they did.
+"""Apply hostile redaction rules while preserving source-character provenance.
 
-Everything is counted as *removals*, and that choice is the whole reward.
+The scorer never infers redaction from an output substring disappearing.  Every
+character starts with its source offset, replacements carry no source offsets, and a
+secret counts as removed only when *all* offsets in its labelled span are gone.  This
+makes deleting or changing one byte of a secret worth zero recall.  The same provenance
+requires every ordinary and decoy character to survive for the clean gate.
 
-The obvious framing — secrets caught, decoys survived, prose kept — pays a model that
+Agent regexes run under one wall-clock budget shared by compilation and every record.
-returns no rules at all: it destroys nothing, so it keeps everything. Measured that way
+Any invalid, excessive, unsafe, over-broad, or timed-out rule invalidates the complete
-an empty ruleset and a ruleset that shreds the corpus scored identically, and neither
+ruleset; valid rules are still measured so the trace explains what happened, but reward
-told a trainer anything. Counting removals fixes it, because a model that removes nothing
+methods apply ``validity`` and therefore pay zero for an invalid submission.
-has no true positives and therefore no precision.
-
-  removed_secrets   true positives — a seeded secret gone from the held-out text.
-  residual_hits     the ones still there. Forge's rule applies to the gate: a residual
-                    hit is exit 1, not a deduction.
-  collateral_hits   decoys destroyed. Each is shape-matched to a secret class, so this is
-                    what a rule keyed on shape pays and one keyed on meaning does not.
-  innocent_lost     ordinary words destroyed. The floor that stops `\\S+ -> [X]`, which
-                    otherwise has perfect recall.
-
-The agent supplies the patterns, so they are hostile input: `regex` is used rather than
-`re` for its `timeout=`, since `re` cannot be interrupted and one nested quantifier would
-otherwise hang a rollout until the harness killed it.
 """
 
 from __future__ import annotations
 
-from dataclasses import dataclass
+import time
+from dataclasses import dataclass, field
+from typing import Callable, Iterable
 
 import regex
 
-from redaction_pressure.corpus import Record
+from redaction_pressure.corpus import Record, Span
 
 MAX_RULES = 24
 MAX_PATTERN_CHARS = 200
-PATTERN_TIMEOUT_S = 0.25
+MAX_MATCHES_PER_RULE_RECORD = 128
+PER_OPERATION_TIMEOUT_S = 0.025
+EPISODE_REGEX_BUDGET_S = 0.75
+ALLOWED_REPLACEMENTS = frozenset({"", "[REDACTED]"})
 
 
 @dataclass(frozen=True)
@@ -40,97 +36,295 @@ class Rule:
     replacement: str
 
 
+@dataclass(frozen=True)
+class ParsedRules:
+    rules: tuple[Rule, ...]
+    errors: tuple[str, ...] = ()
+    submitted: int = 0
+
+
+@dataclass(frozen=True)
+class LocatedSpan:
+    kind: str
+    text: str
+    start: int
+    end: int
+
+    @property
+    def offsets(self) -> range:
+        return range(self.start, self.end)
+
+
+@dataclass(frozen=True)
+class TrackedText:
+    text: str
+    origins: tuple[int | None, ...]
+
+    @classmethod
+    def original(cls, text: str) -> "TrackedText":
+        return cls(text=text, origins=tuple(range(len(text))))
+
+
+@dataclass(frozen=True)
+class CompiledRule:
+    index: int
+    pattern: regex.Pattern
+    replacement: str
+
+
+@dataclass
+class RegexBudget:
+    seconds: float = EPISODE_REGEX_BUDGET_S
+    clock: Callable[[], float] = time.monotonic
+    started: float = field(init=False)
+
+    def __post_init__(self) -> None:
+        self.started = self.clock()
+
+    @property
+    def elapsed(self) -> float:
+        return max(0.0, self.clock() - self.started)
+
+    @property
+    def remaining(self) -> float:
+        return max(0.0, self.seconds - self.elapsed)
+
+    def operation_timeout(self) -> float:
+        remaining = self.remaining
+        if remaining <= 0:
+            raise RegexBudgetExceeded("episode regex budget exhausted")
+        return min(PER_OPERATION_TIMEOUT_S, remaining)
+
+
+class RegexBudgetExceeded(RuntimeError):
+    pass
+
+
+@dataclass(frozen=True)
+class ApplyResult:
+    tracked: TrackedText
+    errors: tuple[str, ...]
+
+
 @dataclass
 class Outcome:
-    """What one ruleset did to one slice."""
+    """Measurements for one submitted ruleset over one held-out slice."""
 
+    removed_secrets: int
+    partial_secrets: int
     residual_hits: int
     secrets_total: int
+    credited_secret_chars: int
+    removed_secret_chars: int
+    secret_chars_total: int
     collateral_hits: int
     decoys_total: int
     innocent_lost: int
     innocent_total: int
     invalid_rules: list[str]
-
+    budget_seconds: float
-    @property
+    budget_elapsed_seconds: float
-    def removed_secrets(self) -> int:
-        return self.secrets_total - self.residual_hits
 
     @property
     def recall(self) -> float:
+        """All-or-nothing span recall; partial secret edits earn no recall."""
         return self.removed_secrets / max(self.secrets_total, 1)
 
     @property
-    def precision(self) -> float:
+    def character_coverage(self) -> float:
-        """Of everything this ruleset destroyed, how much of it should have been.
+        """Diagnostic only: never used as reward because partial deletion is unsafe."""
+        return self.removed_secret_chars / max(self.secret_chars_total, 1)
 
-        Zero for an empty ruleset — it destroyed nothing, so none of what it destroyed
+    @property
-        was a secret. That is the case the survivor-counting version got wrong.
+    def precision(self) -> float:
+        """Of removed source characters, the share in fully removed secret spans.
+
+        Characters removed from a merely damaged secret stay in the denominator but are
+        not true positives. This keeps the historical one-byte exploit at precision zero.
         """
-        destroyed = self.removed_secrets + self.collateral_hits + self.innocent_lost
+        destroyed = self.removed_secret_chars + self.innocent_lost
-        return self.removed_secrets / destroyed if destroyed else 0.0
+        return self.credited_secret_chars / destroyed if destroyed else 0.0
+
+    @property
+    def ordinary_preservation(self) -> float:
+        return 1.0 - self.innocent_lost / max(self.innocent_total, 1)
+
+    @property
+    def validity(self) -> float:
+        """Binary by design: invalid hostile input is never a partially paid policy."""
+        return float(not self.invalid_rules)
 
     @property
     def clean(self) -> bool:
-        """The gate: every secret gone AND no decoy taken with it — the ruleset you could
-        actually ship. `forge redact`'s exit 1, with the collateral clause it implies."""
         return (
-            self.residual_hits == 0
+            self.removed_secrets == self.secrets_total
+            and self.partial_secrets == 0
             and self.collateral_hits == 0
+            and self.innocent_lost == 0
             and self.secrets_total > 0
+            and self.validity == 1.0
         )
 
 
-def compile_rules(rules: list[Rule]) -> tuple[list[tuple[regex.Pattern, str]], list[str]]:
+def compile_rules(
-    """Compile what compiles; report the rest rather than failing the rollout.
+    parsed: ParsedRules,
-
+    budget: RegexBudget,
-    A rollout that dies on one bad pattern teaches nothing — the agent's score should
+) -> tuple[list[CompiledRule], list[str]]:
-    reflect that the rule did no work, which is what dropping it does.
+    """Compile bounded rules and retain every validation error."""
-    """
+    errors = list(parsed.errors)
-    compiled, invalid = [], []
+    rules = list(parsed.rules)
-    for rule in rules[:MAX_RULES]:
+    if len(rules) > MAX_RULES:
+        errors.append(f"too many rules: received {len(rules)}, maximum is {MAX_RULES}")
+    compiled: list[CompiledRule] = []
+    for index, rule in enumerate(rules[:MAX_RULES]):
+        label = f"rule[{index}]"
+        if not rule.pattern:
+            errors.append(f"{label}: pattern must not be empty")
+            continue
         if len(rule.pattern) > MAX_PATTERN_CHARS:
-            invalid.append(f"{rule.pattern[:40]}…: over {MAX_PATTERN_CHARS} characters")
+            errors.append(f"{label}: pattern exceeds {MAX_PATTERN_CHARS} characters")
+            continue
+        if rule.replacement not in ALLOWED_REPLACEMENTS:
+            errors.append(
+                f"{label}: replacement must be empty or [REDACTED]; captures and arbitrary text are unsafe"
+            )
             continue
         try:
-            compiled.append((regex.compile(rule.pattern), rule.replacement))
+            budget.operation_timeout()
+            pattern = regex.compile(rule.pattern)
+            budget.operation_timeout()
+        except RegexBudgetExceeded as exc:
+            errors.append(f"{label}: {exc}")
+            break
         except regex.error as exc:
-            invalid.append(f"{rule.pattern[:40]}: {exc}")
+            errors.append(f"{label}: invalid regex: {exc}")
-    return compiled, invalid
-
-
-def apply_rules(body: str, compiled: list[tuple[regex.Pattern, str]]) -> str:
-    """Rules in the order given, each over the whole text. A pattern that times out is
-    skipped for that record rather than aborting the measurement."""
-    text = body
-    for pattern, replacement in compiled:
-        try:
-            text = pattern.sub(replacement, text, timeout=PATTERN_TIMEOUT_S)
-        except (TimeoutError, regex.error):
             continue
-    return text
+        compiled.append(CompiledRule(index=index, pattern=pattern, replacement=rule.replacement))
+    return compiled, errors
 
 
-def _innocent_text(record: Record) -> str:
+def apply_rules(
-    """The record with every secret and decoy cut out — the part no rule should touch."""
+    body: str,
-    text = record.body
+    compiled: Iterable[CompiledRule],
-    for span in [*record.secrets, *record.decoys]:
+    budget: RegexBudget,
-        text = text.replace(span.text, " ")
+) -> ApplyResult:
-    return text
+    """Apply rules atomically per pattern and carry source offsets through unchanged text."""
+    tracked = TrackedText.original(body)
+    errors: list[str] = []
+    for rule in compiled:
+        label = f"rule[{rule.index}]"
+        try:
+            timeout = budget.operation_timeout()
+            matches = list(rule.pattern.finditer(tracked.text, timeout=timeout))
+            budget.operation_timeout()
+        except (TimeoutError, regex.error):
+            errors.append(f"{label}: regex timed out")
+            continue
+        except RegexBudgetExceeded as exc:
+            errors.append(f"{label}: {exc}")
+            break
+        if len(matches) > MAX_MATCHES_PER_RULE_RECORD:
+            errors.append(
+                f"{label}: produced {len(matches)} matches in one record; maximum is "
+                f"{MAX_MATCHES_PER_RULE_RECORD}"
+            )
+            continue
+
+        text_parts: list[str] = []
+        origins: list[int | None] = []
+        cursor = 0
+        for match in matches:
+            start, end = match.span()
+            text_parts.append(tracked.text[cursor:start])
+            origins.extend(tracked.origins[cursor:start])
+            text_parts.append(rule.replacement)
+            origins.extend([None] * len(rule.replacement))
+            cursor = end
+        text_parts.append(tracked.text[cursor:])
+        origins.extend(tracked.origins[cursor:])
+        tracked = TrackedText(text="".join(text_parts), origins=tuple(origins))
+    return ApplyResult(tracked=tracked, errors=tuple(errors))
 
 
-def measure(records: list[Record], rules: list[Rule]) -> Outcome:
+def _locate(body: str, span: Span) -> LocatedSpan:
-    compiled, invalid = compile_rules(rules)
+    start = body.find(span.text)
-    residual = collateral = secrets = decoys = lost = total = 0
+    if start < 0 or body.find(span.text, start + 1) >= 0:
+        raise ValueError(f"ground-truth span {span.kind!r} must occur exactly once")
+    return LocatedSpan(span.kind, span.text, start, start + len(span.text))
+
+
+def _removed_count(span: LocatedSpan, retained: set[int]) -> int:
+    return sum(1 for offset in span.offsets if offset not in retained)
+
+
+def measure(
+    records: list[Record],
+    parsed: ParsedRules | list[Rule],
+    *,
+    budget_seconds: float = EPISODE_REGEX_BUDGET_S,
+    clock: Callable[[], float] = time.monotonic,
+) -> Outcome:
+    """Measure one ruleset under one episode-wide hostile-regex budget."""
+    if isinstance(parsed, list):
+        parsed = ParsedRules(tuple(parsed), submitted=len(parsed))
+    budget = RegexBudget(seconds=budget_seconds, clock=clock)
+    compiled, errors = compile_rules(parsed, budget)
+    removed_secrets = partial_secrets = residual = 0
+    secrets_total = credited_secret_chars = removed_secret_chars = secret_chars_total = 0
+    collateral = decoys_total = innocent_lost = innocent_total = 0
+
     for record in records:
-        redacted = apply_rules(record.body, compiled)
+        result = apply_rules(record.body, compiled, budget)
-        secrets += len(record.secrets)
+        errors.extend(result.errors)
-        decoys += len(record.decoys)
+        retained = {offset for offset in result.tracked.origins if offset is not None}
-        residual += sum(1 for s in record.secrets if s.text in redacted)
+        secrets = [_locate(record.body, span) for span in record.secrets]
-        collateral += sum(1 for d in record.decoys if d.text not in redacted)
+        decoys = [_locate(record.body, span) for span in record.decoys]
-        # Fidelity is counted over whole words rather than characters: a rule that eats
+        secret_offsets = {offset for span in secrets for offset in span.offsets}
-        # every space would otherwise keep most of its characters and look harmless.
+        decoy_offsets = {offset for span in decoys for offset in span.offsets}
-        innocent = [w for w in _innocent_text(record).split() if w]
+
-        total += len(innocent)
+        for span in secrets:
-        lost += sum(1 for w in innocent if w not in redacted)
+            removed = _removed_count(span, retained)
-    return Outcome(residual, secrets, collateral, decoys, lost, total, invalid)
+            removed_secret_chars += removed
+            secret_chars_total += span.end - span.start
+            secrets_total += 1
+            if removed == span.end - span.start:
+                removed_secrets += 1
+                credited_secret_chars += removed
+            elif removed > 0:
+                partial_secrets += 1
+                residual += 1
+            else:
+                residual += 1
+
+        for span in decoys:
+            decoys_total += 1
+            if _removed_count(span, retained) > 0:
+                collateral += 1
+
+        # Every non-secret source character is protected. Decoys are included here as well
+        # as in their span-level gate, so one altered byte is both visible and costly.
+        protected = set(range(len(record.body))) - secret_offsets
+        innocent_total += len(protected)
+        innocent_lost += len(protected - retained)
+
+        # Keep this assertion close to the accounting: every decoy must be protected.
+        if not decoy_offsets <= protected:
+            raise AssertionError("decoy offsets must be non-secret protected content")
+
+    unique_errors = list(dict.fromkeys(errors))
+    return Outcome(
+        removed_secrets=removed_secrets,
+        partial_secrets=partial_secrets,
+        residual_hits=residual,
+        secrets_total=secrets_total,
+        credited_secret_chars=credited_secret_chars,
+        removed_secret_chars=removed_secret_chars,
+        secret_chars_total=secret_chars_total,
+        collateral_hits=collateral,
+        decoys_total=decoys_total,
+        innocent_lost=innocent_lost,
+        innocent_total=innocent_total,
+        invalid_rules=unique_errors,
+        budget_seconds=budget.seconds,
+        budget_elapsed_seconds=budget.elapsed,
+    )

environments/redaction_pressure/redaction_pressure/taskset.py

@@ -1,79 +1,123 @@
-"""redaction-pressure: write redaction rules that survive records you never read.
+"""Verifiers v1 taskset for provenance-safe redaction rules."""
-
-The agent reads a handful of support tickets seeded with secrets, and returns a ruleset.
-The ruleset is then run against the *next* records the same generator would have produced.
-That gap is the environment: rules fitted to the visible examples score well on the
-visible examples and are worth nothing, and only the held-out slice can tell the two apart.
-
-Three rewards, all counted over what the ruleset *destroyed*:
-
-  recall     seeded secrets removed from the held-out text.
-  precision  of everything destroyed, the share that was a secret. An empty ruleset
-             destroys nothing and therefore scores zero here — which is the point. A
-             first pass scored survivors instead, and an empty ruleset tied the best
-             real attempt at 0.500 because it had kept every decoy and every word.
-  gate       binary: every secret gone and no decoy taken with it. `forge redact`'s
-             exit 1, which is a verdict and not a deduction.
-
-No judge model and no runtime: the reward is a scanner, so a rollout costs one completion
-and grades in milliseconds. Deterministic from the task seed.
-"""
 
 from __future__ import annotations
 
+import hashlib
+import hmac
 import json
+import os
 import re
-from typing import ClassVar
+from typing import Literal
 
-from pydantic import Field
+from pydantic import Field, model_validator
 
 import verifiers.v1 as vf
 
 from redaction_pressure.corpus import build_slices
-from redaction_pressure.scan import MAX_RULES, Rule, measure
+from redaction_pressure.provenance import (
+    CORPUS_VERSION,
+    ENVIRONMENT_VERSION,
+    REWARD_SCHEMA_VERSION,
+    environment_source_sha256,
+    trace_provenance,
+)
+from redaction_pressure.scan import MAX_RULES, ParsedRules, Rule, measure
+
+Split = Literal["train", "dev", "private_eval"]
+PUBLIC_SPLIT_BASES = {"train": 100_000, "dev": 200_000}
+PRIVATE_SEED_ENV_DEFAULT = "REDACTION_PRESSURE_PRIVATE_SEED"
+MAX_TASKS_PER_SPLIT = 10_000
 
 SYSTEM = f"""You are writing redaction rules for a support-ticket corpus.
 
 Return a JSON array of at most {MAX_RULES} rules in one ```json code block. Each rule is
-an object with "pattern" (a Python regular expression) and "replacement" (the text that
+an object with "pattern" (a Python regex) and "replacement". For safety, replacement
-replaces every match).
+must be exactly "[REDACTED]" or the empty string; capture expansion and arbitrary
+replacement text invalidate the complete ruleset.
 
-Your rules will NOT be graded on the tickets below. They will be run against different
+Your rules are graded on different tickets from the same source. A secret earns credit
-tickets from the same source, containing the same kinds of secret and the same kinds of
+only when every source character in its labelled span is removed. Changing one byte is
-lookalike. Removing every secret matters most; destroying text that only resembles a
+still a residual secret. Every non-secret source character, including public lookalikes,
-secret, or shredding ordinary prose, costs you.
+must survive for the clean gate. Invalid, excessive, over-broad, or timed-out rules make
+the rollout invalid and receive zero reward.
 
-Example shape:
+Example:
 ```json
-[{{"pattern": "\\\\bfoo-\\\\d+\\\\b", "replacement": "[REF]"}}]
+[{{"pattern": "\\\\bfoo-\\\\d+\\\\b", "replacement": "[REDACTED]"}}]
 ```"""
 
-_BLOCK = re.compile(r"```(?:json)?\s*\n(.*?)```", re.DOTALL)
+_BLOCK = re.compile(r"```(?:json)?\s*\n(.*?)```", re.DOTALL | re.IGNORECASE)
 
 
-def parse_rules(reply: str) -> list[Rule]:
+def parse_rules(reply: str) -> ParsedRules:
-    """The last JSON array in the reply. A reply with no parsable ruleset is an empty
+    """Parse the last fenced JSON block and retain every structural error."""
-    ruleset, not an error — it scores what doing nothing scores."""
     candidates = _BLOCK.findall(reply or "")
     raw = candidates[-1] if candidates else (reply or "")
     try:
-        parsed = json.loads(raw.strip())
+        payload = json.loads(raw.strip())
-    except json.JSONDecodeError:
+    except json.JSONDecodeError as exc:
-        return []
+        return ParsedRules((), (f"response is not valid JSON: {exc.msg}",), 0)
-    if not isinstance(parsed, list):
+    if not isinstance(payload, list):
-        return []
+        return ParsedRules((), ("response JSON must be an array",), 1)
-    rules = []
+
-    for item in parsed:
+    rules: list[Rule] = []
-        if isinstance(item, dict) and isinstance(item.get("pattern"), str):
+    errors: list[str] = []
-            rules.append(Rule(item["pattern"], str(item.get("replacement", ""))))
+    for index, item in enumerate(payload):
-    return rules
+        if not isinstance(item, dict):
+            errors.append(f"item[{index}] must be an object")
+            continue
+        pattern = item.get("pattern")
+        replacement = item.get("replacement", "")
+        if not isinstance(pattern, str):
+            errors.append(f"item[{index}].pattern must be a string")
+            continue
+        if not isinstance(replacement, str):
+            errors.append(f"item[{index}].replacement must be a string")
+            continue
+        rules.append(Rule(pattern=pattern, replacement=replacement))
+    return ParsedRules(tuple(rules), tuple(errors), len(payload))
+
+
+def _private_key(env_name: str) -> bytes:
+    value = os.environ.get(env_name, "")
+    if len(value) < 32:
+        raise ValueError(
+            f"private_eval requires at least 32 characters in environment variable {env_name}"
+        )
+    return value.encode("utf-8")
+
+
+def seed_for(split: Split, index: int, private_seed_env: str = PRIVATE_SEED_ENV_DEFAULT) -> int:
+    if split in PUBLIC_SPLIT_BASES:
+        return PUBLIC_SPLIT_BASES[split] + index
+    key = _private_key(private_seed_env)
+    digest = hmac.new(
+        key,
+        f"redaction-pressure/v2/private_eval/{index}".encode("utf-8"),
+        hashlib.sha256,
+    ).digest()
+    return int.from_bytes(digest[:8], "big") & ((1 << 63) - 1)
+
+
+def seed_commitment(split: Split, private_seed_env: str = PRIVATE_SEED_ENV_DEFAULT) -> str:
+    if split in PUBLIC_SPLIT_BASES:
+        material = f"redaction-pressure/v2/{split}/{PUBLIC_SPLIT_BASES[split]}".encode()
+    else:
+        material = _private_key(private_seed_env)
+    return hashlib.sha256(material).hexdigest()
 
 
 class RedactionData(vf.TaskData):
     seed: int
-    """Rebuilds both slices exactly; the held-out records are never serialized here."""
+    """Rebuilds both slices exactly; held-out records are generated only for scoring."""
     visible: int
     held_out: int
+    split: Split
+    seed_commitment: str
+    environment_version: str
+    corpus_version: str
+    reward_schema_version: str
+    source_sha256: str
 
 
 class RedactionTask(vf.Task[RedactionData]):
@@ -83,59 +127,81 @@ class RedactionTask(vf.Task[RedactionData]):
 
     @vf.metric
     async def scan(self, trace: vf.Trace) -> dict[str, float]:
-        """Run the ruleset over the held-out slice once; every reward reads this."""
+        """Rebuild the held-out slice once and expose raw diagnostics in the trace."""
         _, held_out = build_slices(self.data.seed, self.data.visible, self.data.held_out)
         outcome = measure(held_out, parse_rules(trace.last_reply))
+        trace.info["environment"] = trace_provenance(
+            split=self.data.split,
+            seed_commitment=self.data.seed_commitment,
+        )
+        trace.info["redaction_errors"] = outcome.invalid_rules
         return {
             "recall": outcome.recall,
             "precision": outcome.precision,
             "clean": float(outcome.clean),
+            "validity": outcome.validity,
+            "character_coverage": outcome.character_coverage,
+            "ordinary_preservation": outcome.ordinary_preservation,
             "residual_hits": float(outcome.residual_hits),
+            "partial_secrets": float(outcome.partial_secrets),
             "collateral_hits": float(outcome.collateral_hits),
             "innocent_lost": float(outcome.innocent_lost),
             "invalid_rules": float(len(outcome.invalid_rules)),
+            "regex_budget_elapsed_seconds": outcome.budget_elapsed_seconds,
         }
 
     @vf.reward(weight=0.35)
     async def recall(self, trace: vf.Trace) -> float:
-        return trace.metrics.get("recall", 0.0)
+        return trace.metrics.get("recall", 0.0) * trace.metrics.get("validity", 0.0)
 
     @vf.reward(weight=0.35)
     async def precision(self, trace: vf.Trace) -> float:
-        return trace.metrics.get("precision", 0.0)
+        return trace.metrics.get("precision", 0.0) * trace.metrics.get("validity", 0.0)
 
     @vf.reward(weight=0.30)
     async def gate(self, trace: vf.Trace) -> float:
-        return trace.metrics.get("clean", 0.0)
+        return trace.metrics.get("clean", 0.0) * trace.metrics.get("validity", 0.0)
 
 
 class RedactionConfig(vf.TasksetConfig):
-    num_tasks: int = Field(64, ge=1)
+    num_tasks: int = Field(64, ge=1, le=MAX_TASKS_PER_SPLIT)
-    visible: int = Field(4, ge=1)
+    visible: int = Field(4, ge=1, le=64)
-    """Records the agent reads."""
+    held_out: int = Field(12, ge=1, le=256)
-    held_out: int = Field(12, ge=1)
+    split: Split = "dev"
-    """Records it is graded on and never sees."""
+    private_seed_env: str = PRIVATE_SEED_ENV_DEFAULT
+
+    @model_validator(mode="after")
+    def validate_private_split(self) -> "RedactionConfig":
+        if self.split == "private_eval":
+            _private_key(self.private_seed_env)
+        return self
 
 
 class RedactionTaskset(vf.Taskset[RedactionTask, RedactionConfig]):
-    SEED_BASE: ClassVar[int] = 90_000
-
     def load(self) -> list[RedactionTask]:
+        commitment = seed_commitment(self.config.split, self.config.private_seed_env)
+        source_hash = environment_source_sha256()
         tasks = []
-        for i in range(self.config.num_tasks):
+        for index in range(self.config.num_tasks):
-            seed = self.SEED_BASE + i
+            seed = seed_for(self.config.split, index, self.config.private_seed_env)
             visible, _ = build_slices(seed, self.config.visible, self.config.held_out)
             tickets = "\n\n".join(record.body for record in visible)
             tasks.append(
                 RedactionTask(
                     RedactionData(
-                        idx=i,
+                        idx=index,
-                        name=f"stream-{seed}",
+                        name=f"{self.config.split}-{index:05d}",
                         prompt=f"Tickets from the stream:\n\n{tickets}\n\nWrite the ruleset.",
                         system_prompt=SYSTEM,
                         seed=seed,
                         visible=self.config.visible,
                         held_out=self.config.held_out,
+                        split=self.config.split,
+                        seed_commitment=commitment,
+                        environment_version=ENVIRONMENT_VERSION,
+                        corpus_version=CORPUS_VERSION,
+                        reward_schema_version=REWARD_SCHEMA_VERSION,
+                        source_sha256=source_hash,
                     ),
                     self.config.task,
                 )

environments/redaction_pressure/tests/__init__.py

@@ -0,0 +1 @@
+"""Tests for redaction-pressure."""

environments/redaction_pressure/tests/helpers.py

@@ -0,0 +1,36 @@
+from __future__ import annotations
+
+import json
+
+from redaction_pressure.scan import Rule
+
+
+ORACLE_RULES = [
+    Rule(r"(?<=Customer pasted their key )sk-live-[a-z0-9]{24}", "[REDACTED]"),
+    Rule(r"(?<=Reply-to was )[a-z]+[.][a-z]+@[a-z]{6}[.]com", "[REDACTED]"),
+    Rule(r"(?<=Callback number )[+]1-[0-9]{3}-[0-9]{3}-[0-9]{4}", "[REDACTED]"),
+    Rule(r"(?<=They read the card )[0-9]{4}(?:-[0-9]{4}){3}", "[REDACTED]"),
+    Rule(r"(?<=Trace pointed at )[a-z]{5}-db-[1-9][.]internal[.]lumbridge", "[REDACTED]"),
+    Rule(
+        r"(?<=Escalated by )(?:Dana|Priya|Marcus|Ines|Tomas|Rui|Nadia|Owen) "
+        r"(?:Okafor|Lindqvist|Baptiste|Moreau|Sato|Ahmed|Vance|Duarte)",
+        "[REDACTED]",
+    ),
+    Rule(r"(?<=Session )[0-9a-f]{8}(?:-[0-9a-f]{4}){3}-[0-9a-f]{12}", "[REDACTED]"),
+]
+
+FIRST_CHARACTER_HACK = [
+    Rule(r"(?<=Customer pasted their key )s", ""),
+    Rule(r"(?<=Reply-to was )[a-z]", ""),
+    Rule(r"(?<=Callback number )[+]", ""),
+    Rule(r"(?<=They read the card )[0-9]", ""),
+    Rule(r"(?<=Trace pointed at )[a-z]", ""),
+    Rule(r"(?<=Escalated by )[A-Z]", ""),
+    Rule(r"(?<=Session )[0-9a-f]", ""),
+]
+
+
+def reply_for(rules: list[Rule]) -> str:
+    return "```json\n" + json.dumps(
+        [{"pattern": rule.pattern, "replacement": rule.replacement} for rule in rules]
+    ) + "\n```"

environments/redaction_pressure/tests/test_scan.py

@@ -0,0 +1,116 @@
+from __future__ import annotations
+
+import unittest
+
+import regex
+
+from redaction_pressure.corpus import build_slices
+from redaction_pressure.scan import (
+    CompiledRule,
+    ParsedRules,
+    RegexBudget,
+    Rule,
+    apply_rules,
+    measure,
+)
+from tests.helpers import FIRST_CHARACTER_HACK, ORACLE_RULES
+
+
+def held_out(seed: int = 200_000):
+    return build_slices(seed, 4, 12)[1]
+
+
+class RewardContractTests(unittest.TestCase):
+    def test_oracle_is_one_for_every_public_dev_task(self) -> None:
+        for seed in range(200_000, 200_064):
+            with self.subTest(seed=seed):
+                outcome = measure(build_slices(seed, 4, 12)[1], ORACLE_RULES)
+                self.assertEqual(outcome.recall, 1.0)
+                self.assertEqual(outcome.precision, 1.0)
+                self.assertTrue(outcome.clean)
+                self.assertEqual(outcome.validity, 1.0)
+
+    def test_inaction_is_zero(self) -> None:
+        outcome = measure(held_out(), [])
+        self.assertEqual(outcome.recall, 0.0)
+        self.assertEqual(outcome.precision, 0.0)
+        self.assertFalse(outcome.clean)
+        self.assertEqual(outcome.validity, 1.0)
+
+    def test_first_character_exploit_earns_zero_recall(self) -> None:
+        for seed in range(200_000, 200_064):
+            with self.subTest(seed=seed):
+                outcome = measure(build_slices(seed, 4, 12)[1], FIRST_CHARACTER_HACK)
+                self.assertEqual(outcome.recall, 0.0)
+                self.assertEqual(outcome.precision, 0.0)
+                self.assertEqual(outcome.partial_secrets, outcome.secrets_total)
+                self.assertEqual(outcome.residual_hits, outcome.secrets_total)
+                self.assertFalse(outcome.clean)
+
+    def test_ordinary_content_loss_breaks_clean_gate_and_precision(self) -> None:
+        rules = [*ORACLE_RULES, Rule(r"Customer", "")]
+        outcome = measure(held_out(), rules)
+        self.assertEqual(outcome.recall, 1.0)
+        self.assertGreater(outcome.innocent_lost, 0)
+        self.assertLess(outcome.precision, 1.0)
+        self.assertFalse(outcome.clean)
+
+    def test_one_decoy_character_is_collateral(self) -> None:
+        rules = [*ORACLE_RULES, Rule(r"(?<=our published sample is )s", "")]
+        outcome = measure(held_out(), rules)
+        self.assertGreater(outcome.collateral_hits, 0)
+        self.assertGreater(outcome.innocent_lost, 0)
+        self.assertFalse(outcome.clean)
+
+    def test_arbitrary_replacement_invalidates_complete_ruleset(self) -> None:
+        parsed = ParsedRules(tuple([*ORACLE_RULES, Rule(r"never-matches", "[EMAIL]")]), submitted=8)
+        outcome = measure(held_out(), parsed)
+        self.assertEqual(outcome.validity, 0.0)
+        self.assertFalse(outcome.clean)
+        self.assertTrue(any("replacement" in error for error in outcome.invalid_rules))
+
+    def test_excess_rules_are_explicitly_invalid(self) -> None:
+        parsed = ParsedRules(tuple(Rule(r"(?!)", "") for _ in range(25)), submitted=25)
+        outcome = measure(held_out(), parsed)
+        self.assertEqual(outcome.validity, 0.0)
+        self.assertTrue(any("too many rules" in error for error in outcome.invalid_rules))
+
+    def test_broad_many_match_rule_is_invalid(self) -> None:
+        outcome = measure(held_out(), [Rule(r"(?=.)", "")])
+        self.assertEqual(outcome.validity, 0.0)
+        self.assertTrue(any("matches in one record" in error for error in outcome.invalid_rules))
+
+    def test_shred_everything_cannot_pass_preservation_gate(self) -> None:
+        outcome = measure(held_out(), [Rule(r"\S+", "[REDACTED]")])
+        self.assertGreater(outcome.innocent_lost, 0)
+        self.assertGreater(outcome.collateral_hits, 0)
+        self.assertFalse(outcome.clean)
+        self.assertLess(outcome.precision, 1.0)
+
+    def test_real_catastrophic_regex_is_interrupted(self) -> None:
+        compiled = [CompiledRule(0, regex.compile(r"(?:a|aa)+$"), "")]
+        result = apply_rules("a" * 20_000 + "!", compiled, RegexBudget(seconds=0.05))
+        self.assertTrue(any("timed out" in error or "budget" in error for error in result.errors))
+        self.assertEqual(result.tracked.text, "a" * 20_000 + "!")
+
+    def test_episode_budget_is_shared_and_bounded(self) -> None:
+        class Clock:
+            def __init__(self) -> None:
+                self.value = 0.0
+
+            def __call__(self) -> float:
+                self.value += 0.2
+                return self.value
+
+        outcome = measure(held_out(), ORACLE_RULES, budget_seconds=0.5, clock=Clock())
+        self.equal_zero_reward_invalid(outcome)
+        self.assertGreaterEqual(outcome.budget_elapsed_seconds, outcome.budget_seconds)
+        self.assertTrue(any("budget exhausted" in error for error in outcome.invalid_rules))
+
+    def equal_zero_reward_invalid(self, outcome) -> None:
+        self.assertEqual(outcome.validity, 0.0)
+        self.assertFalse(outcome.clean)
+
+
+if __name__ == "__main__":
+    unittest.main()

environments/redaction_pressure/tests/test_taskset.py

@@ -0,0 +1,120 @@
+from __future__ import annotations
+
+import asyncio
+import os
+import string
+import unittest
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from pydantic import ValidationError
+
+from redaction_pressure.corpus import build_slices
+from redaction_pressure.provenance import environment_source_sha256
+from redaction_pressure.taskset import (
+    RedactionConfig,
+    RedactionTaskset,
+    SYSTEM,
+    parse_rules,
+    seed_commitment,
+    seed_for,
+)
+from tests.helpers import ORACLE_RULES, reply_for
+
+
+class ParserTests(unittest.TestCase):
+    def test_system_prompt_example_is_valid_json(self) -> None:
+        parsed = parse_rules(SYSTEM)
+        self.assertEqual(parsed.errors, ())
+        self.assertEqual(len(parsed.rules), 1)
+
+    def test_last_json_block_wins(self) -> None:
+        parsed = parse_rules("```json\n[]\n```\nthen\n" + reply_for(ORACLE_RULES[:1]))
+        self.assertEqual(parsed.rules, tuple(ORACLE_RULES[:1]))
+        self.assertEqual(parsed.errors, ())
+        self.assertEqual(parsed.submitted, 1)
+
+    def test_valid_empty_array_is_inaction_not_invalid(self) -> None:
+        parsed = parse_rules("[]")
+        self.assertEqual(parsed.rules, ())
+        self.assertEqual(parsed.errors, ())
+        self.assertEqual(parsed.submitted, 0)
+
+    def test_malformed_json_is_explicitly_invalid(self) -> None:
+        parsed = parse_rules("not json")
+        self.assertFalse(parsed.rules)
+        self.assertTrue(parsed.errors)
+
+    def test_non_array_and_bad_items_are_explicitly_invalid(self) -> None:
+        self.assertTrue(parse_rules('{"pattern": "x"}').errors)
+        parsed = parse_rules('[1, {"pattern": 2}, {"pattern": "x", "replacement": 4}]')
+        self.assertEqual(len(parsed.errors), 3)
+        self.assertEqual(parsed.submitted, 3)
+
+
+class SeedAndProvenanceTests(unittest.TestCase):
+    def test_corpus_is_deterministic(self) -> None:
+        self.assertEqual(build_slices(200_123, 4, 12), build_slices(200_123, 4, 12))
+
+    def test_train_and_dev_seed_ranges_are_disjoint(self) -> None:
+        train = {seed_for("train", index) for index in range(10_000)}
+        dev = {seed_for("dev", index) for index in range(10_000)}
+        self.assertTrue(train.isdisjoint(dev))
+
+    def test_private_split_requires_external_secret(self) -> None:
+        with patch.dict(os.environ, {}, clear=True):
+            with self.assertRaises(ValidationError):
+                RedactionConfig(split="private_eval")
+
+    def test_private_seeds_are_deterministic_keyed_and_committed(self) -> None:
+        key_a = "a" * 32
+        key_b = "b" * 32
+        with patch.dict(os.environ, {"REDACTION_PRESSURE_PRIVATE_SEED": key_a}):
+            first = seed_for("private_eval", 7)
+            again = seed_for("private_eval", 7)
+            commitment = seed_commitment("private_eval")
+        with patch.dict(os.environ, {"REDACTION_PRESSURE_PRIVATE_SEED": key_b}):
+            other = seed_for("private_eval", 7)
+        self.assertEqual(first, again)
+        self.assertNotEqual(first, other)
+        self.assertNotIn(key_a, commitment)
+        self.assertEqual(len(commitment), 64)
+
+    def test_source_hash_is_stable_hex(self) -> None:
+        first = environment_source_sha256()
+        self.assertEqual(first, environment_source_sha256())
+        self.assertEqual(len(first), 64)
+        self.assertTrue(set(first) <= set(string.hexdigits.lower()))
+
+    def test_task_data_carries_replay_identity_without_secret_key(self) -> None:
+        taskset = RedactionTaskset(RedactionConfig(num_tasks=2, split="dev"))
+        tasks = taskset.load()
+        self.assertEqual([task.data.name for task in tasks], ["dev-00000", "dev-00001"])
+        self.assertNotEqual(tasks[0].data.seed, tasks[1].data.seed)
+        self.assertEqual(tasks[0].data.source_sha256, environment_source_sha256())
+        self.assertEqual(len(tasks[0].data.seed_commitment), 64)
+        self.assertEqual(tasks[0].data.environment_version, "0.2.0")
+
+    def test_scoring_records_provenance_and_oracle_metrics(self) -> None:
+        task = RedactionTaskset(RedactionConfig(num_tasks=1, split="dev")).load()[0]
+        trace = SimpleNamespace(last_reply=reply_for(ORACLE_RULES), info={})
+        metrics = asyncio.run(task.scan(trace))
+        self.assertEqual(metrics["recall"], 1.0)
+        self.assertEqual(metrics["precision"], 1.0)
+        self.assertEqual(metrics["clean"], 1.0)
+        self.assertEqual(metrics["validity"], 1.0)
+        self.assertEqual(trace.info["environment"]["source_sha256"], environment_source_sha256())
+        self.assertEqual(trace.info["redaction_errors"], [])
+
+    def test_invalidity_multiplies_all_rewards_to_zero(self) -> None:
+        task = RedactionTaskset(RedactionConfig(num_tasks=1, split="dev")).load()[0]
+        trace = SimpleNamespace(
+            metrics={"recall": 1.0, "precision": 1.0, "clean": 1.0, "validity": 0.0}
+        )
+        self.assertEqual(asyncio.run(task.recall(trace)), 0.0)
+        self.assertEqual(asyncio.run(task.precision(trace)), 0.0)
+        self.assertEqual(asyncio.run(task.gate(trace)), 0.0)
+
+
+if __name__ == "__main__":
+    unittest.main()

environments/redaction_pressure/uv.lock

@@ -5,8 +5,8 @@ resolution-markers = [
     "python_full_version >= '3.14' and sys_platform == 'win32'",
     "python_full_version >= '3.14' and sys_platform != 'win32'",
     "python_full_version >= '3.12' and python_full_version < '3.14' and sys_platform == 'win32'",
-    "python_full_version < '3.12' and sys_platform == 'win32'",
     "python_full_version >= '3.12' and python_full_version < '3.14' and sys_platform != 'win32'",
+    "python_full_version < '3.12' and sys_platform == 'win32'",
     "python_full_version < '3.12' and sys_platform != 'win32'",
 ]
 
@@ -2308,17 +2308,19 @@ wheels = [
 
 [[package]]
 name = "redaction-pressure"
-version = "0.1.0"
+version = "0.2.0"
 source = { editable = "." }
 dependencies = [
+    { name = "pydantic" },
     { name = "regex" },
     { name = "verifiers" },
 ]
 
 [package.metadata]
 requires-dist = [
-    { name = "regex" },
+    { name = "pydantic", specifier = "==2.13.4" },
-    { name = "verifiers" },
+    { name = "regex", specifier = "==2026.7.19" },
+    { name = "verifiers", specifier = "==0.3.0" },
 ]
 
 [[package]]