Spaces:
Sleeping
Sleeping
File size: 19,749 Bytes
c8ebaee e32a33b c8ebaee e32a33b c8ebaee 3149b7e c8ebaee b7aa1f0 c8ebaee b7aa1f0 c8ebaee 37bfd28 c8ebaee 2dedffd c8ebaee 3149b7e 2dedffd 3149b7e c8ebaee 37bfd28 c8ebaee 37bfd28 c8ebaee b7aa1f0 e32a33b b7aa1f0 c8ebaee 37bfd28 c8ebaee 37bfd28 c8ebaee b7aa1f0 e32a33b c8ebaee b7aa1f0 c8ebaee 3149b7e c8ebaee e32a33b c8ebaee e32a33b c8ebaee e32a33b b7aa1f0 c8ebaee 3149b7e c8ebaee 3149b7e c8ebaee 37bfd28 c8ebaee | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 | """OpenEnv Rubric subclasses that power ChargebackOps grading.
Every scoring dimension is a standalone :class:`openenv.core.rubrics.Rubric`
so the whole grader can be introspected via ``named_rubrics``, captured via
``state_dict``, and swapped piecewise (e.g. replace :class:`NoteQualityRubric`
with an ``LLMJudge``). The per-case composite uses :class:`WeightedSum` with
weights that must sum to 1.0.
The rubrics take their inputs via a :class:`GradingContext` dataclass passed
as the ``action`` argument of :meth:`Rubric.forward`. The ``observation``
argument is ignored β ChargebackOps grading operates over deterministic
episode progress, not on the last observation payload. This keeps the rubrics
pure and unit-testable without an environment instance.
Set ``USE_LLM_NOTE_JUDGE=1`` to swap the deterministic
:class:`NoteQualityRubric` for the LLM-backed
:class:`evaluation.llm_note_judge.LLMNoteJudgeRubric` when constructing
:class:`CaseRubric`. The LLM rubric falls back to the deterministic scorer
on any failure, so offline benchmarks remain reproducible without API keys.
"""
from __future__ import annotations
import os
from dataclasses import dataclass
from typing import Any
from openenv.core.rubrics import Gate, Rubric, WeightedSum
try:
from ..scenarios.arbitration import (
ARB_FEE_PER_SIDE,
ARB_ISSUER_WIN_THRESHOLD,
ARB_MERCHANT_WIN_THRESHOLD,
)
from ..scenarios.issuer_model import evidence_strength_score
from ..scenarios.simulation import CaseProgress, InternalCase, TaskScenario
except ImportError: # pragma: no cover
from scenarios.arbitration import (
ARB_FEE_PER_SIDE,
ARB_ISSUER_WIN_THRESHOLD,
ARB_MERCHANT_WIN_THRESHOLD,
)
from scenarios.issuer_model import evidence_strength_score
from scenarios.simulation import CaseProgress, InternalCase, TaskScenario
@dataclass(frozen=True)
class GradingContext:
"""Inputs one per-case rubric evaluation needs."""
case: InternalCase
progress: CaseProgress
step_count: int
@dataclass(frozen=True)
class EpisodeGradingContext:
"""Inputs the episode-level rubric needs."""
task: TaskScenario
progress_by_case: dict[str, CaseProgress]
step_count: int
def _ratio(numerator: int, denominator: int) -> float:
if denominator <= 0:
return 1.0
return max(0.0, min(1.0, numerator / denominator))
def _final_resolution(progress: CaseProgress) -> str:
return progress.final_resolution or "unresolved"
def _contest_is_valid(case: InternalCase) -> bool:
return case.optimal_strategy == "contest" or "contest" in case.acceptable_strategies
class StrategyCorrectnessRubric(Rubric):
"""Score final strategy: optimal=1.0, acceptable=0.35, else 0.0."""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
final = _final_resolution(ctx.progress)
if final == ctx.case.optimal_strategy:
return 1.0
if final in ctx.case.acceptable_strategies:
return 0.35
return 0.0
class EvidenceQualityRubric(Rubric):
"""Score the evidence packet attached to the case.
Zeroes out (vacuous-truth fix) when the agent contests a case that was
never contestable β no evidence quality can rescue a wrong strategy.
"""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
case = ctx.case
progress = ctx.progress
final = _final_resolution(progress)
attached_set = set(progress.attached_evidence_ids)
required_attached = len(attached_set.intersection(case.required_evidence_ids))
helpful_attached = len(attached_set.intersection(case.helpful_evidence_ids))
harmful_attached = len(attached_set.intersection(case.harmful_evidence_ids))
if final == "contest":
if not _contest_is_valid(case):
return 0.0
base = 0.7 * _ratio(required_attached, len(case.required_evidence_ids))
bonus = 0.3 * _ratio(
helpful_attached, max(1, len(case.helpful_evidence_ids))
)
penalty = 0.25 * harmful_attached
return max(0.0, min(1.0, base + bonus - penalty))
if final in {"accept_chargeback", "issue_refund"}:
if case.optimal_strategy == "contest":
return 0.15
return 1.0 if helpful_attached == 0 and harmful_attached == 0 else 0.7
return 0.0
class PacketValidityRubric(Rubric):
"""All-or-nothing: required evidence complete AND no harmful attached."""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
case = ctx.case
progress = ctx.progress
final = _final_resolution(progress)
attached_set = set(progress.attached_evidence_ids)
required_attached = len(attached_set.intersection(case.required_evidence_ids))
harmful_attached = len(attached_set.intersection(case.harmful_evidence_ids))
if final == "contest":
if not _contest_is_valid(case):
return 0.0
if (
required_attached == len(case.required_evidence_ids)
and harmful_attached == 0
):
return 1.0
return 0.0
if final in {"accept_chargeback", "issue_refund"}:
if case.optimal_strategy == "contest":
return 0.0
return 1.0
return 0.0
class DeadlineComplianceRubric(Rubric):
"""1.0 if resolved on time, else 0.0."""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
case = ctx.case
progress = ctx.progress
final = _final_resolution(progress)
if final == "unresolved":
return 0.0
if final == "contest" and progress.merchant_submitted_at_step is not None:
# Long-horizon tasks can delay issuer decisions by several steps.
# Deadline compliance is based on the merchant's submission time,
# not when the issuer eventually responds.
resolution_step = progress.merchant_submitted_at_step
else:
resolution_step = (
progress.resolved_at_step
if progress.resolved_at_step is not None
else ctx.step_count
)
if resolution_step > case.deadline_step:
return 0.0
return 1.0
class CaseAbandonedRubric(Rubric):
"""Hard-constraint rubric: 0.0 if the case was abandoned past deadline.
Used as the inner rubric for the :class:`Gate` wrapped around
:class:`CaseRubric`. The distinction from :class:`DeadlineComplianceRubric`
is intentional:
* :class:`DeadlineComplianceRubric` is the *dimension* β penalises any
late resolution with a 0 in the weighted sum (15% score drop).
* :class:`CaseAbandonedRubric` is the *gate* β hard-zeros the entire
case only when the agent never even attempted to resolve it before
the deadline expired. In real chargeback operations this is the
"no contest, case closed" outcome: the merchant forfeited.
This split means a late-but-completed representment takes the 15%
deadline penalty (and still gets graded on evidence, strategy, and
packet quality), while a case left untouched after the deadline
collapses the entire case score to 0.
"""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
progress = ctx.progress
case = ctx.case
if _final_resolution(progress) != "unresolved":
return 1.0
if (
progress.resolution_status == "pending_issuer_review"
and progress.merchant_submitted_at_step is not None
and progress.merchant_submitted_at_step <= case.deadline_step
):
return 1.0
if ctx.step_count > case.deadline_step:
return 0.0
return 1.0
class EfficiencyRubric(Rubric):
"""Penalise wasted / redundant actions, reward early correct concessions."""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
case = ctx.case
progress = ctx.progress
final = _final_resolution(progress)
wasted_actions = progress.duplicate_queries + progress.invalid_actions
efficiency = max(
0.0,
1.0 - min(0.9, wasted_actions * 0.1 + progress.submit_attempts * 0.05),
)
# Over-querying a concedable case is wasted exploration.
if (
final in {"accept_chargeback", "issue_refund"}
and case.optimal_strategy != "contest"
):
systems_queried = len(progress.revealed_systems)
if systems_queried > 2:
efficiency -= 0.15 * (systems_queried - 2)
# Retrieving policy after the decision was already made is wasted.
if progress.policy_retrieved and progress.resolved_at_step is not None:
if final in {
"accept_chargeback",
"issue_refund",
} and case.optimal_strategy in {
"accept_chargeback",
"issue_refund",
}:
efficiency -= 0.08
# Early correct concession bonus.
if (
final in {"accept_chargeback", "issue_refund"}
and case.optimal_strategy in {"accept_chargeback", "issue_refund"}
and progress.resolved_at_step is not None
and progress.resolved_at_step <= 3
):
efficiency = min(1.0, efficiency + 0.1)
return max(0.0, min(1.0, efficiency))
class OutcomeQualityRubric(Rubric):
"""Discrete outcome quality: optimal=1.0, acceptable=0.4, else 0.0."""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
final = _final_resolution(ctx.progress)
if final == ctx.case.optimal_strategy:
return 1.0
if final in ctx.case.acceptable_strategies:
return 0.4
return 0.0
def _probability_of_merchant_win(score: float) -> float:
"""Map evidence strength to arbitration win probability.
Mirrors the deterministic arbitration ruling: strong packets always win,
weak packets always lose, the ambiguity band is a 50/50 coin flip.
"""
if score >= ARB_MERCHANT_WIN_THRESHOLD:
return 1.0
if score <= ARB_ISSUER_WIN_THRESHOLD:
return 0.0
return 0.5
class EscalationROIRubric(Rubric):
"""Score the merchant's escalate-vs-concede decision on expected value.
Escalation is rational iff ``P(win) * dispute_amount > arb_fee`` β the
arbitration fee is paid by both sides regardless, so the merchant should
only pay it when the expected recovered dispute amount exceeds the fee.
Dimension is vacuous (full credit) for cases that never entered
pre-arbitration, since no escalation decision was taken.
"""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
case = ctx.case
progress = ctx.progress
if progress.round_number < 2 and progress.arbitration_outcome is None:
# Vacuous credit only when the case was never contestable.
# Conceding a contestable case before reaching the issuer review
# is a forfeit on EV grounds, not a smart decision β penalise it.
if case.optimal_strategy == "contest":
expected_contest_recovery = case.amount # P(win) at full evidence
if expected_contest_recovery > ARB_FEE_PER_SIDE:
final = _final_resolution(progress)
if final in {"accept_chargeback", "issue_refund"}:
return 0.0
return 1.0
score = evidence_strength_score(case, progress)
p_win = _probability_of_merchant_win(score)
expected_recovery = p_win * case.amount
escalate_is_positive_ev = expected_recovery > ARB_FEE_PER_SIDE
status = progress.resolution_status
if status == "won_pre_arb":
return 1.0
if status == "conceded_pre_arb":
return 0.0 if escalate_is_positive_ev else 1.0
if progress.arbitration_outcome is not None:
return 1.0 if escalate_is_positive_ev else 0.0
return 0.5
class NoteQualityRubric(Rubric):
"""Text-based representment note scorer (contest-only)."""
def forward(self, action: Any, observation: Any) -> float:
ctx: GradingContext = action
progress = ctx.progress
if _final_resolution(progress) != "contest" or not progress.representment_note:
return 0.0
return grade_representment_note(
progress.representment_note,
ctx.case,
set(progress.attached_evidence_ids),
)
def grade_representment_note(
note: str | None,
case: InternalCase,
attached_ids: set[str],
) -> float:
"""Score a representment note from 0.0 to 1.0.
Evaluates whether the note references required policy claims, mentions
attached evidence, has sufficient substance, and avoids harmful mentions.
"""
if not note or not note.strip():
return 0.0
text = note.lower()
score = 0.0
# Substance: minimum length for a coherent note.
word_count = len(text.split())
if word_count >= 5:
score += 0.2
elif word_count >= 2:
score += 0.1
# Required claims coverage: does the note mention policy requirements?
if case.policy_requirements:
claims_hit = 0
for req in case.policy_requirements:
req_keywords = req.lower().split()
if any(kw in text for kw in req_keywords if len(kw) > 3):
claims_hit += 1
score += 0.5 * _ratio(claims_hit, len(case.policy_requirements))
else:
score += 0.3
# Evidence coherence: does the note reference attached evidence?
evidence_refs = sum(
1
for eid in attached_ids
if eid.lower() in text
or any(
part in text
for part in eid.lower().replace("-", " ").split()
if len(part) > 3
)
)
if evidence_refs > 0:
score += 0.15
# Harmful mention penalty derived from each case's harmful evidence blobs.
harmful_terms: set[str] = set()
for items in case.evidence_by_system.values():
for item in items:
if item.harmful:
for word in (item.title + " " + item.summary).lower().split():
clean = word.strip(".,;:()")
if len(clean) > 3:
harmful_terms.add(clean)
harmful_terms -= {
"was",
"the",
"and",
"for",
"that",
"with",
"from",
"time",
"detail",
}
harmful_hits = sum(1 for term in harmful_terms if term in text)
if harmful_hits > 0:
score -= 0.12 * min(harmful_hits, 3)
return max(0.0, min(1.0, score))
# Weights must match the order of rubrics handed to WeightedSum and sum to 1.0.
CASE_DIMENSION_WEIGHTS: tuple[float, ...] = (
0.20,
0.15,
0.10,
0.10,
0.10,
0.10,
0.05,
0.20,
)
def _resolve_default_note_rubric() -> Rubric:
"""Return the LLM-backed note judge if opted in, else the deterministic one.
Reads ``USE_LLM_NOTE_JUDGE`` lazily so importing this module never triggers
a provider import. The LLM rubric internally falls back to
:class:`NoteQualityRubric` when no provider key is set.
"""
if os.getenv("USE_LLM_NOTE_JUDGE", "").lower() in {"1", "true", "yes"}:
try: # pragma: no cover - import-time guard
from .llm_note_judge import LLMNoteJudgeRubric
except ImportError:
from evaluation.llm_note_judge import LLMNoteJudgeRubric
return LLMNoteJudgeRubric()
return NoteQualityRubric()
CASE_DIMENSION_NAMES: tuple[str, ...] = (
"strategy_correctness",
"evidence_quality",
"packet_validity",
"deadline_compliance",
"efficiency",
"outcome_quality",
"note_quality",
"escalation_roi",
)
class CaseRubric(Rubric):
"""Per-case composite β weighted sum of the seven scoring dimensions,
hard-gated on deadline compliance.
The weighted sum lives in :class:`WeightedSum`. On top of that, a
:class:`Gate` wrapping :class:`DeadlineComplianceRubric` at
``threshold=1.0`` hard-zeros the whole case if the deadline is missed β
in real chargeback operations the best evidence in the world can't save a
case filed too late, so a late resolution must collapse the case score,
not just reduce one dimension. This is a direct use of OpenEnv's
:class:`Gate` primitive and exposes the hard-constraint pattern through
:meth:`named_rubrics`.
"""
def __init__(self, *, note_rubric: Rubric | None = None) -> None:
super().__init__()
if note_rubric is None:
note_rubric = _resolve_default_note_rubric()
self.aggregator = WeightedSum(
rubrics=[
StrategyCorrectnessRubric(),
EvidenceQualityRubric(),
PacketValidityRubric(),
DeadlineComplianceRubric(),
EfficiencyRubric(),
OutcomeQualityRubric(),
note_rubric,
EscalationROIRubric(),
],
weights=list(CASE_DIMENSION_WEIGHTS),
)
# Hard constraint: a case never even *attempted* before the deadline
# expires collapses the entire case score. This wraps
# :class:`CaseAbandonedRubric` (not :class:`DeadlineComplianceRubric`)
# so late-but-completed representments still get dimension-level
# credit while truly abandoned cases are zeroed.
self.deadline_gate = Gate(CaseAbandonedRubric(), threshold=1.0)
def forward(self, action: Any, observation: Any) -> float:
# Always run the aggregator first so per-dimension ``last_score``
# values are fresh for the grader breakdown, even when the gate
# hard-fails the case.
weighted = self.aggregator(action, observation)
if self.deadline_gate(action, observation) < 1.0:
return 0.0
return weighted
def dimension_scores(self) -> dict[str, float]:
"""Return per-dimension scores from the most recent forward pass."""
scores: dict[str, float] = {}
for name, child in zip(CASE_DIMENSION_NAMES, self.aggregator._rubric_list):
scores[name] = (
float(child.last_score) if child.last_score is not None else 0.0
)
return scores
class ChargebackOpsEpisodeRubric(Rubric):
"""Episode-level rubric: aggregate per-case scores weighted by case.weight."""
def __init__(self) -> None:
super().__init__()
self.case_rubric = CaseRubric()
def forward(self, action: Any, observation: Any) -> float:
ctx: EpisodeGradingContext = action
total = 0.0
total_weight = 0.0
for case in ctx.task.cases:
case_ctx = GradingContext(
case=case,
progress=ctx.progress_by_case[case.case_id],
step_count=ctx.step_count,
)
case_score = self.case_rubric(case_ctx, observation)
total += case_score * case.weight
total_weight += case.weight
if total_weight == 0:
return 0.0
return min(1.0, total / total_weight)
|