Spaces:
Running on Zero
Running on Zero
File size: 38,220 Bytes
40b1357 b554e59 40b1357 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 | """Build the checked evaluation dataset from reviewed source references."""
from __future__ import annotations
import argparse
import json
from collections import Counter
from pathlib import Path
from typing import Any
from .evaluation_data import (
EVALUATION_DATASET_ID,
EVALUATION_DATASET_VERSION,
AcousticCondition,
AnnotationEvidence,
AnnotationProvenance,
AnnotationStatus,
AnnotatorType,
CallAnnotation,
ChallengeCall,
ChallengeSegment,
CounterfactualAxis,
CounterfactualPair,
EvaluationDataset,
EvaluationSourceType,
ExistingCallRecord,
ExpectedAttention,
ExpectedFindingLabel,
ExpectedPairEffect,
file_sha256,
validate_evaluation_dataset,
)
from .schemas import FindingPolarity, Speaker
HERE = Path(__file__).resolve().parent
REPO_ROOT = HERE.parents[2]
TRANSCRIPT_DIR = REPO_ROOT / "data" / "sentence_segments" / "banking"
DEFAULT_OUTPUT = HERE / "research" / "evaluation_data_0_1.json"
DEFAULT_SUMMARY = (
HERE / "research" / "evaluation_data_summary_0_1.json"
)
PROFILE_ID = "banking-demo-v1"
def _source_segment(call_id: str, seq_id: int) -> dict[str, Any]:
path = TRANSCRIPT_DIR / f"{call_id}.json"
payload = json.loads(path.read_text(encoding="utf-8"))
return next(
item
for item in payload["sentences"]
if item["seq_id"] == seq_id
)
def _source_evidence(
call_id: str,
seq_id: int,
purpose: str,
) -> AnnotationEvidence:
item = _source_segment(call_id, seq_id)
return AnnotationEvidence(
evidence_id=f"evidence:{call_id}:{seq_id}",
segment_id=item["id"],
seq_id=item["seq_id"],
speaker=Speaker(item["speaker"].lower()),
start_seconds=item["start"],
end_seconds=item["end"],
quote=item["text"],
purpose=purpose,
)
def _seed_provenance() -> AnnotationProvenance:
return AnnotationProvenance(
annotator_id="codex-research-seed",
annotator_type=(
AnnotatorType.MODEL_ASSISTED_RESEARCHER_SEED
),
status=AnnotationStatus.DRAFT,
method=(
"profile-guided transcript review with exact segment "
"citations; no claim of human ground truth"
),
requires_human_review=True,
)
def _finding(
finding_type: str,
polarity: FindingPolarity,
evidence: list[AnnotationEvidence],
rationale: str,
) -> ExpectedFindingLabel:
return ExpectedFindingLabel(
finding_type=finding_type,
polarity=polarity,
evidence_ids=[item.evidence_id for item in evidence],
rationale=rationale,
)
def _existing_record(
call_id: str,
attention: ExpectedAttention,
findings: list[
tuple[
str,
FindingPolarity,
list[tuple[int, str]],
str,
]
],
uncertainty_notes: list[str] | None = None,
) -> ExistingCallRecord:
evidence_by_key: dict[tuple[int, str], AnnotationEvidence] = {}
labels = []
for finding_type, polarity, evidence_specs, rationale in findings:
finding_evidence = []
for seq_id, purpose in evidence_specs:
key = (seq_id, purpose)
evidence_by_key.setdefault(
key,
_source_evidence(call_id, seq_id, purpose),
)
finding_evidence.append(evidence_by_key[key])
labels.append(
_finding(
finding_type,
polarity,
finding_evidence,
rationale,
)
)
transcript_path = (
Path("data")
/ "sentence_segments"
/ "banking"
/ f"{call_id}.json"
)
absolute_path = REPO_ROOT / transcript_path
annotation = CallAnnotation(
annotation_id=f"annotation:{call_id}:seed",
call_id=call_id,
profile_id=PROFILE_ID,
source_type=EvaluationSourceType.EXISTING_CALL,
expected_attention=attention,
expected_findings=labels,
evidence=list(evidence_by_key.values()),
uncertainty_notes=uncertainty_notes or [],
provenance=_seed_provenance(),
)
return ExistingCallRecord(
call_id=call_id,
transcript_path=transcript_path.as_posix(),
transcript_sha256=file_sha256(absolute_path),
annotation=annotation,
)
def build_existing_records() -> list[ExistingCallRecord]:
"""Create draft labels whose citations remain tied to source files."""
return [
_existing_record(
"en_CA_Banking_1586889",
ExpectedAttention.NOT_REQUIRED,
[
(
"control.identity_verification_completed",
FindingPolarity.POSITIVE,
[(11, "Identity verification is explicitly framed.")],
"The agent requests verification before account action.",
),
(
"control.recurring_schedule_confirmed",
FindingPolarity.POSITIVE,
[(28, "Amount, accounts, frequency, and start date are recapped.")],
"The recurring transfer schedule is explicit.",
),
(
"outcome.transfer_confirmed",
FindingPolarity.POSITIVE,
[(104, "The completed setup is summarized.")],
"The call contains a clear transaction recap.",
),
],
),
_existing_record(
"en_CA_Banking_1588683",
ExpectedAttention.NOT_REQUIRED,
[
(
"control.identity_verification_completed",
FindingPolarity.POSITIVE,
[(6, "The agent begins a multi-item verification exchange.")],
"The exchange occurs before the account is opened.",
),
(
"control.recurring_schedule_confirmed",
FindingPolarity.POSITIVE,
[(69, "The recurring dates and setup are stated.")],
"The recurring schedule is explicit.",
),
(
"outcome.transfer_confirmed",
FindingPolarity.POSITIVE,
[(83, "The agent verifies acceptance and scheduled dates.")],
"The customer receives confirmation of the setup.",
),
],
[
(
"The project profile treats the multi-item exchange as "
"verification; it does not assess whether requesting a "
"full card security code is appropriate operational "
"policy."
)
],
),
_existing_record(
"en_CA_Banking_1590992",
ExpectedAttention.NOT_REQUIRED,
[
(
"control.identity_verification_completed",
FindingPolarity.POSITIVE,
[(39, "Security questions begin before execution.")],
"The agent performs a security-question exchange.",
),
(
"control.transfer_authorization_obtained",
FindingPolarity.POSITIVE,
[(158, "The customer accepts the immediate transfer.")],
"Approval follows the transfer discussion.",
),
(
"outcome.transfer_confirmed",
FindingPolarity.POSITIVE,
[(177, "The customer confirms receipt of the notification.")],
"The notification provides transaction confirmation.",
),
],
),
_existing_record(
"en_CA_Banking_1592237",
ExpectedAttention.NOT_REQUIRED,
[
(
"control.identity_verification_completed",
FindingPolarity.POSITIVE,
[(12, "The agent frames date of birth as security verification.")],
"Identity is checked before account opening.",
),
(
"control.account_opening_consent_obtained",
FindingPolarity.POSITIVE,
[(80, "The customer agrees to open the checking account.")],
"The customer gives explicit opening consent.",
),
(
"outcome.account_opening_confirmed",
FindingPolarity.POSITIVE,
[(89, "The agent confirms setup and next steps.")],
"The outcome and email follow-up are stated.",
),
],
[
(
"The transcript also introduces a savings account late "
"in the call; human review should decide whether it is "
"in scope for the selected checking-account intent."
)
],
),
_existing_record(
"en_US_General_Banking_1584540",
ExpectedAttention.NOT_REQUIRED,
[
(
"process.loan_request_confirmed",
FindingPolarity.POSITIVE,
[(62, "The agent summarizes purpose and amount.")],
"The equipment-financing purpose is established.",
),
(
"process.loan_eligibility_explained",
FindingPolarity.POSITIVE,
[(95, "The agent qualifies the likely eligibility outcome.")],
"The explanation avoids promising approval.",
),
(
"outcome.loan_next_step_confirmed",
FindingPolarity.POSITIVE,
[(100, "The agent offers information and a later reapplication path.")],
"A route to improve and return is provided.",
),
],
[
(
"The current applicability fixture classifies this as "
"a general inquiry, although the agent collects "
"financial details and gives an eligibility opinion. "
"A human must adjudicate the inquiry-versus-application "
"boundary."
)
],
),
_existing_record(
"en_US_General_Banking_1586157",
ExpectedAttention.REQUIRED,
[
(
"process.recurring_change_path_unclear",
FindingPolarity.NEGATIVE,
[(126, "Only expiration extension is explained after setup.")],
"The call does not explain how to change, pause, or cancel the recurring instruction.",
),
(
"control.identity_verification_completed",
FindingPolarity.POSITIVE,
[(8, "Date of birth is requested before account action.")],
"The agent completes a verification exchange.",
),
(
"control.recurring_schedule_confirmed",
FindingPolarity.POSITIVE,
[(121, "Amount, source, destination, frequency, and date are recapped.")],
"The recurring schedule is explicit.",
),
],
),
_existing_record(
"en_US_General_Banking_1586678",
ExpectedAttention.REQUIRED,
[
(
"control.identity_verification_missing",
FindingPolarity.NEGATIVE,
[(9, "The first account request asks for full account numbers.")],
"Account identifiers are collected without a customer-verification exchange.",
),
(
"control.transfer_authorization_missing",
FindingPolarity.NEGATIVE,
[(46, "The agent selects a schedule without later obtaining explicit approval.")],
"No explicit authorization follows the material transfer details.",
),
(
"process.recurring_change_path_explained",
FindingPolarity.POSITIVE,
[(51, "The agent explains available change and pause operations.")],
"The customer receives a usable modification path.",
),
],
),
_existing_record(
"en_US_General_Banking_1586893",
ExpectedAttention.NOT_REQUIRED,
[
(
"control.identity_verification_completed",
FindingPolarity.POSITIVE,
[(10, "Identity verification is explicitly requested.")],
"The verification exchange precedes account action.",
),
(
"process.transfer_terms_explained",
FindingPolarity.POSITIVE,
[(60, "Fees and modification paths are stated.")],
"Material recurring-transfer terms are explained.",
),
(
"outcome.transfer_confirmed",
FindingPolarity.POSITIVE,
[(102, "The agent confirms the schedule is active.")],
"The transfer setup has a clear completion statement.",
),
],
),
_existing_record(
"en_US_General_Banking_1587139",
ExpectedAttention.REQUIRED,
[
(
"escalation.explicit_dissatisfaction",
FindingPolarity.NEGATIVE,
[(60, "The customer directly calls the situation ridiculous.")],
"The explicit wording meets the review-level text rule.",
),
(
"control.identity_verification_completed",
FindingPolarity.POSITIVE,
[(9, "Date of birth is requested before the transaction.")],
"The call contains a verification exchange.",
),
(
"control.transfer_authorization_obtained",
FindingPolarity.POSITIVE,
[(79, "The customer accepts the expedited transfer.")],
"The customer gives explicit approval after timing is explained.",
),
],
[
(
"The existing acoustic result indicates later recovery, "
"but this seed does not treat that model-derived signal "
"as human-validated ground truth."
)
],
),
_existing_record(
"en_US_General_Banking_1587700",
ExpectedAttention.REQUIRED,
[
(
"control.transfer_authorization_missing",
FindingPolarity.NEGATIVE,
[(64, "The agent announces execution without a fresh approval after the details.")],
"The transcript lacks explicit authorization after amount, accounts, and timing are established.",
),
(
"control.transfer_accounts_confirmed",
FindingPolarity.POSITIVE,
[(58, "The destination account is selected explicitly.")],
"The source and destination are made unambiguous.",
),
(
"outcome.transfer_confirmed",
FindingPolarity.POSITIVE,
[(142, "The confirmation recaps amount and accounts.")],
"The customer receives a detailed completion record.",
),
],
[
(
"Human review should decide whether name plus account "
"number constitutes a clear identity-verification "
"exchange under the project profile."
)
],
),
]
def _challenge_provenance(axis: CounterfactualAxis) -> AnnotationProvenance:
return AnnotationProvenance(
annotator_id="controlled-challenge-author",
annotator_type=AnnotatorType.SYNTHETIC_AUTHOR,
status=AnnotationStatus.ADJUDICATED,
method="known-answer counterfactual authored from one declared change",
requires_human_review=False,
ground_truth_basis=(
f"the {axis.value} manipulation is explicit and all declared "
"invariants are checked"
),
)
def _challenge_segments(
call_id: str,
turns: list[tuple[Speaker, str]],
*,
seconds_per_turn: float = 4.0,
) -> list[ChallengeSegment]:
segments = []
for index, (speaker, text) in enumerate(turns, start=1):
start = round((index - 1) * seconds_per_turn, 2)
end = round(start + min(3.0, seconds_per_turn), 2)
segments.append(
ChallengeSegment(
segment_id=f"{call_id}:segment:{index}",
seq_id=index,
speaker=speaker,
start_seconds=start,
end_seconds=end,
text=text,
)
)
return segments
def _challenge_annotation(
call_id: str,
axis: CounterfactualAxis,
attention: ExpectedAttention,
segments: list[ChallengeSegment],
finding_specs: list[
tuple[str, FindingPolarity, list[int], str]
],
) -> CallAnnotation:
evidence = []
findings = []
by_seq = {item.seq_id: item for item in segments}
for finding_type, polarity, seq_ids, rationale in finding_specs:
finding_evidence = []
for seq_id in seq_ids:
segment = by_seq[seq_id]
item = AnnotationEvidence(
evidence_id=f"evidence:{call_id}:{seq_id}",
segment_id=segment.segment_id,
seq_id=segment.seq_id,
speaker=segment.speaker,
start_seconds=segment.start_seconds,
end_seconds=segment.end_seconds,
quote=segment.text,
purpose=f"Known-answer evidence for {finding_type}.",
)
if item.evidence_id not in {
known.evidence_id for known in evidence
}:
evidence.append(item)
finding_evidence.append(item)
findings.append(
_finding(
finding_type,
polarity,
finding_evidence,
rationale,
)
)
return CallAnnotation(
annotation_id=f"annotation:{call_id}:known-answer",
call_id=call_id,
profile_id=PROFILE_ID,
source_type=EvaluationSourceType.CONTROLLED_CHALLENGE,
expected_attention=attention,
expected_findings=findings,
evidence=evidence,
provenance=_challenge_provenance(axis),
)
def _challenge_call(
call_id: str,
title: str,
axis: CounterfactualAxis,
turns: list[tuple[Speaker, str]],
attention: ExpectedAttention,
finding_specs: list[
tuple[str, FindingPolarity, list[int], str]
],
*,
acoustic: AcousticCondition = AcousticCondition.CALM,
seconds_per_turn: float = 4.0,
) -> ChallengeCall:
segments = _challenge_segments(
call_id,
turns,
seconds_per_turn=seconds_per_turn,
)
return ChallengeCall(
call_id=call_id,
title=title,
intent_ids=["transfer.one_time"],
facts={
"request.executes_transaction": True,
"transfer.executed_during_call": True,
"policy.requires_identity_challenge": True,
},
acoustic_condition=acoustic,
segments=segments,
annotation=_challenge_annotation(
call_id,
axis,
attention,
segments,
finding_specs,
),
)
def _clean_turns() -> list[tuple[Speaker, str]]:
return [
(Speaker.AGENT, "Thank you for calling North Star Bank. My name is Maya."),
(Speaker.CUSTOMER, "Please transfer 200 dollars from savings to checking today."),
(Speaker.AGENT, "Before I access the accounts, please confirm your date of birth."),
(Speaker.CUSTOMER, "It is May 4, 1988."),
(Speaker.AGENT, "I have 200 dollars from savings to checking today, with no fee."),
(Speaker.CUSTOMER, "Yes, please proceed."),
(Speaker.AGENT, "The 200 dollar transfer is complete and available today."),
]
def build_challenges() -> tuple[
list[ChallengeCall],
list[CounterfactualPair],
]:
positive_clean = [
(
"control.identity_verification_completed",
FindingPolarity.POSITIVE,
[3],
"Identity is checked before account access.",
),
(
"control.transfer_authorization_obtained",
FindingPolarity.POSITIVE,
[6],
"The customer explicitly authorizes the transfer.",
),
(
"outcome.transfer_confirmed",
FindingPolarity.POSITIVE,
[7],
"The agent confirms the completed outcome.",
),
]
calls = []
pairs = []
baseline = _clean_turns() + [
(Speaker.AGENT, "A manager reviewed the account settings yesterday.")
]
variant = _clean_turns() + [
(Speaker.CUSTOMER, "I want to speak to a manager about this.")
]
calls.extend(
[
_challenge_call(
"challenge-transcript-context",
"Manager word used as account context",
CounterfactualAxis.TRANSCRIPT,
baseline,
ExpectedAttention.NOT_REQUIRED,
positive_clean,
),
_challenge_call(
"challenge-transcript-request",
"Customer explicitly requests a manager",
CounterfactualAxis.TRANSCRIPT,
variant,
ExpectedAttention.REQUIRED,
[
(
"escalation.manager_requested",
FindingPolarity.NEGATIVE,
[8],
"The customer explicitly asks for a manager.",
),
*positive_clean,
],
),
]
)
pairs.append(
CounterfactualPair(
pair_id="pair:transcript:manager-attribution",
axis=CounterfactualAxis.TRANSCRIPT,
baseline_call_id="challenge-transcript-context",
variant_call_id="challenge-transcript-request",
controlled_change=(
"A contextual agent statement becomes an explicit customer "
"manager request."
),
invariant_dimensions=[
"intent_and_facts",
"acoustic_condition",
],
expected_effect=ExpectedPairEffect.ATTENTION_CHANGES,
)
)
delivery_turns = _clean_turns() + [
(Speaker.CUSTOMER, "I am very frustrated that this took so long.")
]
delivery_findings = [
(
"escalation.explicit_dissatisfaction",
FindingPolarity.NEGATIVE,
[8],
"The explicit text requires review in both delivery conditions.",
),
*positive_clean,
]
calls.extend(
[
_challenge_call(
"challenge-delivery-calm",
"Explicit dissatisfaction with calm delivery",
CounterfactualAxis.DELIVERY,
delivery_turns,
ExpectedAttention.REQUIRED,
delivery_findings,
acoustic=AcousticCondition.CALM,
),
_challenge_call(
"challenge-delivery-elevated",
"Explicit dissatisfaction with persistent elevation",
CounterfactualAxis.DELIVERY,
delivery_turns,
ExpectedAttention.REQUIRED,
delivery_findings,
acoustic=AcousticCondition.PERSISTENT_ELEVATION,
),
]
)
pairs.append(
CounterfactualPair(
pair_id="pair:delivery:acoustic-support",
axis=CounterfactualAxis.DELIVERY,
baseline_call_id="challenge-delivery-calm",
variant_call_id="challenge-delivery-elevated",
controlled_change=(
"Only the authored acoustic condition changes from calm "
"to persistent elevation."
),
invariant_dimensions=[
"intent_and_facts",
"transcript",
"attention",
],
expected_effect=ExpectedPairEffect.SUPPORT_CHANGES,
)
)
outcome_base = _clean_turns()
outcome_variant = _clean_turns()
outcome_variant[-1] = (
Speaker.AGENT,
"I have noted your transfer request.",
)
calls.extend(
[
_challenge_call(
"challenge-outcome-confirmed",
"Completed transfer has a clear confirmation",
CounterfactualAxis.OUTCOME,
outcome_base,
ExpectedAttention.NOT_REQUIRED,
positive_clean,
),
_challenge_call(
"challenge-outcome-unclear",
"Completed transfer lacks outcome confirmation",
CounterfactualAxis.OUTCOME,
outcome_variant,
ExpectedAttention.REQUIRED,
[
(
"outcome.transfer_confirmation_missing",
FindingPolarity.NEGATIVE,
[7],
"The closing statement does not say whether execution succeeded.",
),
*positive_clean[:2],
],
),
]
)
pairs.append(
CounterfactualPair(
pair_id="pair:outcome:completion-confirmation",
axis=CounterfactualAxis.OUTCOME,
baseline_call_id="challenge-outcome-confirmed",
variant_call_id="challenge-outcome-unclear",
controlled_change=(
"The completion statement becomes an ambiguous request "
"acknowledgement."
),
invariant_dimensions=[
"intent_and_facts",
"acoustic_condition",
],
expected_effect=ExpectedPairEffect.ATTENTION_CHANGES,
)
)
identity_base = _clean_turns()
identity_variant = _clean_turns()
identity_variant[2] = (
Speaker.AGENT,
"I can see the savings and checking accounts on your profile.",
)
identity_variant[3] = (
Speaker.CUSTOMER,
"Those are the correct accounts.",
)
calls.extend(
[
_challenge_call(
"challenge-control-identity-present",
"Identity verification precedes account action",
CounterfactualAxis.CONTROL,
identity_base,
ExpectedAttention.NOT_REQUIRED,
positive_clean,
),
_challenge_call(
"challenge-control-identity-missing",
"Account action proceeds without identity verification",
CounterfactualAxis.CONTROL,
identity_variant,
ExpectedAttention.REQUIRED,
[
(
"control.identity_verification_missing",
FindingPolarity.NEGATIVE,
[3],
"The agent accesses account details without a verification exchange.",
),
*positive_clean[1:],
],
),
]
)
pairs.append(
CounterfactualPair(
pair_id="pair:control:identity-verification",
axis=CounterfactualAxis.CONTROL,
baseline_call_id="challenge-control-identity-present",
variant_call_id="challenge-control-identity-missing",
controlled_change=(
"The identity request and response become unverified "
"account acknowledgement."
),
invariant_dimensions=[
"intent_and_facts",
"acoustic_condition",
],
expected_effect=ExpectedPairEffect.ATTENTION_CHANGES,
)
)
auth_base = _clean_turns()
auth_variant = _clean_turns()
auth_variant[5] = (
Speaker.AGENT,
"I am processing that transfer now.",
)
calls.extend(
[
_challenge_call(
"challenge-control-authorization-present",
"Explicit approval follows material terms",
CounterfactualAxis.CONTROL,
auth_base,
ExpectedAttention.NOT_REQUIRED,
positive_clean,
),
_challenge_call(
"challenge-control-authorization-missing",
"Execution proceeds without explicit approval",
CounterfactualAxis.CONTROL,
auth_variant,
ExpectedAttention.REQUIRED,
[
(
"control.transfer_authorization_missing",
FindingPolarity.NEGATIVE,
[6],
"The agent executes without customer approval after the recap.",
),
positive_clean[0],
positive_clean[2],
],
),
]
)
pairs.append(
CounterfactualPair(
pair_id="pair:control:transfer-authorization",
axis=CounterfactualAxis.CONTROL,
baseline_call_id="challenge-control-authorization-present",
variant_call_id="challenge-control-authorization-missing",
controlled_change=(
"The customer's explicit approval becomes an agent "
"execution statement."
),
invariant_dimensions=[
"intent_and_facts",
"acoustic_condition",
],
expected_effect=ExpectedPairEffect.ATTENTION_CHANGES,
)
)
recovery_turns = _clean_turns() + [
(Speaker.CUSTOMER, "I am very frustrated about the delay."),
(Speaker.AGENT, "I understand. The transfer is complete now."),
(Speaker.CUSTOMER, "Thank you for resolving it."),
]
recovery_negative = (
"escalation.explicit_dissatisfaction",
FindingPolarity.NEGATIVE,
[8],
"Explicit dissatisfaction remains reviewable after recovery.",
)
calls.extend(
[
_challenge_call(
"challenge-recovery-observed",
"Elevated customer delivery settles before close",
CounterfactualAxis.RECOVERY,
recovery_turns,
ExpectedAttention.REQUIRED,
[
recovery_negative,
*positive_clean,
(
"escalation.recovery_observed",
FindingPolarity.POSITIVE,
[10],
"The authored delivery condition settles after resolution.",
),
],
acoustic=AcousticCondition.RECOVERED_ELEVATION,
),
_challenge_call(
"challenge-recovery-unresolved",
"Elevated customer delivery remains unresolved",
CounterfactualAxis.RECOVERY,
recovery_turns,
ExpectedAttention.REQUIRED,
[recovery_negative, *positive_clean],
acoustic=AcousticCondition.UNRESOLVED_ELEVATION,
),
]
)
pairs.append(
CounterfactualPair(
pair_id="pair:recovery:settled-versus-unresolved",
axis=CounterfactualAxis.RECOVERY,
baseline_call_id="challenge-recovery-observed",
variant_call_id="challenge-recovery-unresolved",
controlled_change=(
"Only the authored late-call acoustic trajectory changes "
"from recovered to unresolved."
),
invariant_dimensions=[
"intent_and_facts",
"transcript",
"attention",
],
expected_effect=ExpectedPairEffect.RECOVERY_CHANGES,
)
)
duration_turns = _clean_turns()
calls.extend(
[
_challenge_call(
"challenge-duration-short",
"Clean transfer at ordinary spacing",
CounterfactualAxis.DURATION,
duration_turns,
ExpectedAttention.NOT_REQUIRED,
positive_clean,
seconds_per_turn=4.0,
),
_challenge_call(
"challenge-duration-long",
"Same clean transfer with long spacing",
CounterfactualAxis.DURATION,
duration_turns,
ExpectedAttention.NOT_REQUIRED,
positive_clean,
seconds_per_turn=18.0,
),
]
)
pairs.append(
CounterfactualPair(
pair_id="pair:duration:spacing-invariance",
axis=CounterfactualAxis.DURATION,
baseline_call_id="challenge-duration-short",
variant_call_id="challenge-duration-long",
controlled_change=(
"Only segment timing and total duration are expanded."
),
invariant_dimensions=[
"intent_and_facts",
"transcript",
"acoustic_condition",
"attention",
],
expected_effect=ExpectedPairEffect.NO_DECISION_CHANGE,
)
)
return calls, pairs
def build_dataset() -> EvaluationDataset:
challenges, pairs = build_challenges()
return EvaluationDataset(
dataset_id=EVALUATION_DATASET_ID,
dataset_version=EVALUATION_DATASET_VERSION,
profile_id=PROFILE_ID,
existing_calls=build_existing_records(),
challenge_calls=challenges,
counterfactual_pairs=pairs,
)
def build_summary(dataset: EvaluationDataset) -> dict[str, Any]:
existing_statuses = Counter(
item.annotation.provenance.status.value
for item in dataset.existing_calls
)
existing_attention = Counter(
item.annotation.expected_attention.value
for item in dataset.existing_calls
)
axes = Counter(
item.axis.value for item in dataset.counterfactual_pairs
)
challenge_attention = Counter(
item.annotation.expected_attention.value
for item in dataset.challenge_calls
)
return {
"schema_version": "1.0",
"dataset_id": dataset.dataset_id,
"dataset_version": dataset.dataset_version,
"profile_id": dataset.profile_id,
"existing_calls": {
"count": len(dataset.existing_calls),
"annotation_status_counts": dict(
sorted(existing_statuses.items())
),
"expected_attention_counts": dict(
sorted(existing_attention.items())
),
"human_adjudicated_count": sum(
item.annotation.provenance.annotator_type
== AnnotatorType.HUMAN
and item.annotation.provenance.status
== AnnotationStatus.ADJUDICATED
for item in dataset.existing_calls
),
"requires_human_review_count": sum(
item.annotation.provenance.requires_human_review
for item in dataset.existing_calls
),
},
"controlled_challenges": {
"call_count": len(dataset.challenge_calls),
"pair_count": len(dataset.counterfactual_pairs),
"pair_axis_counts": dict(sorted(axes.items())),
"expected_attention_counts": dict(
sorted(challenge_attention.items())
),
"known_answer_count": sum(
item.annotation.provenance.status
== AnnotationStatus.ADJUDICATED
for item in dataset.challenge_calls
),
},
}
def main() -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--output", type=Path, default=DEFAULT_OUTPUT)
parser.add_argument(
"--summary-output",
type=Path,
default=DEFAULT_SUMMARY,
)
args = parser.parse_args()
dataset = build_dataset()
validate_evaluation_dataset(dataset, REPO_ROOT)
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(
json.dumps(dataset.model_dump(mode="json"), indent=2) + "\n",
encoding="utf-8",
)
args.summary_output.parent.mkdir(parents=True, exist_ok=True)
args.summary_output.write_text(
json.dumps(build_summary(dataset), indent=2) + "\n",
encoding="utf-8",
)
print(
f"Wrote {len(dataset.existing_calls)} existing-call labels, "
f"{len(dataset.challenge_calls)} challenge calls, and "
f"{len(dataset.counterfactual_pairs)} pairs."
)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|