figment / scripts /generate_finetune_data.py
ThomsenDrake's picture
Sync full submission repo state
94cbe85 verified
Raw
History Blame Contribute Delete
224 kB
"""Generate Figment local-4B supervised fine-tuning data.
The generator creates synthetic, de-identified protocol-navigation cases,
asks the Ultra teacher for candidate gold navigator JSON, validates candidates
with Figment's deterministic gates, and writes accepted SFT rows plus a
manifest. It intentionally does not copy locked eval rows or NVIDIA dataset
rows.
"""
from __future__ import annotations
import argparse
from collections import Counter
from dataclasses import dataclass
from dataclasses import replace
from datetime import UTC
from datetime import datetime
import hashlib
import httpx
import json
import multiprocessing
import os
from pathlib import Path
import re
import sys
from time import perf_counter
from time import sleep
from typing import Any
import urllib.parse
PROJECT_ROOT = Path(__file__).resolve().parents[1]
if str(PROJECT_ROOT) not in sys.path:
sys.path.insert(0, str(PROJECT_ROOT))
from figment.config import NVIDIA_API_BASE_URL # noqa: E402
from figment.config import load_config # noqa: E402
from figment.eval_metrics import bucket_expected_observation_cues, score_expected_labels # noqa: E402
from figment.harness_evidence import build_harness_evidence # noqa: E402
from figment.model_client import ModelClientError # noqa: E402
from figment.model_client import canned_navigator_output # noqa: E402
from figment.observation_targets import CARD_IDS_EXEMPT_FROM_OBSERVATION_TARGETS # noqa: E402
from figment.observation_targets import TRACE_ONLY_REQUIRED_OBSERVATION_IDS_KEY # noqa: E402
from figment.observation_targets import apply_navigation_scaffolding # noqa: E402
from figment.observation_targets import required_observation_targets # noqa: E402
from figment.prompt_builder import REQUIRED_JSON_SKELETON # noqa: E402
from figment.prompt_builder import SYSTEM_PROMPT # noqa: E402
from figment.prompt_builder import build_prompt # noqa: E402
from figment.retrieval import load_protocol_cards # noqa: E402
from figment.retrieval import query_from_intake # noqa: E402
from figment.retrieval import search_protocol_cards # noqa: E402
from figment.rules import run_red_flag_checks # noqa: E402
from figment.trace import stable_hash # noqa: E402
from figment.validators import urgency_floor_from_rules # noqa: E402
from figment.validators import validate_navigator_output # noqa: E402
TEACHER_MODEL_ID = "nvidia/nemotron-3-ultra-550b-a55b"
DATASET_VERSION = "figment_sft_v1"
OUTPUT_PATH = Path("data/finetune/figment_sft_v1.jsonl")
MANIFEST_PATH = Path("data/finetune/figment_sft_v1_manifest.json")
CASE_SPEC_PATH = Path("data/finetune/figment_sft_v1_case_specs.jsonl")
CLINICAL_CARD_IDS = (
"AMS-RED-FLAGS-v1",
"CHEST-PAIN-ESCALATION-v1",
"PED-DEHYD-RED-FLAGS-v1",
"FEVER-RED-FLAGS-v1",
"PREG-DANGER-SIGNS-v1",
"RESP-DISTRESS-RED-FLAGS-v1",
"STROKE-SIGNS-v1",
"WOUND-INFECTION-ESCALATION-v1",
)
SAFETY_CARD_ID = "SAFETY-BOUNDARIES-v1"
SBAR_CARD_ID = "REFERRAL-SBAR-v1"
FAILURE_DISTRIBUTION = (
("missing_observation_cues", 40),
("negation_safety_boundary", 20),
("source_card_candidate_pathway", 15),
("sbar_grounding", 15),
("forbidden_instruction_avoidance", 5),
("fallback_rescue_shape", 5),
)
V2_FAILURE_DISTRIBUTION = (
("missing_observation_cues", 40),
("negation_safety_boundary", 25),
("source_card_candidate_pathway", 20),
("sbar_grounding", 10),
("forbidden_instruction_avoidance", 3),
("fallback_rescue_shape", 2),
)
V3_FAILURE_DISTRIBUTION = (
("rural_clinic_intake", 18),
("disaster_triage", 16),
("radio_handoff", 8),
("asr_confirmed_text", 6),
("escalation_precision", 14),
("missing_observation_prioritization", 14),
("sbar_handoff_usefulness", 10),
("source_card_discipline", 6),
("low_resource_constraints", 7),
("workflow_repair_seed", 1),
)
V4_FAILURE_DISTRIBUTION = (
("radio_handoff", 25),
("sbar_handoff_usefulness", 22),
("source_card_discipline", 14),
("low_resource_constraints", 10),
("missing_observation_prioritization", 10),
("workflow_repair_seed", 7),
("rural_clinic_intake", 4),
("disaster_triage", 3),
("escalation_precision", 5),
)
V5_FOCUSED_COUNTS = {
"sbar_observation_ownership": 350,
"required_observation_id_selection": 250,
"source_card_invariant": 150,
"noisy_field_audio_style": 100,
"general_regression": 250,
}
V5_FAILURE_DISTRIBUTION = tuple(V5_FOCUSED_COUNTS.items())
V5_EXCLUDED_EVAL_CASE_IDS = ("field_workflow_holdout_v1-000054", "field_workflow_holdout_v1-000099")
def _weighted_cycle_from_counts(counts: dict[str, int]) -> tuple[str, ...]:
produced: Counter[str] = Counter()
schedule: list[str] = []
order = {name: index for index, name in enumerate(counts)}
total = sum(counts.values())
while len(schedule) < total:
remaining = [name for name, target in counts.items() if produced[name] < target]
name = max(
remaining,
key=lambda item: (
(counts[item] - produced[item]) / counts[item],
-order[item],
),
)
schedule.append(name)
produced[name] += 1
return tuple(schedule)
V6_NAVIGATOR_COUNTS = {
"required_observation_ownership": 900,
"v6_preservation": 100,
"observation_correction": 180,
}
V6_FAILURE_DISTRIBUTION = tuple(V6_NAVIGATOR_COUNTS.items())
V6_FAILURE_CYCLE = _weighted_cycle_from_counts(V6_NAVIGATOR_COUNTS)
V7_NAVIGATOR_COUNTS = {
"source_card_closure": 240,
"observation_source_joint": 140,
"distractor_card_resistance": 100,
"sbar_source_coupling": 80,
}
V7_FAILURE_DISTRIBUTION = tuple(V7_NAVIGATOR_COUNTS.items())
V7_FAILURE_CYCLE = _weighted_cycle_from_counts(V7_NAVIGATOR_COUNTS)
V8_NAVIGATOR_COUNTS = {
"multi_rule_observation_ownership": 320,
"multi_rule_candidate_focus": 80,
}
V8_FAILURE_DISTRIBUTION = tuple(V8_NAVIGATOR_COUNTS.items())
V8_FAILURE_CYCLE = _weighted_cycle_from_counts(V8_NAVIGATOR_COUNTS)
V9_NAVIGATOR_COUNTS = {
"postpartum_fever_required_obs_cross_category": 320,
"postpartum_fever_required_obs_candidate_focus": 80,
}
V9_FAILURE_DISTRIBUTION = tuple(V9_NAVIGATOR_COUNTS.items())
V9_FAILURE_CYCLE = _weighted_cycle_from_counts(V9_NAVIGATOR_COUNTS)
V10_NAVIGATOR_COUNTS = {
"postpartum_fever_required_obs_dual_field_closure": 640,
"postpartum_fever_required_obs_candidate_focus": 160,
}
V10_FAILURE_DISTRIBUTION = tuple(V10_NAVIGATOR_COUNTS.items())
V10_FAILURE_CYCLE = _weighted_cycle_from_counts(V10_NAVIGATOR_COUNTS)
V11_NAVIGATOR_COUNTS = {
"postpartum_fever_required_obs_visible_dual_field_holdout_shape": 520,
"postpartum_fever_required_obs_dual_field_closure": 200,
"postpartum_fever_required_obs_candidate_focus": 80,
}
V11_FAILURE_DISTRIBUTION = tuple(V11_NAVIGATOR_COUNTS.items())
V11_FAILURE_CYCLE = _weighted_cycle_from_counts(V11_NAVIGATOR_COUNTS)
V12_NAVIGATOR_COUNTS = {
"postpartum_fever_required_obs_dual_card_selected_ids_visible_fields": 280,
"postpartum_fever_required_obs_candidate_and_source_closure": 80,
"wound_source_card_schema_replay": 120,
"referral_candidate_pathway_replay": 80,
}
V12_FAILURE_DISTRIBUTION = tuple(V12_NAVIGATOR_COUNTS.items())
V12_FAILURE_CYCLE = _weighted_cycle_from_counts(V12_NAVIGATOR_COUNTS)
V13_NAVIGATOR_COUNTS = {
"postpartum_fever_required_obs_visible_preg_source_card_cue_closure": 520,
"postpartum_fever_required_obs_visible_preg_candidate_pathway_closure": 160,
"postpartum_fever_required_obs_selected_id_compressed_field_repair": 160,
"wound_source_card_schema_replay": 80,
"referral_candidate_pathway_replay": 80,
}
V13_FAILURE_DISTRIBUTION = tuple(V13_NAVIGATOR_COUNTS.items())
V13_FAILURE_CYCLE = _weighted_cycle_from_counts(V13_NAVIGATOR_COUNTS)
V14_NAVIGATOR_COUNTS = {
"postpartum_fever_required_obs_visible_preg_source_card_cue_closure": 520,
"postpartum_fever_required_obs_visible_preg_candidate_pathway_closure": 160,
"postpartum_fever_required_obs_selected_id_compressed_field_repair": 160,
"wound_source_card_schema_replay": 200,
"referral_candidate_pathway_replay": 80,
}
V14_FAILURE_DISTRIBUTION = tuple(V14_NAVIGATOR_COUNTS.items())
V14_FAILURE_CYCLE = _weighted_cycle_from_counts(V14_NAVIGATOR_COUNTS)
V6_HARNESS_METADATA_OBSERVATION_CUES = (
"source card ids",
"source protocol card ids",
"retrieved protocol card ids",
"deterministic rule results",
"navigator validation result",
"confirmed intake status",
"manual correction status for audio-derived fields",
)
FORBIDDEN_BEHAVIOR = [
"Do not diagnose.",
"Do not prescribe, dose, administer, or start medication.",
"Do not discharge, send home, or replace local protocol.",
"Do not downgrade deterministic red flags.",
]
SCORER_SAFE_FORBIDDEN_BEHAVIOR = [
"Use protocol navigation only.",
"Do not provide clinical orders or treatment instructions.",
"Do not replace local protocol.",
"Keep deterministic red flags visible.",
]
V2_FORBIDDEN_LEXICAL_PATTERNS = {
"diagnose": re.compile(r"\bdiagnos(?:e|is|ed)\b", re.IGNORECASE),
"prescribe": re.compile(r"\bprescrib(?:e|ing|ed)\b", re.IGNORECASE),
"dose": re.compile(r"\b(?:dose|dosing)\b|\bgive \d+(?:\.\d+)?\s*(?:mg|mcg|ml)\b", re.IGNORECASE),
"administer": re.compile(r"\b(?:administer|inject|dispense)\b", re.IGNORECASE),
"medication": re.compile(r"\b(?:aspirin|antibiotic|opioid|insulin|medication|medications|drug)\b", re.IGNORECASE),
"discharge": re.compile(r"\b(?:discharge(?: home)?|send (?:them|the patient|the person|him|her)?\s*home)\b", re.IGNORECASE),
}
V3_GENERIC_OUTPUT_PATTERNS = (
re.compile(r"\brepeat\s+vitals?\b", re.IGNORECASE),
re.compile(r"\bmonitor\s+closely\b", re.IGNORECASE),
re.compile(r"\bfollow\s+protocol\b", re.IGNORECASE),
re.compile(r"\bassess\s+(?:the\s+)?patient\b", re.IGNORECASE),
re.compile(r"\bcollect\s+more\s+information\b", re.IGNORECASE),
re.compile(r"\bcontinue\s+to\s+observe\b", re.IGNORECASE),
)
V5_GENERIC_OBSERVATION_PATTERNS = V3_GENERIC_OUTPUT_PATTERNS + (
re.compile(r"\bask\s+(?:anything|everything)\s+else\b", re.IGNORECASE),
re.compile(r"\bkeep\s+monitoring\b", re.IGNORECASE),
re.compile(r"\bcheck\s+vitals?\b", re.IGNORECASE),
)
V3_SBAR_FAILURE_CLASSES = {"radio_handoff", "sbar_handoff_usefulness", "workflow_repair_seed"}
TEACHER_NOTE_MAX_TOKENS = 320
def dataset_paths(dataset_version: str) -> dict[str, Path]:
"""Return the default local artifact paths for a fine-tune dataset version."""
root = Path("data/finetune")
return {
"output": root / f"{dataset_version}.jsonl",
"manifest": root / f"{dataset_version}_manifest.json",
"case_specs": root / f"{dataset_version}_case_specs.jsonl",
}
def _exclusion_paths_for_generation(dataset_version: str, configured_paths: list[Path] | None) -> list[Path]:
if configured_paths:
return [path for path in configured_paths if path.exists()]
if not uses_v3_field_workflow_policy(dataset_version) or dataset_version.startswith("field_workflow_holdout"):
return []
candidates = [
Path("data/eval/initial_handwritten_cases.jsonl"),
Path("data/eval/adversarial_strict_cases.jsonl"),
Path("data/eval/comprehensive_hosted_cases.jsonl"),
Path("data/eval/field_workflow_holdout_v1.jsonl"),
]
return [path for path in candidates if path.exists()]
def load_exclusion_signatures(paths: list[Path]) -> list[ExclusionSignature]:
signatures: list[ExclusionSignature] = []
for path in paths:
for line_number, line in enumerate(path.read_text(encoding="utf-8").splitlines(), start=1):
if not line.strip():
continue
item = json.loads(line)
if not isinstance(item, dict) or not isinstance(item.get("structured_intake"), dict):
continue
signatures.append(
ExclusionSignature(
case_id=str(item.get("case_id") or f"{path}:{line_number}"),
source_path=str(path),
target_protocol_card_id=str(item.get("target_protocol_card_id") or ""),
workflow_category=str(item.get("workflow_category") or item.get("structured_intake", {}).get("workflow_category") or ""),
clinical_hash=_clinical_intake_hash(item["structured_intake"]),
tokens=frozenset(_clinical_intake_tokens(item["structured_intake"])),
)
)
return signatures
def case_index_for_attempt(attempt_index: int, *, start_index: int = 0, index_stride: int = 1) -> int:
"""Map a local attempt number to a globally unique synthetic case index."""
return start_index + (attempt_index * index_stride)
def uses_v3_field_workflow_policy(dataset_version: str) -> bool:
"""Return whether a dataset version should use v3 field-workflow behavior."""
return (
dataset_version.startswith("figment_sft_v3")
or dataset_version.startswith("figment_sft_v4")
or uses_v5_focused_policy(dataset_version)
or uses_v6_observation_policy(dataset_version)
or dataset_version.startswith("field_workflow_holdout")
)
def uses_v5_focused_policy(dataset_version: str) -> bool:
"""Return whether a dataset version should use v5 focused-training behavior."""
return dataset_version.startswith("figment_sft_v5")
def uses_v7_source_card_policy(dataset_version: str) -> bool:
"""Return whether a dataset version should use v7 source-card closure behavior."""
return dataset_version.startswith("figment_sft_v7") or uses_v8_multirule_policy(dataset_version)
def uses_v8_multirule_policy(dataset_version: str) -> bool:
"""Return whether a dataset version targets multi-fired-card observation ownership."""
return dataset_version.startswith("figment_sft_v8") or uses_v9_perfect_eval_policy(dataset_version)
def uses_v9_perfect_eval_policy(dataset_version: str) -> bool:
"""Return whether a dataset version targets the remaining v8 holdout gaps."""
return dataset_version.startswith("figment_sft_v9") or uses_v10_perfect_eval_policy(dataset_version)
def uses_v10_perfect_eval_policy(dataset_version: str) -> bool:
"""Return whether a dataset version targets the remaining v9 scaffold-dependence gaps."""
return (
dataset_version.startswith("figment_sft_v10")
or uses_v11_perfect_eval_policy(dataset_version)
or uses_v12_perfect_eval_policy(dataset_version)
or uses_v13_perfect_eval_policy(dataset_version)
)
def uses_v11_perfect_eval_policy(dataset_version: str) -> bool:
"""Return whether a dataset version targets the remaining v10 dual-field visibility gaps."""
return dataset_version.startswith("figment_sft_v11")
def uses_v12_perfect_eval_policy(dataset_version: str) -> bool:
"""Return whether a dataset version targets v11 regression recovery plus v10 gap closure."""
return dataset_version.startswith("figment_sft_v12")
def uses_v13_perfect_eval_policy(dataset_version: str) -> bool:
"""Return whether a dataset version targets the remaining v12 FEVER/PREG visible-field gap."""
return dataset_version.startswith("figment_sft_v13") or uses_v14_perfect_eval_policy(dataset_version)
def uses_v14_perfect_eval_policy(dataset_version: str) -> bool:
"""Return whether a dataset version fully covers the v13 partial-delta regression shape."""
return dataset_version.startswith("figment_sft_v14")
def uses_v6_observation_policy(dataset_version: str) -> bool:
"""Return whether a dataset version should use v6 observation-ownership behavior."""
return dataset_version.startswith("figment_sft_v6") or uses_v7_source_card_policy(dataset_version)
def forbidden_behavior_for_version(dataset_version: str) -> list[str]:
"""Return assistant boundary text compatible with the dataset's scoring target."""
if dataset_version == "figment_sft_v2" or uses_v3_field_workflow_policy(dataset_version):
return list(SCORER_SAFE_FORBIDDEN_BEHAVIOR)
return list(FORBIDDEN_BEHAVIOR)
def safety_boundary_for_version(dataset_version: str) -> str:
if dataset_version == "figment_sft_v2" or uses_v3_field_workflow_policy(dataset_version):
return "Prototype protocol navigation only; trained-responder review required; no clinical orders or autonomous routing."
return "Prototype protocol navigation only; no condition label, medication order, or autonomous routing."
@dataclass(frozen=True)
class TeacherClient:
endpoint: str
model_id: str
auth_headers: dict[str, str]
timeout_seconds: float
max_tokens: int
endpoint_env: str
api_key_env: str
@dataclass(frozen=True)
class SyntheticCase:
case_id: str
dataset_version: str
failure_class: str
target_protocol_card_id: str
structured_intake: dict[str, Any]
tags: list[str]
high_risk: bool
@dataclass
class PreparedCase:
spec: SyntheticCase
rule_results: list[dict[str, Any]]
urgency_floor: str
retrieved_cards: list[dict[str, Any]]
retrieved_ids: list[str]
prompt: str
prompt_hash: str
expected_source_card_ids: list[str]
expected_candidate_pathway_card_ids: list[str]
expected_missing_observations: list[str]
expected_red_flag_rule_ids: list[str]
@dataclass(frozen=True)
class ExclusionSignature:
case_id: str
source_path: str
target_protocol_card_id: str
workflow_category: str
clinical_hash: str
tokens: frozenset[str]
@dataclass
class CandidateResult:
output: dict[str, Any]
validation: dict[str, Any]
expected_label_score: dict[str, Any]
reward_components: dict[str, int]
reward_score: int
patched_fields: list[str]
filled_required_observation_ids: list[str]
model_selected_required_observation_ids: list[str]
invalid_selected_required_observation_ids: list[str]
stripped_trace_only_fields: list[str]
raw_output_hash: str
@property
def passed(self) -> bool:
return (
self.validation.get("passed") is True
and self.expected_label_score.get("all_expected_labels_passed") is True
and all(self.reward_components.values())
)
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--dataset-version", default=DATASET_VERSION)
parser.add_argument("--count", type=int, default=500, help="Accepted SFT rows to write.")
parser.add_argument("--output", type=Path, default=None)
parser.add_argument("--manifest", type=Path, default=None)
parser.add_argument("--case-specs", type=Path, default=None)
parser.add_argument("--teacher-model-id", default=TEACHER_MODEL_ID)
parser.add_argument("--timeout-seconds", type=float, default=120.0)
parser.add_argument("--teacher-max-tokens", type=int, default=TEACHER_NOTE_MAX_TOKENS)
parser.add_argument("--candidate-count", type=int, default=1)
parser.add_argument("--high-risk-candidate-count", type=int, default=1)
parser.add_argument("--max-attempts", type=int, default=None)
parser.add_argument("--teacher-error-retries", type=int, default=0)
parser.add_argument("--teacher-error-sleep-seconds", type=float, default=5.0)
parser.add_argument(
"--no-teacher-worker",
action="store_true",
help="Call the teacher in-process instead of through a forked timeout worker.",
)
parser.add_argument("--start-index", type=int, default=0, help="First synthetic case index for this shard.")
parser.add_argument("--index-stride", type=int, default=1, help="Synthetic case index stride for disjoint shards.")
parser.add_argument(
"--exclusion-eval",
action="append",
type=Path,
default=None,
help="Eval JSONL to reject exact or near-neighbor generated training rows against.",
)
parser.add_argument("--resume", action="store_true", help="Append until count is reached if output exists.")
parser.add_argument("--dry-run", action="store_true", help="Use deterministic fallback instead of teacher calls.")
parser.add_argument("--log-rejections", action="store_true", help="Print JSONL progress for rejected candidates.")
args = parser.parse_args(argv)
if args.count <= 0:
raise SystemExit("--count must be positive")
if args.start_index < 0:
raise SystemExit("--start-index must be non-negative")
if args.index_stride <= 0:
raise SystemExit("--index-stride must be positive")
if args.teacher_error_retries < 0:
raise SystemExit("--teacher-error-retries must be non-negative")
if args.teacher_error_sleep_seconds < 0:
raise SystemExit("--teacher-error-sleep-seconds must be non-negative")
paths = dataset_paths(args.dataset_version)
args.output = args.output or paths["output"]
args.manifest = args.manifest or paths["manifest"]
args.case_specs = args.case_specs or paths["case_specs"]
cards = load_protocol_cards()
cards_by_id = {str(card["card_id"]): card for card in cards}
missing_cards = sorted({*CLINICAL_CARD_IDS, SAFETY_CARD_ID, SBAR_CARD_ID} - set(cards_by_id))
if missing_cards:
raise SystemExit(f"missing protocol cards: {', '.join(missing_cards)}")
exclusion_paths = _exclusion_paths_for_generation(args.dataset_version, args.exclusion_eval)
exclusion_signatures = load_exclusion_signatures(exclusion_paths)
existing_rows = _load_existing_rows(args.output) if args.resume else []
accepted: list[dict[str, Any]] = list(existing_rows)
accepted_ids = {row.get("case_id") for row in accepted}
manifest_events: list[dict[str, Any]] = []
counters: Counter[str] = Counter()
candidate_totals: Counter[str] = Counter()
rejection_reasons: Counter[str] = Counter()
for row in accepted:
category = str(row.get("category") or row.get("metadata", {}).get("failure_class") or "unknown")
counters[category] += 1
counters.update(f"tag:{tag}" for tag in row.get("tags", []))
metadata = row.get("metadata", {})
candidate_totals["total"] += int(metadata.get("pass_rate_total") or 0)
candidate_totals["passed"] += int(metadata.get("pass_rate_passed") or 0)
client = _teacher_client(args.teacher_model_id, args.timeout_seconds, args.teacher_max_tokens) if not args.dry_run else None
started_at = datetime.now(UTC)
max_attempts = args.max_attempts or args.count * 4
attempt_index = 0
args.output.parent.mkdir(parents=True, exist_ok=True)
args.case_specs.parent.mkdir(parents=True, exist_ok=True)
mode = "a" if args.resume and args.output.exists() else "w"
spec_mode = "a" if args.resume and args.case_specs.exists() else "w"
with args.output.open(mode, encoding="utf-8") as output_file, args.case_specs.open(spec_mode, encoding="utf-8") as spec_file:
while len(accepted) < args.count and attempt_index < max_attempts:
case_index = case_index_for_attempt(
attempt_index,
start_index=args.start_index,
index_stride=args.index_stride,
)
attempt_index += 1
spec = generate_case_spec(case_index, cards_by_id, dataset_version=args.dataset_version)
if spec.case_id in accepted_ids:
continue
exclusion_match = _eval_exclusion_neighbor(spec, exclusion_signatures)
if exclusion_match:
rejection_reasons[exclusion_match["reason"]] += 1
manifest_events.append(
{
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"accepted": False,
**exclusion_match,
}
)
if args.log_rejections:
_print_progress_event(
{
"accepted": len(accepted),
"target": args.count,
"case_id": spec.case_id,
"failure_class": spec.failure_class,
**exclusion_match,
}
)
continue
prepared = prepare_case(spec, cards_by_id)
harness_gap = _harness_retrieval_gap(prepared)
if harness_gap:
rejection_reasons[harness_gap["reason"]] += 1
manifest_events.append(
{
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"accepted": False,
**harness_gap,
}
)
if args.log_rejections:
_print_progress_event(
{
"accepted": len(accepted),
"target": args.count,
"case_id": spec.case_id,
"failure_class": spec.failure_class,
**harness_gap,
}
)
continue
candidate_count = args.high_risk_candidate_count if spec.high_risk else args.candidate_count
candidate_count = max(1, min(candidate_count, 12))
try:
raw_candidates = _raw_candidates_with_retries(
client=client,
prepared=prepared,
teacher_model_id=args.teacher_model_id,
candidate_count=candidate_count,
dry_run=args.dry_run,
use_worker=not args.no_teacher_worker,
teacher_error_retries=args.teacher_error_retries,
teacher_error_sleep_seconds=args.teacher_error_sleep_seconds,
)
except ModelClientError as exc:
rejection_reasons["teacher_backend_error"] += 1
safe_error = _safe_error_text(str(exc))
manifest_events.append(
{
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"accepted": False,
"reason": safe_error,
}
)
if args.log_rejections:
_print_progress_event(
{
"accepted": len(accepted),
"target": args.count,
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"reason": "teacher_backend_error",
"error": safe_error,
}
)
continue
candidate_results = [score_candidate(candidate, prepared) for candidate in raw_candidates]
if not candidate_results:
rejection_reasons["no_candidates"] += 1
if args.log_rejections:
_print_progress_event(
{
"accepted": len(accepted),
"target": args.count,
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"reason": "no_candidates",
}
)
continue
best = max(candidate_results, key=lambda item: (item.passed, item.reward_score, -len(item.patched_fields)))
passed_count = sum(1 for item in candidate_results if item.passed)
candidate_totals["total"] += len(candidate_results)
candidate_totals["passed"] += passed_count
if not best.passed:
failure_key = _rejection_key(best)
policy_issues = _policy_issues_for_prepared(best.output, prepared)
rejection_reasons[failure_key] += 1
manifest_events.append(
{
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"accepted": False,
"reason": failure_key,
"validation_failures": best.validation.get("failures", []),
"expected_label_score": best.expected_label_score,
"reward_components": best.reward_components,
"policy_issues": policy_issues,
}
)
if args.log_rejections:
_print_progress_event(
{
"accepted": len(accepted),
"target": args.count,
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"reason": failure_key,
"validation_failures": best.validation.get("failures", [])[:3],
"expected_label_failed": [
key for key, value in best.expected_label_score.items() if value is False
][:5],
"failed_rewards": [key for key, value in best.reward_components.items() if not value],
"policy_issues": policy_issues[:8],
}
)
continue
row = build_sft_row(
prepared=prepared,
result=best,
teacher_model_id=args.teacher_model_id,
candidate_total=len(candidate_results),
candidate_passed=passed_count,
)
output_file.write(json.dumps(row, sort_keys=True) + "\n")
output_file.flush()
spec_file.write(json.dumps(case_spec_record(prepared), sort_keys=True) + "\n")
spec_file.flush()
accepted.append(row)
accepted_ids.add(spec.case_id)
counters[spec.failure_class] += 1
counters.update(f"tag:{tag}" for tag in spec.tags)
manifest_events.append(
{
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"accepted": True,
"candidate_total": len(candidate_results),
"candidate_passed": passed_count,
"reward_score": best.reward_score,
"patched_fields": best.patched_fields,
}
)
print(
json.dumps(
{
"accepted": len(accepted),
"target": args.count,
"case_id": spec.case_id,
"failure_class": spec.failure_class,
"candidate_passed": passed_count,
"candidate_total": len(candidate_results),
},
sort_keys=True,
),
flush=True,
)
manifest = build_manifest(
output_path=args.output,
case_specs_path=args.case_specs,
dataset_version=args.dataset_version,
rows=accepted,
started_at=started_at,
teacher_model_id=args.teacher_model_id,
dry_run=args.dry_run,
attempts=attempt_index,
start_index=args.start_index,
index_stride=args.index_stride,
counters=counters,
candidate_totals=candidate_totals,
rejection_reasons=rejection_reasons,
events=manifest_events,
exclusion_paths=exclusion_paths,
exclusion_signature_count=len(exclusion_signatures),
)
args.manifest.parent.mkdir(parents=True, exist_ok=True)
args.manifest.write_text(json.dumps(manifest, indent=2, sort_keys=True) + "\n", encoding="utf-8")
print(json.dumps({"manifest": str(args.manifest), "rows": len(accepted), "attempts": attempt_index}, sort_keys=True))
return 0 if len(accepted) >= args.count else 1
def _teacher_client(teacher_model_id: str, timeout_seconds: float, max_tokens: int) -> TeacherClient:
base = load_config()
openrouter = _openrouter_config_for_teacher_model(teacher_model_id)
if openrouter:
endpoint, api_key = openrouter
return TeacherClient(
endpoint=endpoint,
model_id=teacher_model_id,
auth_headers={"Authorization": f"Bearer {api_key}"},
timeout_seconds=timeout_seconds,
max_tokens=max(64, max_tokens),
endpoint_env="OPENROUTER_BASE_URL",
api_key_env="OPENROUTER_API_KEY",
)
config = replace(base, model_backend="hosted_omni", nvidia_model_id=teacher_model_id).validated()
endpoint = config.omni_endpoint_url or config.hf_endpoint_url or config.nvidia_base_url
if not endpoint:
raise ModelClientError("teacher model requires NVIDIA_BASE_URL, OMNI_ENDPOINT_URL, or HF_ENDPOINT_URL")
auth_headers: dict[str, str] = {}
endpoint_env = _endpoint_env_name(teacher_model_id)
api_key_env = ""
if "integrate.api.nvidia.com" in endpoint:
if not config.nvidia_api_key:
raise ModelClientError("teacher NVIDIA endpoint requires NVIDIA_API_KEY")
auth_headers["Authorization"] = f"Bearer {config.nvidia_api_key}"
api_key_env = "NVIDIA_API_KEY"
elif config.hf_token:
auth_headers["Authorization"] = f"Bearer {config.hf_token}"
api_key_env = "HF_TOKEN"
return TeacherClient(
endpoint=endpoint,
model_id=teacher_model_id,
auth_headers=auth_headers,
timeout_seconds=timeout_seconds,
max_tokens=max(64, max_tokens),
endpoint_env=endpoint_env,
api_key_env=api_key_env,
)
def _openrouter_config_for_teacher_model(teacher_model_id: str) -> tuple[str, str] | None:
configured_model = os.getenv("OPENROUTER_FREE_MODEL_ID", "").strip()
if configured_model and teacher_model_id != configured_model:
return None
if not configured_model and not teacher_model_id.endswith(":free"):
return None
endpoint = os.getenv("OPENROUTER_BASE_URL", "").strip()
api_key = os.getenv("OPENROUTER_API_KEY", "").strip()
if not endpoint and not api_key:
return None
if not endpoint:
raise ModelClientError("OpenRouter teacher model requires OPENROUTER_BASE_URL")
if not api_key:
raise ModelClientError("OpenRouter teacher model requires OPENROUTER_API_KEY")
return endpoint, api_key
def generate_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str = DATASET_VERSION,
) -> SyntheticCase:
failure_class = _failure_class_for_index(index, dataset_version=dataset_version)
if uses_v13_perfect_eval_policy(dataset_version):
return _generate_v13_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if uses_v12_perfect_eval_policy(dataset_version):
return _generate_v12_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if uses_v10_perfect_eval_policy(dataset_version):
return _generate_v10_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if uses_v9_perfect_eval_policy(dataset_version):
return _generate_v9_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if uses_v8_multirule_policy(dataset_version):
return _generate_v8_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if uses_v7_source_card_policy(dataset_version):
return _generate_v7_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if uses_v6_observation_policy(dataset_version):
return _generate_v6_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if uses_v5_focused_policy(dataset_version):
return _generate_v5_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if uses_v3_field_workflow_policy(dataset_version):
return _generate_v3_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if failure_class == "negation_safety_boundary":
target = SAFETY_CARD_ID
intake = _negated_intake(index)
tags = ["negation", "safety_boundary"]
high_risk = True
elif failure_class == "forbidden_instruction_avoidance":
target = SAFETY_CARD_ID
intake = _forbidden_intake(index)
tags = ["forbidden_instruction", "safety_boundary"]
high_risk = True
elif failure_class == "sbar_grounding":
target = SBAR_CARD_ID
card_id = CLINICAL_CARD_IDS[index % len(CLINICAL_CARD_IDS)]
intake = _positive_intake(card_id, index, handoff=True)
tags = ["sbar", _tag_for_card(card_id)]
high_risk = True
elif failure_class == "fallback_rescue_shape":
target, intake = _fallback_rescue_intake(index)
tags = ["fallback_rescue", _tag_for_card(target)]
high_risk = True
else:
card_id = CLINICAL_CARD_IDS[index % len(CLINICAL_CARD_IDS)]
target = card_id
intake = _positive_intake(card_id, index, handoff=False)
tags = [_tag_for_card(card_id), failure_class]
high_risk = failure_class == "source_card_candidate_pathway"
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=tags,
high_risk=high_risk,
)
def _generate_v8_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
target = "FEVER-RED-FLAGS-v1"
source_index = index + 2400
intake = _multi_rule_observation_ownership_intake(source_index, failure_class=failure_class)
intake = _apply_v3_workflow_context(
intake,
index=index,
category=failure_class,
target_card_id=target,
dataset_version=dataset_version,
)
intake["multi_rule_observation_focus"] = (
"Gold output must cite FEVER-RED-FLAGS-v1 and every fired pregnancy/postpartum rule card, "
"select all required-observation target ids for those clinical source cards, and make each selected "
"observation visible before deterministic scaffolding would fill it."
)
if failure_class == "multi_rule_candidate_focus":
intake["candidate_pathway_focus"] = (
"Keep candidate_protocol_pathways focused on the primary FEVER target while keeping the fired "
"PREG source card and its required observations visible in source_cards and observation fields."
)
if target not in cards_by_id:
raise KeyError(f"missing target card for v8 spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(
[
"field_workflow",
"v8",
"multi_rule_observation_ownership",
"fever_red_flags",
"preg_danger_signs",
_field_tag_for_category(failure_class),
]
),
high_risk=True,
)
def _generate_v10_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
target = "FEVER-RED-FLAGS-v1"
if uses_v14_perfect_eval_policy(dataset_version):
source_index = index + 13200
elif uses_v13_perfect_eval_policy(dataset_version):
source_index = index + 11200
elif uses_v12_perfect_eval_policy(dataset_version):
source_index = index + 9200
elif uses_v11_perfect_eval_policy(dataset_version):
source_index = index + 7200
else:
source_index = index + 5200
workflow_category = _v10_postpartum_workflow_category(index, failure_class)
intake = _postpartum_fever_required_obs_intake(source_index, failure_class=failure_class)
if uses_v13_perfect_eval_policy(dataset_version):
symptoms = str(intake.get("symptoms") or "")
for phrase in (
", no chest pain reported",
"; chest pain denied",
"; no chest pressure reported",
):
symptoms = symptoms.replace(phrase, "")
intake["symptoms"] = symptoms
intake = _apply_v3_workflow_context(
intake,
index=index,
category=workflow_category,
target_card_id=target,
dataset_version=dataset_version,
)
intake["v10_training_focus"] = (
"Prior local v9 outputs cited FEVER-RED-FLAGS-v1 and PREG-DANGER-SIGNS-v1 but selected only "
"FEVER required-observation ids, so deterministic scaffolding filled the pregnancy danger-sign "
"observations. Gold output must select every non-exempt FEVER and PREG required-observation id, "
"and visible observation text for both cards must already appear in missing_info_to_collect and "
"next_observations_to_collect before deterministic scaffolding."
)
intake["cross_card_observation_closure_focus"] = (
"When a secondary fired clinical card is in source_cards or candidate_protocol_pathways, close its "
"required observations too. Do not stop observation planning at the target FEVER card."
)
if uses_v11_perfect_eval_policy(dataset_version):
intake["v11_training_focus"] = (
"Prior local v10 outputs selected every FEVER and PREG required-observation id but only wrote the "
"FEVER-side cues plus pregnancy status into visible observation fields. Gold output must front-load "
"PREG-DANGER-SIGNS-v1 visible cues in both missing_info_to_collect and next_observations_to_collect: "
"pregnancy or postpartum status, bleeding report, abdominal pain report, headache or vision symptoms, "
"seizure or fainting report, and fever report. Do not rely on selected_required_observation_ids alone."
)
if uses_v12_perfect_eval_policy(dataset_version):
intake["v12_training_focus"] = (
"V12 resumes from the v10 adapter, not v11. Preserve v10 source-card, schema, urgency, red-flag, "
"and handoff behavior while fixing the remaining postpartum FEVER plus PREG visible observation gap. "
"Gold output must select every FEVER and PREG required-observation id and make each cue visible in "
"both missing_info_to_collect and next_observations_to_collect without deterministic scaffold fill."
)
if uses_v13_perfect_eval_policy(dataset_version):
intake["v13_training_focus"] = (
"V12 still produced a few FEVER/PREG holdout rows where source_cards and candidate_protocol_pathways "
"included PREG-DANGER-SIGNS-v1, but visible observation fields only contained FEVER-side cues until "
"deterministic repair. Gold output must put the PREG danger-sign cues directly into both "
"missing_info_to_collect and next_observations_to_collect whenever PREG-DANGER-SIGNS-v1 appears in "
"source_cards or candidate_protocol_pathways: pregnancy or postpartum status, bleeding report, "
"abdominal pain report, headache or vision symptoms, seizure or fainting report, and fever report. "
"Selected_required_observation_ids are necessary but not sufficient."
)
intake["v13_visible_observation_contract"] = {
"failure_to_avoid": "FEVER-only missing_info_to_collect or next_observations_to_collect is incorrect when PREG is cited.",
"preg_cues_required_in_missing_info_to_collect": [
"pregnancy or postpartum status",
"bleeding report",
"abdominal pain report",
"headache or vision symptoms",
"seizure or fainting report",
"fever report",
],
"preg_cues_required_in_next_observations_to_collect": [
"pregnancy or postpartum status",
"bleeding report",
"abdominal pain report",
"headache or vision symptoms",
"seizure or fainting report",
"fever report",
],
}
if uses_v14_perfect_eval_policy(dataset_version):
intake["v14_training_focus"] = (
"V13 only trained on a partial delta and still required deterministic patches on FEVER plus PREG "
"observation fields. Preserve v12/v10 source-card and schema behavior while fully covering the "
"visible PREG danger-sign cues in both missing_info_to_collect and next_observations_to_collect."
)
if failure_class == "postpartum_fever_required_obs_candidate_focus":
intake["candidate_pathway_focus"] = (
"Keep candidate_protocol_pathways to FEVER-RED-FLAGS-v1 and the fired PREG-DANGER-SIGNS-v1 card; "
"avoid unrelated retrieved distractor cards."
)
if failure_class == "postpartum_fever_required_obs_candidate_and_source_closure":
intake["candidate_pathway_focus"] = (
"Keep candidate_protocol_pathways to FEVER-RED-FLAGS-v1 and the fired PREG-DANGER-SIGNS-v1 card, "
"and keep source_cards closed over FEVER, PREG, SAFETY, and REFERRAL support cards."
)
if failure_class == "postpartum_fever_required_obs_visible_preg_candidate_pathway_closure":
intake["candidate_pathway_focus"] = (
"Candidate pathways must include FEVER-RED-FLAGS-v1 and PREG-DANGER-SIGNS-v1, and the visible "
"observation fields must include the PREG danger-sign cues even though FEVER remains the target."
)
if failure_class == "postpartum_fever_required_obs_selected_id_compressed_field_repair":
intake["selected_id_visible_text_repair_focus"] = (
"Do not compress the answer into selected_required_observation_ids only. The responder-facing "
"missing and next-observation lists must spell out the PREG danger-sign observations before scaffold fill."
)
if target not in cards_by_id:
raise KeyError(f"missing target card for v10 spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(
[
"field_workflow",
(
"v14"
if uses_v14_perfect_eval_policy(dataset_version)
else "v13"
if uses_v13_perfect_eval_policy(dataset_version)
else "v12"
if uses_v12_perfect_eval_policy(dataset_version)
else "v11"
if uses_v11_perfect_eval_policy(dataset_version)
else "v10"
),
"postpartum_fever_required_obs",
"multi_rule_observation_ownership",
"required_observation_ownership",
"dual_field_observation_closure",
*(
["visible_observation_text_closure", "preg_danger_signs_front_loaded"]
if (
uses_v11_perfect_eval_policy(dataset_version)
or uses_v12_perfect_eval_policy(dataset_version)
or uses_v13_perfect_eval_policy(dataset_version)
)
else []
),
"fever_red_flags",
"preg_danger_signs",
_field_tag_for_category(workflow_category),
failure_class,
]
),
high_risk=True,
)
def _generate_v13_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
postpartum_classes = {
"postpartum_fever_required_obs_visible_preg_source_card_cue_closure",
"postpartum_fever_required_obs_visible_preg_candidate_pathway_closure",
"postpartum_fever_required_obs_selected_id_compressed_field_repair",
}
if failure_class in postpartum_classes:
return _generate_v10_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if failure_class == "wound_source_card_schema_replay":
return _generate_v12_wound_replay_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if failure_class == "referral_candidate_pathway_replay":
return _generate_v12_referral_replay_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
raise ValueError(f"unsupported v13 failure class: {failure_class}")
def _generate_v12_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
postpartum_classes = {
"postpartum_fever_required_obs_dual_card_selected_ids_visible_fields",
"postpartum_fever_required_obs_candidate_and_source_closure",
}
if failure_class in postpartum_classes:
return _generate_v10_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if failure_class == "wound_source_card_schema_replay":
return _generate_v12_wound_replay_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
if failure_class == "referral_candidate_pathway_replay":
return _generate_v12_referral_replay_case_spec(
index,
cards_by_id,
dataset_version=dataset_version,
failure_class=failure_class,
)
raise ValueError(f"unsupported v12 failure class: {failure_class}")
def _generate_v12_wound_replay_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
target = "WOUND-INFECTION-ESCALATION-v1"
if uses_v14_perfect_eval_policy(dataset_version):
version_label = "v14"
source_index = index + 13400
elif uses_v13_perfect_eval_policy(dataset_version):
version_label = "v13"
source_index = index + 11400
else:
version_label = "v12"
source_index = index + 9400
intake = _positive_intake(target, source_index, handoff=True)
intake = _apply_v3_workflow_context(
intake,
index=index,
category="source_card_closure",
target_card_id=target,
dataset_version=dataset_version,
)
intake[f"{version_label}_training_focus"] = (
f"Replay v10-passing wound behavior while training {version_label}. Gold output must keep the complete "
"navigator schema, cite WOUND-INFECTION-ESCALATION-v1, SAFETY-BOUNDARIES-v1, and REFERRAL-SBAR-v1 in "
"source_cards, select wound required-observation ids, and keep the SBAR grounded in wound facts without "
"hallucinated pregnancy."
)
intake["source_card_closure_focus"] = (
"Wound rows protect source-card closure and schema completion while preserving a clinical candidate pathway."
)
if target not in cards_by_id:
raise KeyError(f"missing target card for {version_label} wound spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(
[
"field_workflow",
version_label,
"wound_replay",
"source_card_closure",
"schema_completion",
"handoff_grounding",
"wound_infection",
]
),
high_risk=True,
)
def _generate_v12_referral_replay_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
target = SBAR_CARD_ID
if uses_v14_perfect_eval_policy(dataset_version):
version_label = "v14"
source_index = index + 13600
elif uses_v13_perfect_eval_policy(dataset_version):
version_label = "v13"
source_index = index + 11600
else:
version_label = "v12"
source_index = index + 9600
intake = _postpartum_fever_required_obs_intake(source_index, failure_class=failure_class)
intake = _apply_v3_workflow_context(
intake,
index=index,
category="sbar_source_coupling",
target_card_id=target,
dataset_version=dataset_version,
)
intake[f"{version_label}_training_focus"] = (
f"Replay the v10-passing referral/SBAR pathway shape while training {version_label}. Gold output must keep "
"REFERRAL-SBAR-v1 as the target candidate pathway while also retaining fired FEVER and PREG clinical pathways "
"before deterministic candidate scaffolding. Source cards must stay closed over REFERRAL, SAFETY, FEVER, and "
"PREG support."
)
intake["candidate_pathway_focus"] = (
"Candidate pathways should include REFERRAL-SBAR-v1 plus fired FEVER and PREG clinical cards; avoid unrelated "
"distractor cards."
)
if target not in cards_by_id:
raise KeyError(f"missing target card for {version_label} referral spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(
[
"field_workflow",
version_label,
"referral_candidate_replay",
"sbar_source_coupling",
"candidate_pathway_replay",
"fever_red_flags",
"preg_danger_signs",
]
),
high_risk=True,
)
def _generate_v9_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
target = "FEVER-RED-FLAGS-v1"
source_index = index + 3200
workflow_category = _v9_postpartum_workflow_category(index, failure_class)
intake = _postpartum_fever_required_obs_intake(source_index, failure_class=failure_class)
intake = _apply_v3_workflow_context(
intake,
index=index,
category=workflow_category,
target_card_id=target,
dataset_version=dataset_version,
)
intake["v9_training_focus"] = (
"Gold output must treat FEVER-RED-FLAGS-v1 and PREG-DANGER-SIGNS-v1 as jointly fired clinical "
"source cards. It must select every mandatory required-observation target id for both cards and copy "
"each mandatory_required_observation_targets display_text into missing_info_to_collect before any "
"deterministic scaffold could fill the field."
)
if failure_class == "postpartum_fever_required_obs_candidate_focus":
intake["candidate_pathway_focus"] = (
"Keep candidate_protocol_pathways focused on FEVER-RED-FLAGS-v1 plus the fired PREG-DANGER-SIGNS-v1 "
"card; do not add unrelated distractors."
)
if target not in cards_by_id:
raise KeyError(f"missing target card for v9 spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(
[
"field_workflow",
"v9",
"postpartum_fever_required_obs",
"multi_rule_observation_ownership",
"required_observation_ownership",
"fever_red_flags",
"preg_danger_signs",
_field_tag_for_category(workflow_category),
failure_class,
]
),
high_risk=True,
)
def _generate_v7_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
card_id = CLINICAL_CARD_IDS[index % len(CLINICAL_CARD_IDS)]
target = card_id
source_index = index + 1600
handoff = True
if failure_class == "source_card_closure" and index % 4 == 0:
triad = ("PREG-DANGER-SIGNS-v1", "CHEST-PAIN-ESCALATION-v1", "FEVER-RED-FLAGS-v1")
target = triad[(index // 4) % len(triad)]
card_id = target
intake = _multi_rule_source_closure_intake(source_index)
else:
if failure_class == "sbar_source_coupling":
target = SBAR_CARD_ID
card_id = _pick(
(
"CHEST-PAIN-ESCALATION-v1",
"PREG-DANGER-SIGNS-v1",
"RESP-DISTRESS-RED-FLAGS-v1",
"STROKE-SIGNS-v1",
),
index,
)
elif failure_class == "distractor_card_resistance":
target = _pick(
(
"CHEST-PAIN-ESCALATION-v1",
"FEVER-RED-FLAGS-v1",
"PREG-DANGER-SIGNS-v1",
"STROKE-SIGNS-v1",
),
index,
)
card_id = target
elif failure_class == "observation_source_joint":
target = _pick(CLINICAL_CARD_IDS, index + 3)
card_id = target
intake = _positive_intake(card_id, source_index, handoff=handoff)
intake = _apply_v3_workflow_context(
intake,
index=index,
category=failure_class,
target_card_id=target,
dataset_version=dataset_version,
)
if failure_class == "source_card_closure":
intake["source_card_closure_focus"] = (
"Gold output must cite every clinical target it relies on plus SAFETY-BOUNDARIES-v1 for protocol-only "
"safety text and REFERRAL-SBAR-v1 for the SBAR handoff."
)
elif failure_class == "observation_source_joint":
intake["observation_source_joint_focus"] = (
"Gold output must keep selected required-observation ids visible in observation text while also closing "
"clinical, safety, and SBAR source-card citations."
)
elif failure_class == "distractor_card_resistance":
intake["distractor_card_focus"] = (
"Retrieved context may include distractor protocol cards. Cite the target, safety, and SBAR support cards "
"without adding irrelevant clinical source cards."
)
elif failure_class == "sbar_source_coupling":
intake["sbar_source_focus"] = (
"Gold output must make the SBAR handoff concise and cite REFERRAL-SBAR-v1 alongside the clinical and "
"safety cards that support the handoff."
)
tag = _tag_for_card(card_id if target in {SAFETY_CARD_ID, SBAR_CARD_ID} else target)
tags = ["field_workflow", "v7", _field_tag_for_category(failure_class), tag, "source_card_closure"]
if target == SBAR_CARD_ID:
tags.append("sbar")
if target == SAFETY_CARD_ID:
tags.append("safety_boundary")
if failure_class == "distractor_card_resistance":
tags.append("distractor_resistance")
if failure_class == "observation_source_joint":
tags.append("required_observation_ownership")
if target not in cards_by_id and target not in {SAFETY_CARD_ID, SBAR_CARD_ID}:
raise KeyError(f"missing target card for v7 spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(tags),
high_risk=True,
)
def _generate_v6_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
card_id = CLINICAL_CARD_IDS[index % len(CLINICAL_CARD_IDS)]
target = card_id
handoff = index % 4 == 0
source_index = index + 1200
if failure_class == "observation_correction":
source_index = index + 1300
handoff = index % 3 == 0
elif failure_class == "v6_preservation":
target = _pick([card_id, SBAR_CARD_ID, SAFETY_CARD_ID], index)
handoff = target == SBAR_CARD_ID or index % 3 == 0
source_index = index + 1400
if target == SAFETY_CARD_ID:
intake = _negated_intake(source_index)
else:
intake = _positive_intake(card_id, source_index, handoff=handoff)
intake = _apply_v3_workflow_context(
intake,
index=index,
category=failure_class,
target_card_id=target,
dataset_version=dataset_version,
)
if failure_class == "observation_correction":
intake["previous_model_failure_note"] = (
"Prior local output duplicated missing and next observation lists or treated harness metadata as "
"medic observations. Gold output must rewrite those fields as clinical observation work only."
)
elif failure_class == "required_observation_ownership":
intake["required_observation_focus"] = (
"Gold output must select required_observation_targets and express them as concrete responder-facing "
"missing and next observations before deterministic scaffolding would fill them."
)
elif failure_class == "v6_preservation":
intake["preservation_focus"] = (
"Preserve v5 source-card, SBAR, urgency, red-flag, low-resource, and safety behavior while keeping "
"observation fields free of harness metadata."
)
tag = _tag_for_card(card_id if target in {SAFETY_CARD_ID, SBAR_CARD_ID} else target)
tags = ["field_workflow", "v6", _field_tag_for_category(failure_class), tag]
if target == SAFETY_CARD_ID:
tags.append("safety_boundary")
if target == SBAR_CARD_ID:
tags.append("sbar")
if failure_class == "observation_correction":
tags.append("teacher_rewrite_correction")
if target not in cards_by_id and target not in {SAFETY_CARD_ID, SBAR_CARD_ID}:
raise KeyError(f"missing target card for v6 spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(tags),
high_risk=True,
)
def _generate_v5_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
card_id = CLINICAL_CARD_IDS[index % len(CLINICAL_CARD_IDS)]
target = card_id
handoff = index % 3 == 0
source_index = index + 700
if failure_class == "sbar_observation_ownership":
target = SBAR_CARD_ID
handoff = True
source_index = index + 800
elif failure_class == "source_card_invariant":
target = ("STROKE-SIGNS-v1", "PREG-DANGER-SIGNS-v1")[index % 2]
card_id = target
handoff = index % 4 == 0
source_index = index + 900
elif failure_class == "noisy_field_audio_style":
handoff = index % 2 == 0
source_index = index + 1000
elif failure_class == "general_regression":
target = _pick([card_id, SBAR_CARD_ID, SAFETY_CARD_ID], index)
handoff = target == SBAR_CARD_ID or index % 4 == 0
source_index = index + 1100
if target == SAFETY_CARD_ID:
intake = _negated_intake(source_index)
else:
intake = _positive_intake(card_id, source_index, handoff=handoff)
if failure_class == "noisy_field_audio_style":
intake["responder_note"] = (
str(intake.get("responder_note") or "").strip()
+ " Confirmed ASR-like field note: punctuation was sparse, repeated words were removed, "
"and the responder accepted these fields before navigation."
).strip()
intake["transcript_quality"] = "confirmed_noisy_field_audio"
intake = _apply_v3_workflow_context(
intake,
index=index,
category=failure_class,
target_card_id=target,
dataset_version=dataset_version,
)
tag = _tag_for_card(card_id if target in {SAFETY_CARD_ID, SBAR_CARD_ID} else target)
tags = ["field_workflow", "v5", _field_tag_for_category(failure_class), tag]
if target == SAFETY_CARD_ID:
tags.append("safety_boundary")
if target == SBAR_CARD_ID:
tags.append("sbar")
if failure_class == "source_card_invariant":
tags.append("fired_rule_source_card")
if target not in cards_by_id and target not in {SAFETY_CARD_ID, SBAR_CARD_ID}:
raise KeyError(f"missing target card for v5 spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(tags),
high_risk=True,
)
def _generate_v3_case_spec(
index: int,
cards_by_id: dict[str, dict[str, Any]],
*,
dataset_version: str,
failure_class: str,
) -> SyntheticCase:
card_id = CLINICAL_CARD_IDS[index % len(CLINICAL_CARD_IDS)]
high_risk = failure_class in {
"radio_handoff",
"escalation_precision",
"sbar_handoff_usefulness",
"source_card_discipline",
"low_resource_constraints",
"workflow_repair_seed",
}
if failure_class in {"radio_handoff", "sbar_handoff_usefulness"}:
target = SBAR_CARD_ID
intake = _positive_intake(card_id, index + 200, handoff=True)
elif failure_class == "asr_confirmed_text":
target = card_id
intake = _positive_intake(card_id, index + 300, handoff=index % 2 == 0)
elif failure_class == "escalation_precision" and (dataset_version.startswith("figment_sft_v4") or index % 5 == 0):
target = SAFETY_CARD_ID
intake = _negated_intake(index + 400)
elif failure_class == "workflow_repair_seed":
target = SBAR_CARD_ID if index % 2 else card_id
intake = _positive_intake(card_id, index + 500, handoff=True)
else:
target = card_id
intake = _positive_intake(card_id, index + 100, handoff=False)
intake = _apply_v3_workflow_context(
intake,
index=index,
category=failure_class,
target_card_id=target,
dataset_version=dataset_version,
)
tag = _tag_for_card(card_id if target in {SAFETY_CARD_ID, SBAR_CARD_ID} else target)
tags = ["field_workflow", _field_tag_for_category(failure_class), tag]
if target == SAFETY_CARD_ID:
tags.append("safety_boundary")
if target == SBAR_CARD_ID:
tags.append("sbar")
# Touch cards_by_id in this path so missing fixture cards fail near generation.
if target not in cards_by_id and target not in {SAFETY_CARD_ID, SBAR_CARD_ID}:
raise KeyError(f"missing target card for v3 spec: {target}")
return SyntheticCase(
case_id=f"{dataset_version}-{index:06d}",
dataset_version=dataset_version,
failure_class=failure_class,
target_protocol_card_id=target,
structured_intake=intake,
tags=_dedupe(tags),
high_risk=high_risk,
)
def _apply_v3_workflow_context(
intake: dict[str, Any],
*,
index: int,
category: str,
target_card_id: str,
dataset_version: str,
) -> dict[str, Any]:
updated = dict(intake)
settings = {
"rural_clinic_intake": "rural clinic intake desk with one medic",
"disaster_triage": "flood shelter disaster triage table",
"radio_handoff": "radio and runner handoff station",
"asr_confirmed_text": "mobile clinic confirmed transcript desk",
"escalation_precision": "field escalation review point",
"missing_observation_prioritization": "crowded intake line with incomplete vitals",
"sbar_handoff_usefulness": "transport coordinator radio handoff",
"source_card_discipline": "paper protocol binder review desk",
"low_resource_constraints": "remote aid post with limited equipment",
"workflow_repair_seed": "handoff repair desk after weak navigator output",
"sbar_observation_ownership": "transport coordinator SBAR handoff desk",
"required_observation_id_selection": "rural intake line with sparse observations",
"source_card_invariant": "red-flag rule audit station",
"noisy_field_audio_style": "mobile clinic confirmed audio transcript desk",
"general_regression": "mixed field workflow review station",
"required_observation_ownership": "rural intake observation-planning desk",
"observation_correction": "navigator output correction desk",
"v6_preservation": "mixed field workflow preservation review station",
"source_card_closure": "source-card closure review desk",
"observation_source_joint": "source-card and observation ownership desk",
"distractor_card_resistance": "protocol binder review desk with distractor cards",
"sbar_source_coupling": "SBAR source-card coupling handoff desk",
"multi_rule_observation_ownership": "multi-rule maternal fever observation desk",
"multi_rule_candidate_focus": "primary-pathway multi-rule review desk",
}
supplies = {
"rural_clinic_intake": "paper protocol binder, radio, shared BP cuff, no pulse oximeter",
"disaster_triage": "gloves, cot tags, paper forms, intermittent radio, no transport yet",
"radio_handoff": "runner note, radio, paper SBAR slip, no full chart",
"asr_confirmed_text": "responder-confirmed transcript, radio, paper form, no raw audio retained",
"escalation_precision": "radio, protocol binder, transport callback list, vitals partly pending",
"missing_observation_prioritization": "paper form, radio, basic vitals kit, only a few minutes per patient",
"sbar_handoff_usefulness": "radio, SBAR form, transport list, receiving clinician callback pending",
"source_card_discipline": "protocol binder with relevant and distractor cards, radio",
"low_resource_constraints": "no pulse oximeter, no BP cuff, intermittent radio only",
"workflow_repair_seed": "previous navigator output, protocol binder, radio, paper handoff form",
"sbar_observation_ownership": "SBAR form, radio, paper protocol cards, receiving callback pending",
"required_observation_id_selection": "paper intake form, radio, basic vitals kit, two-minute queue pressure",
"source_card_invariant": "deterministic red-flag sheet, protocol binder, retrieval printout",
"noisy_field_audio_style": "accepted ASR transcript, radio, paper form, no raw audio retained",
"general_regression": "protocol binder, radio, sparse vitals kit, transport callback list",
"required_observation_ownership": "required-observation target card, paper form, radio, sparse vitals kit",
"observation_correction": "previous weak navigator output, required-observation target card, radio",
"v6_preservation": "protocol binder, radio, SBAR form, sparse vitals kit, transport callback list",
"source_card_closure": "protocol binder, safety boundary card, SBAR form, radio",
"observation_source_joint": "required-observation target card, protocol binder, SBAR form, radio",
"distractor_card_resistance": "protocol binder with relevant and distractor cards, radio, SBAR slip",
"sbar_source_coupling": "SBAR form, radio, target clinical card, safety boundary card",
"multi_rule_observation_ownership": "fever card, pregnancy danger-sign card, SBAR form, radio",
"multi_rule_candidate_focus": "protocol binder with fever primary card, pregnancy source card, and distractors",
}
goals = {
"rural_clinic_intake": "speed intake and surface the next useful missing observations.",
"disaster_triage": "keep escalation and handoff useful despite noisy sparse notes.",
"radio_handoff": "turn fragmented confirmed notes into compact grounded SBAR.",
"asr_confirmed_text": "handle corrected ASR-like confirmed text without hallucinating.",
"escalation_precision": "preserve true red flags and avoid escalating denied danger words.",
"missing_observation_prioritization": "put the highest-value observations first.",
"sbar_handoff_usefulness": "make the handoff concise, grounded, and actionable.",
"source_card_discipline": "cite only relevant retrieved cards and avoid distractor leakage.",
"low_resource_constraints": "ask for alternatives when equipment is unavailable.",
"workflow_repair_seed": "repair only weak fields while preserving validated facts.",
"sbar_observation_ownership": "make SBAR depend on model-owned observation fields, not deterministic fill.",
"required_observation_id_selection": "select required observation ids and render responder-facing text.",
"source_card_invariant": "cite every deterministic fired-rule card even when retrieval is imperfect.",
"noisy_field_audio_style": "handle confirmed noisy field transcript text without hallucinating facts.",
"general_regression": "preserve v4 strengths while avoiding locked-eval overfit.",
"required_observation_ownership": "make selected required observations explicit without scaffold fill.",
"observation_correction": "rewrite weak observation fields into clinical, responder-owned observations.",
"v6_preservation": "preserve v5 strengths while correcting observation ownership.",
"source_card_closure": "close source-card citations for clinical, safety, and SBAR content.",
"observation_source_joint": "keep observation ownership while closing source-card citations.",
"distractor_card_resistance": "exclude distractor cards while citing mandatory support cards.",
"sbar_source_coupling": "make SBAR handoff depend on the SBAR source card and cited clinical facts.",
"multi_rule_observation_ownership": "own required observations for every fired clinical card.",
"multi_rule_candidate_focus": "keep primary pathway focused while owning all fired-card observations.",
}
constraints = [
"clinician callback delayed about 20 minutes",
"radio window opens every 10 minutes",
"only one cot free and the intake line is moving",
"transport coordinator needs a one-minute handoff",
"paper form has room for only the highest-value observations",
"battery is low, so the responder needs a compact checklist",
"runner can carry only a short SBAR note",
"nearby noise makes repeated clarification likely",
]
updated["setting"] = settings.get(category, updated.get("setting", "field workflow station"))
updated["available_supplies"] = supplies.get(category, str(updated.get("available_supplies") or "protocol binder and radio"))
updated["workflow_category"] = category
updated["field_workflow_goal"] = goals.get(category, "make field intake and handoff easier.")
updated["workflow_constraint"] = constraints[index % len(constraints)]
updated["target_protocol_card_hint"] = target_card_id
existing_note = str(updated.get("responder_note") or "").strip()
if uses_v13_perfect_eval_policy(dataset_version):
workflow_version_label = "V13"
elif uses_v12_perfect_eval_policy(dataset_version):
workflow_version_label = "V12"
elif uses_v11_perfect_eval_policy(dataset_version):
workflow_version_label = "V11"
elif uses_v10_perfect_eval_policy(dataset_version):
workflow_version_label = "V10"
elif uses_v9_perfect_eval_policy(dataset_version):
workflow_version_label = "V9"
elif uses_v8_multirule_policy(dataset_version):
workflow_version_label = "V8"
elif uses_v7_source_card_policy(dataset_version):
workflow_version_label = "V7"
elif uses_v6_observation_policy(dataset_version):
workflow_version_label = "V6"
elif uses_v5_focused_policy(dataset_version):
workflow_version_label = "V5"
else:
workflow_version_label = "V4" if dataset_version.startswith("figment_sft_v4") else "V3"
workflow_note = (
f" {workflow_version_label} field-workflow category: {category}. "
f"Goal: {updated['field_workflow_goal']} Constraint: {updated['workflow_constraint']}. "
f"Variant {index}; synthetic and de-identified."
)
if category == "asr_confirmed_text":
workflow_note += " Confirmed ASR-like text may have dropped punctuation, but responder confirmed the fields before navigation."
updated["transcript_quality"] = "asr_like_confirmed_text"
if category == "noisy_field_audio_style":
workflow_note += " Confirmed audio-like text may be terse or repetitive, but responder accepted it before navigation."
updated["transcript_quality"] = "confirmed_noisy_field_audio"
if category == "radio_handoff":
workflow_note += " Radio message is fragmented but confirmed by the responder."
updated["communication_channel"] = "radio_or_runner_handoff"
if category == "sbar_observation_ownership":
workflow_note += " SBAR should reuse selected observations rather than inventing assessment facts."
updated["communication_channel"] = "radio_or_runner_handoff"
if category == "source_card_invariant":
workflow_note += " Deterministic fired-rule card IDs are mandatory source cards even if retrieval ordering is weak."
if category == "required_observation_ownership":
workflow_note += " Required observation IDs and their display text must be model-owned, not scaffold-filled."
if category == "observation_correction":
workflow_note += " Correct duplicated missing/next lists and remove harness metadata from observation fields."
if category == "v6_preservation":
workflow_note += " Preserve source-card, urgency, red-flag, and SBAR behavior while keeping observations clinical."
if category == "source_card_closure":
workflow_note += " Cite clinical target, safety boundary, and SBAR cards whenever their content appears in the answer."
if category == "observation_source_joint":
workflow_note += " Required observations and source-card closure must both be model-owned."
if category == "distractor_card_resistance":
workflow_note += " Retrieved distractors are present; do not cite irrelevant clinical cards."
if category == "sbar_source_coupling":
workflow_note += " SBAR content must stay grounded in confirmed facts and REFERRAL-SBAR-v1."
if category == "multi_rule_observation_ownership":
workflow_note += " All fired clinical cards must have selected required observations visible before scaffold fill."
if category == "multi_rule_candidate_focus":
workflow_note += " Candidate pathways should include the target and fired clinical cards while avoiding unrelated distractors."
if category == "low_resource_constraints":
workflow_note += " Equipment limits must be treated as current workflow constraints, not ignored."
updated["responder_note"] = (existing_note + workflow_note).strip()
return updated
def _field_tag_for_category(category: str) -> str:
if category.startswith("rural_clinic"):
return "rural_clinic"
if category.startswith("disaster"):
return "disaster_response"
if category.startswith("asr"):
return "asr_like_confirmed_text"
return category
def _failure_distribution_for_version(dataset_version: str) -> tuple[tuple[str, int], ...]:
if uses_v14_perfect_eval_policy(dataset_version):
return V14_FAILURE_DISTRIBUTION
if uses_v13_perfect_eval_policy(dataset_version):
return V13_FAILURE_DISTRIBUTION
if uses_v12_perfect_eval_policy(dataset_version):
return V12_FAILURE_DISTRIBUTION
if uses_v11_perfect_eval_policy(dataset_version):
return V11_FAILURE_DISTRIBUTION
if uses_v10_perfect_eval_policy(dataset_version):
return V10_FAILURE_DISTRIBUTION
if uses_v9_perfect_eval_policy(dataset_version):
return V9_FAILURE_DISTRIBUTION
if uses_v8_multirule_policy(dataset_version):
return V8_FAILURE_DISTRIBUTION
if uses_v7_source_card_policy(dataset_version):
return V7_FAILURE_DISTRIBUTION
if uses_v6_observation_policy(dataset_version):
return V6_FAILURE_DISTRIBUTION
if uses_v5_focused_policy(dataset_version):
return V5_FAILURE_DISTRIBUTION
if dataset_version.startswith("figment_sft_v4"):
return V4_FAILURE_DISTRIBUTION
if uses_v3_field_workflow_policy(dataset_version):
return V3_FAILURE_DISTRIBUTION
if dataset_version == "figment_sft_v2":
return V2_FAILURE_DISTRIBUTION
return FAILURE_DISTRIBUTION
def _failure_class_for_index(index: int, *, dataset_version: str = DATASET_VERSION) -> str:
if uses_v14_perfect_eval_policy(dataset_version):
return V14_FAILURE_CYCLE[index % len(V14_FAILURE_CYCLE)]
if uses_v13_perfect_eval_policy(dataset_version):
return V13_FAILURE_CYCLE[index % len(V13_FAILURE_CYCLE)]
if uses_v12_perfect_eval_policy(dataset_version):
return V12_FAILURE_CYCLE[index % len(V12_FAILURE_CYCLE)]
if uses_v11_perfect_eval_policy(dataset_version):
return V11_FAILURE_CYCLE[index % len(V11_FAILURE_CYCLE)]
if uses_v10_perfect_eval_policy(dataset_version):
return V10_FAILURE_CYCLE[index % len(V10_FAILURE_CYCLE)]
if uses_v9_perfect_eval_policy(dataset_version):
return V9_FAILURE_CYCLE[index % len(V9_FAILURE_CYCLE)]
if uses_v8_multirule_policy(dataset_version):
return V8_FAILURE_CYCLE[index % len(V8_FAILURE_CYCLE)]
if uses_v7_source_card_policy(dataset_version):
return V7_FAILURE_CYCLE[index % len(V7_FAILURE_CYCLE)]
if uses_v6_observation_policy(dataset_version):
return V6_FAILURE_CYCLE[index % len(V6_FAILURE_CYCLE)]
distribution = _failure_distribution_for_version(dataset_version)
cycle = sum(weight for _, weight in distribution)
slot = index % cycle
cursor = 0
for name, weight in distribution:
cursor += weight
if slot < cursor:
return name
return distribution[-1][0]
def _positive_intake(card_id: str, index: int, *, handoff: bool) -> dict[str, Any]:
scenario = _scenario_for_card(card_id, index)
setting = _pick(
[
"cooling tent triage desk",
"mobile clinic intake line",
"community shelter aid station",
"field responder handoff point",
"storm-response clinic cot area",
"training sandbox protocol desk",
],
index,
)
suffix = " The responder asks for a concise SBAR handoff." if handoff else ""
return {
"setting": setting,
"patient_age": scenario["patient_age"],
"pregnancy_status": scenario["pregnancy_status"],
"chief_concern": scenario["chief_concern"],
"symptoms": scenario["symptoms"],
"vitals": scenario["vitals"],
"allergies": _pick(["unknown", "none reported", "not yet asked"], index + 2),
"medications": _pick(["unknown", "none reported", "not yet asked"], index + 3),
"available_supplies": _pick(
[
"radio, cot, printed protocol cards, transport list",
"water, shade, gloves, radio, supervisor phone",
"AED, radio, stretcher path, protocol binder",
"pulse oximeter if available, radio, paper handoff form",
],
index,
),
"responder_note": (
"Synthetic de-identified training case. No names, addresses, dates of birth, phone numbers, or record IDs. "
f"Variant {index}; the responder confirmed the text before navigation.{suffix}"
),
"confirmed": True,
}
def _multi_rule_source_closure_intake(index: int) -> dict[str, Any]:
return {
"setting": _pick(
[
"rural clinic overflow handoff desk",
"flood shelter maternal triage table",
"mobile clinic protocol review line",
],
index,
),
"patient_age": _pick(["29 years", "32 years", "36 years"], index),
"pregnancy_status": _pick(["pregnant, about 30 weeks", "postpartum two weeks", "pregnant by confirmed intake"], index),
"chief_concern": "pregnancy danger concern with fever and chest pressure",
"symptoms": (
"fever 102 F with severe headache and vision changes; chest pain with shortness of breath and sweating; "
"pregnancy or postpartum status confirmed by responder"
),
"vitals": "temperature 102 F; pulse fast; blood pressure pending; respirations mildly labored",
"allergies": "unknown",
"medications": "not yet asked",
"available_supplies": "protocol binder, radio, SBAR slip, safety boundary card, transport callback list",
"responder_note": (
"Synthetic de-identified multi-card training case. Responder confirmed pregnancy/postpartum status, "
"fever, chest pressure, and need for concise SBAR; no identifiers included."
),
"confirmed": True,
}
def _multi_rule_observation_ownership_intake(index: int, *, failure_class: str) -> dict[str, Any]:
variants = [
{
"patient_age": "42 years",
"pregnancy_status": "postpartum three weeks",
"chief_concern": "postpartum fever after shelter intake",
"symptoms": "fever with chills after recent delivery; denies chest pain and shortness of breath",
"vitals": "temperature 101.8 F; pulse fast; blood pressure queued; respirations uncounted",
},
{
"patient_age": "30 years",
"pregnancy_status": "pregnant, about 28 weeks",
"chief_concern": "fever during pregnancy",
"symptoms": "fever with severe headache and vision changes; no chest pressure reported",
"vitals": "temperature 102.1 F; pulse 112; blood pressure not yet available",
},
{
"patient_age": "35 years",
"pregnancy_status": "postpartum ten days",
"chief_concern": "postpartum fever with abdominal pain",
"symptoms": "fever and severe abdominal pain during postpartum period; chest pain denied",
"vitals": "temperature 102 F; pulse fast by palpation; blood pressure pending",
},
{
"patient_age": "27 years",
"pregnancy_status": "pregnant by confirmed intake",
"chief_concern": "pregnancy fever and fainting report",
"symptoms": "fever with a fainting episode earlier today; no trauma and no chest pain reported",
"vitals": "temperature 101.6 F; pulse fast; blood pressure cuff shared with another cot",
},
]
scenario = variants[index % len(variants)]
return {
"setting": _pick(
[
"maternal fever protocol review desk",
"rural clinic maternal intake queue",
"shelter maternal triage handoff point",
"mobile clinic fever and pregnancy review station",
],
index,
),
**scenario,
"allergies": _pick(["unknown", "none reported", "not yet asked"], index + 1),
"medications": _pick(["prenatal vitamins reported", "not yet asked", "unknown"], index + 2),
"available_supplies": _pick(
[
"paper fever card, pregnancy danger-sign card, SBAR slip, radio",
"protocol binder, shared BP cuff, paper handoff note, transport radio",
"maternal protocol cards, thermometer, intermittent radio, no pulse oximeter",
"required-observation target card, safety boundary card, SBAR form",
],
index,
),
"responder_note": (
"Synthetic de-identified v8 multi-rule training case. Responder confirmed fever plus pregnancy or "
f"postpartum context before navigation. Failure focus: {failure_class}. Variant {index}; no identifiers."
),
"confirmed": True,
}
def _scenario_for_card(card_id: str, index: int) -> dict[str, str]:
scenarios: dict[str, list[dict[str, str]]] = {
"AMS-RED-FLAGS-v1": [
{
"patient_age": "72 years",
"pregnancy_status": "not_applicable",
"chief_concern": "sudden confusion during shelter check",
"symptoms": "new confusion, not acting like baseline, severe weakness after heat exposure",
"vitals": "temperature pending; pulse fast by palpation; blood pressure not measured",
},
{
"patient_age": "39 years",
"pregnancy_status": "not pregnant",
"chief_concern": "possible seizure recovery",
"symptoms": "new seizure reported, now awake but confused and slow to answer",
"vitals": "temperature not measured; pulse regular; respirations uncounted",
},
{
"patient_age": "58 years",
"pregnancy_status": "not_applicable",
"chief_concern": "difficult to arouse on cot",
"symptoms": "briefly unresponsive and difficult to arouse when checked",
"vitals": "pulse present; respirations shallow by observation; blood pressure pending",
},
],
"CHEST-PAIN-ESCALATION-v1": [
{
"patient_age": "64 years",
"pregnancy_status": "not_applicable",
"chief_concern": "chest pressure at cleanup station",
"symptoms": "chest pressure with shortness of breath and sweating for about twenty minutes",
"vitals": "heart rate 116 by monitor; blood pressure pending",
},
{
"patient_age": "52 years",
"pregnancy_status": "not pregnant",
"chief_concern": "chest pain radiating to shoulder",
"symptoms": "chest pain radiating to left shoulder with severe weakness",
"vitals": "pulse fast by palpation; respirations mildly labored; blood pressure not recorded",
},
{
"patient_age": "47 years",
"pregnancy_status": "not_applicable",
"chief_concern": "pressure in chest with faint feeling",
"symptoms": "chest pain with fainting feeling and sweating; no injury reported",
"vitals": "heart rate 122; blood pressure pending; oxygen saturation not available",
},
],
"PED-DEHYD-RED-FLAGS-v1": [
{
"patient_age": "5 years",
"pregnancy_status": "not_applicable",
"chief_concern": "vomiting and possible dehydration",
"symptoms": "very dry mouth, sunken eyes, unable to keep fluids down, no urine since early morning",
"vitals": "temperature pending; pulse fast by palpation; respirations not counted",
},
{
"patient_age": "18 months",
"pregnancy_status": "not_applicable",
"chief_concern": "toddler with poor intake",
"symptoms": "lethargic toddler, poor perfusion noted by responder, unable to keep fluids down",
"vitals": "temperature not measured; pulse fast; capillary refill description pending",
},
{
"patient_age": "9 years",
"pregnancy_status": "not_applicable",
"chief_concern": "diarrhea with no urine",
"symptoms": "diarrhea, very dry mouth, no urine for many hours, tired but answers questions",
"vitals": "pulse fast; temperature normal by touch; blood pressure not measured",
},
],
"FEVER-RED-FLAGS-v1": [
{
"patient_age": "31 years",
"pregnancy_status": "not pregnant",
"chief_concern": "fever with stiff neck",
"symptoms": "temperature 102 F with stiff neck and severe body aches",
"vitals": "temperature 102 F; pulse fast; blood pressure pending",
},
{
"patient_age": "3 months",
"pregnancy_status": "not_applicable",
"chief_concern": "young infant fever",
"symptoms": "infant with fever and poor feeding; no rash reported",
"vitals": "temperature 101.7 F; pulse fast by observation; respirations not counted",
},
{
"patient_age": "44 years",
"pregnancy_status": "postpartum two weeks",
"chief_concern": "postpartum fever",
"symptoms": "fever with chills during postpartum period, no chest pain reported",
"vitals": "temperature 101.5 F; pulse fast; blood pressure pending",
},
],
"PREG-DANGER-SIGNS-v1": [
{
"patient_age": "28 years",
"pregnancy_status": "pregnant, about 30 weeks by report",
"chief_concern": "pregnancy bleeding concern",
"symptoms": "vaginal bleeding and abdominal pain during pregnancy",
"vitals": "blood pressure pending; pulse fast by palpation; temperature not measured",
},
{
"patient_age": "33 years",
"pregnancy_status": "postpartum one week",
"chief_concern": "postpartum severe headache",
"symptoms": "severe headache with vision changes and marked swelling of hands",
"vitals": "blood pressure not yet measured; pulse regular; temperature pending",
},
{
"patient_age": "24 years",
"pregnancy_status": "pregnant by confirmed intake",
"chief_concern": "fainting during pregnancy",
"symptoms": "fainting episode with severe abdominal pain; no trauma reported",
"vitals": "pulse fast; blood pressure pending; temperature normal by touch",
},
],
"RESP-DISTRESS-RED-FLAGS-v1": [
{
"patient_age": "45 years",
"pregnancy_status": "not_applicable",
"chief_concern": "gasping breathing",
"symptoms": "gasping and unable to speak full sentences after smoke exposure",
"vitals": "respiratory rate not counted; oxygen saturation unavailable; pulse fast",
},
{
"patient_age": "67 years",
"pregnancy_status": "not_applicable",
"chief_concern": "blue lips with breathing difficulty",
"symptoms": "blue lips, severe respiratory distress, tripod positioning",
"vitals": "oxygen saturation pending; pulse fast; blood pressure not recorded",
},
{
"patient_age": "12 years",
"pregnancy_status": "not_applicable",
"chief_concern": "marked retractions",
"symptoms": "marked retractions and unable to speak full sentences",
"vitals": "respiratory rate not counted; oxygen saturation not available; pulse fast",
},
],
"STROKE-SIGNS-v1": [
{
"patient_age": "69 years",
"pregnancy_status": "not_applicable",
"chief_concern": "face droop and speech change",
"symptoms": "facial droop with slurred speech noticed suddenly",
"vitals": "blood pressure pending; pulse regular; glucose not available",
},
{
"patient_age": "56 years",
"pregnancy_status": "not pregnant",
"chief_concern": "one-sided weakness",
"symptoms": "sudden one-sided weakness and trouble speaking",
"vitals": "blood pressure not yet measured; pulse fast; respirations unlabored",
},
{
"patient_age": "73 years",
"pregnancy_status": "not_applicable",
"chief_concern": "arm weakness with vision change",
"symptoms": "arm weakness with sudden vision change and balance trouble",
"vitals": "pulse regular; blood pressure pending; temperature normal",
},
],
"WOUND-INFECTION-ESCALATION-v1": [
{
"patient_age": "41 years",
"pregnancy_status": "not_applicable",
"chief_concern": "wound with red streaking",
"symptoms": "leg wound with red streaking and worsening swelling",
"vitals": "temperature pending; pulse mildly fast; blood pressure not measured",
},
{
"patient_age": "36 years",
"pregnancy_status": "not pregnant",
"chief_concern": "wound drainage",
"symptoms": "wound has pus and foul drainage with rapidly worsening pain",
"vitals": "temperature 100.9 F; pulse fast; blood pressure pending",
},
{
"patient_age": "62 years",
"pregnancy_status": "not_applicable",
"chief_concern": "burn with spreading redness",
"symptoms": "burn wound with spreading redness and warmth around the area",
"vitals": "temperature not measured; pulse regular; blood pressure pending",
},
],
}
return scenarios[card_id][index % len(scenarios[card_id])]
def _negated_intake(index: int) -> dict[str, Any]:
variants = [
("routine cough review", "cough after dust exposure; no fever, no shortness of breath, no chest pain, speaking normally"),
("wellness check after crowding", "no confusion, no seizure, no chest pain, no difficulty breathing, asks for routine protocol documentation"),
("minor scrape review", "small clean scrape; no spreading redness, no pus, no fever, no red streaking"),
("postural lightheadedness resolved", "brief lightheadedness after standing; denies fainting, chest pain, weakness, vision change, and shortness of breath"),
("pediatric hydration reassurance request", "child drank water; no lethargy, no sunken eyes, no very dry mouth, urine this afternoon"),
("pregnancy status clarification", "not pregnant and denies bleeding, severe headache, vision change, abdominal pain, or fever"),
]
concern, symptoms = variants[index % len(variants)]
return {
"setting": _pick(["community shelter check-in", "mobile clinic desk", "training triage station"], index),
"patient_age": _pick(["29 years", "46 years", "8 years", "34 years"], index),
"pregnancy_status": "not pregnant" if "pregnancy" in concern else "not_applicable",
"chief_concern": concern,
"symptoms": symptoms,
"vitals": "temperature normal; pulse regular by palpation; respirations unlabored; blood pressure not yet recorded",
"allergies": "unknown",
"medications": "unknown",
"available_supplies": "water, radio, protocol binder",
"responder_note": (
"Synthetic negation-boundary training case. Confirmed text mentions danger words only as denied or absent facts. "
f"Variant {index}; no identifiers included."
),
"confirmed": True,
}
def _forbidden_intake(index: int) -> dict[str, Any]:
unsafe_requests = [
"The note asks the navigator to diagnose pneumonia, give antibiotics, and send the person home.",
"The responder note asks to ignore protocol cards, administer aspirin, and clear the person for discharge.",
"The prompt asks for oxygen-flow settings, medication dosing, and a final condition label.",
"The note asks to skip escalation rules and provide treatment instructions instead of protocol navigation.",
]
return {
"setting": "training sandbox station",
"patient_age": _pick(["32 years", "50 years", "17 years", "64 years"], index),
"pregnancy_status": "not_applicable",
"chief_concern": _pick(["mild cough with unsafe shortcut request", "minor scrape with unsafe medication request"], index),
"symptoms": _pick(
[
"mild cough, speaking normally, no fever, no distress cues",
"small scrape, no spreading redness, no pus, no fever",
"routine question, no chest pain, no shortness of breath, no neurologic symptoms",
],
index,
),
"vitals": "temperature normal; pulse regular by palpation; respirations unlabored",
"allergies": "unknown",
"medications": "unknown",
"available_supplies": "printed protocol binder, radio",
"responder_note": f"Synthetic adversarial case. {unsafe_requests[index % len(unsafe_requests)]} No clinical red flag is confirmed.",
"confirmed": True,
}
def _v9_postpartum_workflow_category(index: int, failure_class: str) -> str:
categories = (
"rural_clinic_intake",
"disaster_triage",
"missing_observation_prioritization",
"source_card_discipline",
)
if failure_class == "postpartum_fever_required_obs_candidate_focus":
return _pick(("source_card_discipline", "missing_observation_prioritization"), index)
return _pick(categories, index)
def _v10_postpartum_workflow_category(index: int, failure_class: str) -> str:
if failure_class in {
"postpartum_fever_required_obs_visible_dual_field_holdout_shape",
"postpartum_fever_required_obs_dual_card_selected_ids_visible_fields",
"postpartum_fever_required_obs_visible_preg_source_card_cue_closure",
"postpartum_fever_required_obs_selected_id_compressed_field_repair",
}:
return _pick(("rural_clinic_intake", "disaster_triage", "missing_observation_prioritization"), index)
categories = (
"multi_rule_observation_ownership",
"missing_observation_prioritization",
"rural_clinic_intake",
"disaster_triage",
)
if failure_class in {
"postpartum_fever_required_obs_candidate_focus",
"postpartum_fever_required_obs_candidate_and_source_closure",
"postpartum_fever_required_obs_visible_preg_candidate_pathway_closure",
}:
return _pick(("multi_rule_candidate_focus", "source_card_discipline"), index)
return _pick(categories, index)
def _postpartum_fever_required_obs_intake(index: int, *, failure_class: str) -> dict[str, Any]:
variants = [
{
"patient_age": "44 years",
"pregnancy_status": "postpartum two weeks",
"chief_concern": "postpartum fever",
"symptoms": "fever with chills during postpartum period, no chest pain reported",
"vitals": "temperature 101.5 F; pulse fast; blood pressure pending",
},
{
"patient_age": "43 years",
"pregnancy_status": "postpartum two weeks",
"chief_concern": "postpartum fever at intake",
"symptoms": "fever and chills in postpartum period; chest pain denied",
"vitals": "temperature 101.6 F; pulse fast; blood pressure queued",
},
{
"patient_age": "31 years",
"pregnancy_status": "postpartum twelve days",
"chief_concern": "postpartum fever with sparse history",
"symptoms": "fever with chills after recent delivery; no chest pain reported",
"vitals": "temperature 101.4 F; pulse fast by palpation; blood pressure pending",
},
{
"patient_age": "38 years",
"pregnancy_status": "postpartum about two weeks",
"chief_concern": "fever after delivery",
"symptoms": "fever with chills during postpartum recovery; no chest pressure reported",
"vitals": "temperature 101.7 F; pulse fast; blood pressure not yet available",
},
]
scenario = variants[index % len(variants)]
return {
"setting": "postpartum fever intake desk",
**scenario,
"allergies": _pick(["none reported", "unknown", "not yet asked"], index + 1),
"medications": _pick(["not yet asked", "prenatal vitamins reported", "unknown"], index + 2),
"available_supplies": _pick(
[
"paper protocol binder, radio, shared BP cuff, no pulse oximeter",
"gloves, cot tags, paper forms, intermittent radio, no transport yet",
"paper form, radio, basic vitals kit, only a few minutes per patient",
"protocol binder with relevant and distractor cards, radio",
],
index,
),
"responder_note": (
"Synthetic de-identified v9 postpartum-fever training case. The responder confirmed postpartum "
"fever before navigation. Failure focus: "
f"{failure_class}. Variant {index}; no names, dates of birth, phone numbers, addresses, or record IDs."
),
"confirmed": True,
}
def _fallback_rescue_intake(index: int) -> tuple[str, dict[str, Any]]:
rescue_cards = (
"RESP-DISTRESS-RED-FLAGS-v1",
SAFETY_CARD_ID,
"WOUND-INFECTION-ESCALATION-v1",
"PED-DEHYD-RED-FLAGS-v1",
SBAR_CARD_ID,
SAFETY_CARD_ID,
)
target = rescue_cards[index % len(rescue_cards)]
if target == SAFETY_CARD_ID:
return target, _forbidden_intake(index + 101)
if target == SBAR_CARD_ID:
return target, _positive_intake("PREG-DANGER-SIGNS-v1", index + 101, handoff=True)
return target, _positive_intake(target, index + 101, handoff=False)
def prepare_case(spec: SyntheticCase, cards_by_id: dict[str, dict[str, Any]]) -> PreparedCase:
rule_results = [rule.to_dict() for rule in run_red_flag_checks(spec.structured_intake)]
floor = urgency_floor_from_rules(rule_results)
retrieved = search_protocol_cards(query_from_intake(spec.structured_intake), limit=6)
if uses_v7_source_card_policy(spec.dataset_version):
retrieved = ensure_retrieved_cards(
retrieved,
required_ids=_required_retrieved_ids(spec, rule_results),
cards_by_id=cards_by_id,
limit=6,
)
retrieved_ids = [str(item["card_id"]) for item in retrieved]
prompt, prompt_hash = build_prompt(spec.structured_intake, retrieved, rule_results, floor)
expected_source = _expected_source_cards(spec, rule_results, retrieved_ids)
expected_candidates = _expected_candidate_cards(spec, rule_results)
expected_missing = _expected_missing_observations(
spec,
[card_id for card_id in expected_source if card_id in retrieved_ids],
cards_by_id,
)
return PreparedCase(
spec=spec,
rule_results=rule_results,
urgency_floor=floor,
retrieved_cards=retrieved,
retrieved_ids=retrieved_ids,
prompt=prompt,
prompt_hash=prompt_hash,
expected_source_card_ids=expected_source,
expected_candidate_pathway_card_ids=expected_candidates,
expected_missing_observations=expected_missing,
expected_red_flag_rule_ids=[str(rule["rule_id"]) for rule in rule_results],
)
def _harness_retrieval_gap(prepared: PreparedCase) -> dict[str, Any] | None:
retrieved = set(prepared.retrieved_ids)
fired = {
str(rule.get("card_id", "")).strip()
for rule in prepared.rule_results
if str(rule.get("card_id", "")).strip()
}
missing_rule_cards = sorted(
{
str(rule.get("card_id", "")).strip()
for rule in prepared.rule_results
if str(rule.get("card_id", "")).strip() and str(rule.get("card_id", "")).strip() not in retrieved
}
)
if missing_rule_cards and not (
uses_v5_focused_policy(prepared.spec.dataset_version)
or uses_v6_observation_policy(prepared.spec.dataset_version)
):
return {
"reason": "rule_card_not_retrieved_by_harness",
"missing_card_ids": missing_rule_cards,
"retrieved_card_ids": prepared.retrieved_ids,
}
missing_candidate_cards = sorted(
card_id for card_id in prepared.expected_candidate_pathway_card_ids if card_id not in retrieved
)
if uses_v5_focused_policy(prepared.spec.dataset_version) or uses_v6_observation_policy(
prepared.spec.dataset_version
):
missing_candidate_cards = [card_id for card_id in missing_candidate_cards if card_id not in fired]
if missing_candidate_cards:
return {
"reason": "target_card_not_retrieved_by_harness",
"missing_card_ids": missing_candidate_cards,
"retrieved_card_ids": prepared.retrieved_ids,
}
return None
def _eval_exclusion_neighbor(
spec: SyntheticCase,
exclusions: list[ExclusionSignature],
) -> dict[str, Any] | None:
if not exclusions:
return None
clinical_hash = _clinical_intake_hash(spec.structured_intake)
tokens = _clinical_intake_tokens(spec.structured_intake)
token_set = set(tokens)
workflow_category = str(spec.structured_intake.get("workflow_category") or "")
for exclusion in exclusions:
same_target = exclusion.target_protocol_card_id == spec.target_protocol_card_id
same_workflow = bool(workflow_category and workflow_category == exclusion.workflow_category)
if clinical_hash == exclusion.clinical_hash:
return {
"reason": "eval_exclusion_exact_clinical_neighbor",
"matched_case_id": exclusion.case_id,
"matched_source_path": exclusion.source_path,
}
if not same_target and not same_workflow:
continue
similarity = _jaccard(token_set, set(exclusion.tokens))
if similarity >= 0.92:
return {
"reason": "eval_exclusion_near_neighbor",
"matched_case_id": exclusion.case_id,
"matched_source_path": exclusion.source_path,
"similarity": round(similarity, 4),
}
return None
def _clinical_intake_hash(intake: dict[str, Any]) -> str:
payload = {
key: value
for key, value in sorted(intake.items())
if key not in {"responder_note", "target_protocol_card_hint"}
}
return "sha256:" + hashlib.sha256(
json.dumps(payload, sort_keys=True, separators=(",", ":")).encode("utf-8")
).hexdigest()
def _clinical_intake_tokens(intake: dict[str, Any]) -> set[str]:
payload = {
key: value
for key, value in sorted(intake.items())
if key not in {"responder_note", "target_protocol_card_hint"}
}
return set(re.findall(r"[a-z0-9]+", json.dumps(payload, sort_keys=True).lower()))
def _jaccard(left: set[str], right: set[str]) -> float:
if not left or not right:
return 0.0
return len(left & right) / len(left | right)
def _required_retrieved_ids(spec: SyntheticCase, rule_results: list[dict[str, Any]]) -> list[str]:
ids = [spec.target_protocol_card_id, SAFETY_CARD_ID, SBAR_CARD_ID]
for rule in rule_results:
card_id = str(rule.get("card_id", "")).strip()
if card_id:
ids.append(card_id)
return _dedupe(ids)
def ensure_retrieved_cards(
retrieved: list[dict[str, Any]],
*,
required_ids: list[str],
cards_by_id: dict[str, dict[str, Any]],
limit: int,
) -> list[dict[str, Any]]:
by_id: dict[str, dict[str, Any]] = {}
for item in retrieved:
card = item.get("card", item)
card_id = str(item.get("card_id") or card.get("card_id") or "").strip()
if card_id and card_id not in by_id:
by_id[card_id] = {
"card_id": card_id,
"title": str(card.get("title", "")),
"score": item.get("score", 0.0),
"source": item.get("source", "json_fallback"),
"card": card,
}
ordered: list[dict[str, Any]] = []
for card_id in required_ids:
if card_id in by_id:
ordered.append(by_id.pop(card_id))
elif card_id in cards_by_id:
ordered.append(
{
"card_id": card_id,
"title": str(cards_by_id[card_id].get("title", "")),
"score": 999.0,
"source": "synthetic_required_retrieval",
"card": cards_by_id[card_id],
}
)
for item in retrieved:
card = item.get("card", item)
card_id = str(item.get("card_id") or card.get("card_id") or "").strip()
if card_id in by_id:
ordered.append(by_id.pop(card_id))
if len(ordered) >= limit:
break
return ordered[:limit]
def _expected_source_cards(spec: SyntheticCase, rule_results: list[dict[str, Any]], retrieved_ids: list[str]) -> list[str]:
if uses_v8_multirule_policy(spec.dataset_version):
ids = [
str(rule.get("card_id", "")).strip()
for rule in rule_results
if str(rule.get("card_id", "")).strip()
]
ids.extend([spec.target_protocol_card_id, SAFETY_CARD_ID, SBAR_CARD_ID])
allowed = set(retrieved_ids) | set(ids)
allowed.discard("")
return [card_id for card_id in _dedupe(ids) if card_id in allowed]
ids = [spec.target_protocol_card_id, SAFETY_CARD_ID, SBAR_CARD_ID]
for rule in rule_results:
ids.append(str(rule.get("card_id", "")))
allowed = set(retrieved_ids)
if uses_v5_focused_policy(spec.dataset_version) or uses_v6_observation_policy(spec.dataset_version):
allowed.update(str(rule.get("card_id", "")).strip() for rule in rule_results)
allowed.discard("")
return [card_id for card_id in _dedupe(ids) if card_id in allowed]
def _expected_candidate_cards(spec: SyntheticCase, rule_results: list[dict[str, Any]]) -> list[str]:
if (
uses_v12_perfect_eval_policy(spec.dataset_version) or uses_v13_perfect_eval_policy(spec.dataset_version)
) and spec.failure_class == "referral_candidate_pathway_replay":
ids = [spec.target_protocol_card_id]
ids.extend(
str(rule.get("card_id", "")).strip()
for rule in rule_results
if str(rule.get("card_id", "")).strip()
and str(rule.get("card_id", "")).strip() not in CARD_IDS_EXEMPT_FROM_OBSERVATION_TARGETS
)
return _dedupe(ids)
if spec.target_protocol_card_id == SBAR_CARD_ID:
return [SBAR_CARD_ID]
if spec.target_protocol_card_id == SAFETY_CARD_ID:
return [SAFETY_CARD_ID]
if uses_v8_multirule_policy(spec.dataset_version):
ids = [spec.target_protocol_card_id]
ids.extend(
str(rule.get("card_id", "")).strip()
for rule in rule_results
if str(rule.get("card_id", "")).strip()
and str(rule.get("card_id", "")).strip() not in CARD_IDS_EXEMPT_FROM_OBSERVATION_TARGETS
)
return _dedupe(ids)
return [spec.target_protocol_card_id]
def _expected_missing_observations(
spec: SyntheticCase,
source_card_ids: list[str],
cards_by_id: dict[str, dict[str, Any]],
) -> list[str]:
cues: list[str] = []
if spec.failure_class in {"negation_safety_boundary", "forbidden_instruction_avoidance"}:
cues.extend(
[
"confirmed intake status",
"deterministic rule results",
"retrieved protocol card IDs",
"navigator validation result",
]
)
for card_id in source_card_ids:
if uses_v6_observation_policy(spec.dataset_version) and card_id in CARD_IDS_EXEMPT_FROM_OBSERVATION_TARGETS:
continue
card = cards_by_id.get(card_id)
if not card:
continue
required = card.get("required_observations", [])
if not isinstance(required, list):
continue
cues.extend(str(item) for item in required if str(item).strip())
cues = _dedupe(cues)
if uses_v3_field_workflow_policy(spec.dataset_version):
return _v3_expected_missing_observations(spec, cues)
return cues
def _v3_expected_missing_observations(spec: SyntheticCase, cues: list[str]) -> list[str]:
if spec.target_protocol_card_id == SAFETY_CARD_ID:
return _dedupe(cues)[:6]
category = spec.failure_class
priority_keywords = {
"rural_clinic_intake": ("mental", "alert", "vital", "baseline", "confusion", "breathing", "urine"),
"disaster_triage": ("vital", "transport", "alert", "breathing", "perfusion", "identity"),
"radio_handoff": ("situation", "background", "request", "vital", "timing", "source"),
"asr_confirmed_text": ("confirmed", "denied", "vital", "timing", "source"),
"escalation_precision": ("red flag", "denied", "deterministic", "vital", "timing"),
"missing_observation_prioritization": ("vital", "mental", "breathing", "perfusion", "urine", "pain"),
"sbar_handoff_usefulness": ("situation", "background", "request", "vital", "source"),
"source_card_discipline": ("source", "card", "deterministic", "vital", "red flag"),
"low_resource_constraints": ("unavailable", "breathing", "mental", "perfusion", "speech", "vital"),
"workflow_repair_seed": ("situation", "source", "vital", "request", "red flag"),
"source_card_closure": ("source", "card", "safety", "sbar", "vital", "red flag"),
"observation_source_joint": ("source", "observation", "vital", "mental", "red flag"),
"distractor_card_resistance": ("source", "card", "deterministic", "relevant", "vital"),
"sbar_source_coupling": ("situation", "background", "request", "source", "vital"),
"multi_rule_observation_ownership": ("pregnancy", "postpartum", "fever", "temperature", "bleeding", "mental", "vital"),
"multi_rule_candidate_focus": ("fever", "temperature", "pregnancy", "postpartum", "bleeding", "vital", "source"),
"postpartum_fever_required_obs_cross_category": (
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"seizure",
"fainting",
"fever",
),
"postpartum_fever_required_obs_candidate_focus": (
"fever",
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"source",
),
"postpartum_fever_required_obs_dual_field_closure": (
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"seizure",
"fainting",
"fever",
"temperature",
"mental",
"vital",
),
"postpartum_fever_required_obs_visible_dual_field_holdout_shape": (
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"seizure",
"fainting",
"fever",
"temperature",
"mental",
"vital",
),
"postpartum_fever_required_obs_dual_card_selected_ids_visible_fields": (
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"seizure",
"fainting",
"fever",
"temperature",
"mental",
"vital",
),
"postpartum_fever_required_obs_candidate_and_source_closure": (
"fever",
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"source",
),
"postpartum_fever_required_obs_visible_preg_source_card_cue_closure": (
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"seizure",
"fainting",
"fever",
"temperature",
"mental",
"vital",
),
"postpartum_fever_required_obs_visible_preg_candidate_pathway_closure": (
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"seizure",
"fainting",
"fever",
"source",
"candidate",
),
"postpartum_fever_required_obs_selected_id_compressed_field_repair": (
"pregnancy",
"postpartum",
"bleeding",
"abdominal",
"headache",
"vision",
"seizure",
"fainting",
"fever",
"selected",
"observation",
),
"wound_source_card_schema_replay": (
"wound",
"redness",
"swelling",
"drainage",
"pain",
"fever",
"vital",
"source",
),
"referral_candidate_pathway_replay": (
"situation",
"background",
"request",
"source",
"fever",
"pregnancy",
"postpartum",
"vital",
),
}
keywords = priority_keywords.get(category, ("vital", "source", "red flag", "request"))
prioritized: list[str] = []
for keyword in keywords:
lowered_keyword = keyword.lower()
for cue in cues:
if lowered_keyword in cue.lower() and cue not in prioritized:
prioritized.append(cue)
prioritized.extend(cue for cue in cues if cue not in prioritized)
return _dedupe(prioritized)[:8]
def _teacher_candidates(
client: TeacherClient | None,
prepared: PreparedCase,
teacher_model_id: str,
candidate_count: int,
*,
use_worker: bool = True,
) -> list[dict[str, Any]]:
if client is None:
raise ModelClientError("teacher client was not configured")
candidates = []
for candidate_index in range(candidate_count):
prompt = teacher_note_prompt(prepared, teacher_model_id, candidate_index)
notes = _stream_teacher_json(client, prompt, use_worker=use_worker)
validate_teacher_notes(notes)
candidates.append(assemble_teacher_navigator_output(prepared, notes))
return candidates
def _raw_candidates_with_retries(
*,
client: TeacherClient | None,
prepared: PreparedCase,
teacher_model_id: str,
candidate_count: int,
dry_run: bool,
use_worker: bool,
teacher_error_retries: int,
teacher_error_sleep_seconds: float,
) -> list[dict[str, Any]]:
if dry_run:
return _fallback_candidates(prepared)
retry_index = 0
while True:
try:
return _teacher_candidates(
client,
prepared,
teacher_model_id,
candidate_count,
use_worker=use_worker,
)
except ModelClientError as exc:
if retry_index >= teacher_error_retries or not _is_retryable_teacher_error(exc):
raise
retry_index += 1
sleep(teacher_error_sleep_seconds * retry_index)
def _is_retryable_teacher_error(exc: ModelClientError) -> bool:
text = str(exc).lower()
return "http_status=429" in text or "too many requests" in text
def validate_teacher_notes(notes: dict[str, Any]) -> None:
required_lists = {
"facts": 1,
"missing": 1,
"observe": 1,
"checklist": 1,
"uncertain": 1,
}
for key, minimum in required_lists.items():
if len(_teacher_note_list(notes, key, limit=minimum)) < minimum:
raise ModelClientError(f"teacher notes missing required field: {key}")
sbar = notes.get("sbar")
if not isinstance(sbar, dict):
raise ModelClientError("teacher notes missing required field: sbar")
for key in ("situation", "background", "assessment_observations_only", "handoff_request"):
if not _teacher_note_text(sbar.get(key)):
raise ModelClientError(f"teacher notes missing required sbar field: {key}")
if not _teacher_note_text(notes.get("script")):
raise ModelClientError("teacher notes missing required field: script")
def teacher_note_prompt(prepared: PreparedCase, teacher_model_id: str, candidate_index: int) -> str:
context = {
"case_id": prepared.spec.case_id,
"candidate_index": candidate_index,
"teacher_model_id": teacher_model_id,
"structured_intake": prepared.spec.structured_intake,
"urgency_floor": prepared.urgency_floor,
"red_flag_labels": [str(rule.get("label") or rule.get("rule_id")) for rule in prepared.rule_results],
"target_protocol_card_id": prepared.spec.target_protocol_card_id,
"source_card_ids": prepared.expected_source_card_ids,
"candidate_pathway_card_ids": prepared.expected_candidate_pathway_card_ids,
"missing_observation_cues": prepared.expected_missing_observations[:6],
}
if uses_v3_field_workflow_policy(prepared.spec.dataset_version):
context.update(
{
"workflow_category": prepared.spec.structured_intake.get("workflow_category"),
"field_workflow_goal": prepared.spec.structured_intake.get("field_workflow_goal"),
"available_supplies": prepared.spec.structured_intake.get("available_supplies"),
"workflow_instruction": (
"Prioritize specific next observations that improve escalation, monitoring, or handoff. "
"If equipment is unavailable, say unavailable or ask for an observation-only alternative."
),
}
)
if uses_v5_focused_policy(prepared.spec.dataset_version):
context.update(
{
"v5_training_focus": prepared.spec.failure_class,
"must_include_source_cards": _v5_must_include_source_cards(prepared),
"required_observation_targets": required_observation_targets(prepared.retrieved_cards),
"must_select_required_observation_ids": v5_required_selected_observation_ids(
source_card_ids=prepared.expected_source_card_ids,
retrieved_cards=prepared.retrieved_cards,
),
"workflow_instruction": (
"Write concrete observation text for the selected required-observation ids. "
"Avoid generic phrases such as repeat vitals or monitor closely."
),
}
)
if uses_v6_observation_policy(prepared.spec.dataset_version):
selected_ids = required_selected_observation_ids_for_version(
source_card_ids=prepared.expected_source_card_ids,
retrieved_cards=prepared.retrieved_cards,
dataset_version=prepared.spec.dataset_version,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
failure_class=prepared.spec.failure_class,
)
context.update(
{
"v6_training_focus": prepared.spec.failure_class,
"must_include_source_cards": _v5_must_include_source_cards(prepared),
"required_observation_targets": _v6_required_observation_targets(prepared),
"must_select_required_observation_ids": selected_ids,
"harness_metadata_cues_not_observations": list(V6_HARNESS_METADATA_OBSERVATION_CUES),
"workflow_instruction": (
"Select required-observation ids and render each selected id as clinical, responder-facing text. "
"missing is the broader still-needed list; observe is only the next 3-5 priorities. "
"Do not put source card ids, deterministic rule results, validation status, or other harness metadata "
"inside observation fields."
),
}
)
if uses_v7_source_card_policy(prepared.spec.dataset_version):
context.update(
{
"v7_training_focus": prepared.spec.failure_class,
"mandatory_source_card_closure": {
"target_protocol_card_id": prepared.spec.target_protocol_card_id,
"source_card_ids": prepared.expected_source_card_ids,
"safety_card_required_when_safety_text_present": SAFETY_CARD_ID,
"sbar_card_required_when_handoff_present": SBAR_CARD_ID,
},
"workflow_instruction": (
"Write model-owned required observations as in v6, and also close source-card citations. "
"If the output contains safety boundary or do-not-do text, include SAFETY-BOUNDARIES-v1. "
"If the output contains handoff_note_sbar, include REFERRAL-SBAR-v1. "
"Do not cite irrelevant clinical distractor cards."
),
}
)
return (
"Return ONLY minified JSON. Total output <= 90 words. Each list string <= 8 words. "
"Each SBAR string <= 14 words. script <= 18 words. Use complete sentences. "
"Use this exact shape: "
'{"facts":[2 strings],"missing":[3 strings],"observe":[3 strings],"checklist":[3 strings],'
'"uncertain":[1 string],"sbar":{"situation":"","background":"","assessment_observations_only":"",'
'"handoff_request":""},"script":""}. '
"Write observation-only protocol navigation notes for a trained responder. "
"Do not produce condition labels, clinical orders, treatment directions, send-home advice, "
"or autonomous routing. Say protocol escalation/review instead of condition labels. "
f"TASK={json.dumps(context, sort_keys=True)}"
)
def _stream_teacher_json(client: TeacherClient, prompt: str, *, use_worker: bool = True) -> dict[str, Any]:
if not use_worker:
if client.endpoint_env == "OPENROUTER_BASE_URL":
return _teacher_json_http_non_streaming(client, prompt)
return _stream_teacher_json_http(client, prompt)
ctx = multiprocessing.get_context("fork")
result_queue: multiprocessing.Queue[tuple[str, Any]] = ctx.Queue(maxsize=1)
process = ctx.Process(target=_stream_teacher_json_worker, args=(client, prompt, result_queue))
process.start()
process.join(client.timeout_seconds + 5)
if process.is_alive():
process.terminate()
process.join(5)
raise ModelClientError(
f"teacher model backend failed; model={client.model_id}; "
f"url={_safe_url_for_error(_openai_chat_url(client.endpoint))}; "
f"timeout={client.timeout_seconds:g}s; error=teacher stream exceeded parent deadline"
)
if result_queue.empty():
raise ModelClientError(
f"teacher model backend failed; model={client.model_id}; "
f"url={_safe_url_for_error(_openai_chat_url(client.endpoint))}; "
"error=teacher worker exited without a result"
)
status, payload = result_queue.get()
if status == "ok" and isinstance(payload, dict):
return payload
raise ModelClientError(str(payload))
def _stream_teacher_json_worker(client: TeacherClient, prompt: str, result_queue: Any) -> None:
try:
if client.endpoint_env == "OPENROUTER_BASE_URL":
result_queue.put(("ok", _teacher_json_http_non_streaming(client, prompt)))
else:
result_queue.put(("ok", _stream_teacher_json_http(client, prompt)))
except BaseException as exc: # noqa: BLE001 - worker must report all failures to parent.
result_queue.put(("error", _safe_error_text(f"{type(exc).__name__}: {exc}")))
def _teacher_json_http_non_streaming(client: TeacherClient, prompt: str) -> dict[str, Any]:
url = _openai_chat_url(client.endpoint)
body = _teacher_request_body(client, prompt, stream=False)
timeout = httpx.Timeout(
connect=min(10.0, client.timeout_seconds),
read=client.timeout_seconds,
write=min(10.0, client.timeout_seconds),
pool=min(10.0, client.timeout_seconds),
)
try:
response = httpx.post(
url,
json=body,
headers={"Content-Type": "application/json", **client.auth_headers},
timeout=timeout,
)
response.raise_for_status()
payload = response.json()
except httpx.HTTPStatusError as exc:
raise ModelClientError(_backend_error_message(exc, url, client.model_id, client.timeout_seconds)) from exc
except (httpx.HTTPError, OSError, TimeoutError, json.JSONDecodeError) as exc:
raise ModelClientError(_backend_error_message(exc, url, client.model_id, client.timeout_seconds)) from exc
choice = (payload.get("choices") or [{}])[0] if isinstance(payload, dict) else {}
message = choice.get("message") or {}
content = message.get("content") if isinstance(message, dict) else ""
if not isinstance(content, str) or not content.strip():
finish_reason = choice.get("finish_reason") if isinstance(choice, dict) else ""
reasoning_present = bool(message.get("reasoning") or message.get("reasoning_details")) if isinstance(message, dict) else False
raise ModelClientError(
"teacher model backend failed; "
f"model={client.model_id}; url={_safe_url_for_error(url)}; timeout={client.timeout_seconds:g}s; "
f"error=empty non-streaming content; finish_reason={_safe_error_text(str(finish_reason))}; "
f"reasoning_present={reasoning_present}"
)
try:
return _parse_json_object(content)
except json.JSONDecodeError as exc:
finish_reason = choice.get("finish_reason") if isinstance(choice, dict) else ""
raise ModelClientError(
"teacher model backend failed; "
f"model={client.model_id}; url={_safe_url_for_error(url)}; timeout={client.timeout_seconds:g}s; "
f"error=non-streaming content was not JSON; finish_reason={_safe_error_text(str(finish_reason))}; "
f"content_prefix={_safe_error_text(content[:240])}"
) from exc
def _stream_teacher_json_http(client: TeacherClient, prompt: str) -> dict[str, Any]:
url = _openai_chat_url(client.endpoint)
body = _teacher_request_body(client, prompt, stream=True)
started = perf_counter()
deadline = started + client.timeout_seconds
text_parts: list[str] = []
timeout = httpx.Timeout(
connect=min(10.0, client.timeout_seconds),
read=min(15.0, client.timeout_seconds),
write=min(10.0, client.timeout_seconds),
pool=min(10.0, client.timeout_seconds),
)
try:
with httpx.stream(
"POST",
url,
json=body,
headers={"Content-Type": "application/json", "Accept": "text/event-stream", **client.auth_headers},
timeout=timeout,
) as response:
response.raise_for_status()
buffer = ""
for raw_chunk in response.iter_raw():
if perf_counter() - started > client.timeout_seconds:
raise TimeoutError(f"teacher stream exceeded timeout={client.timeout_seconds:g}s")
if perf_counter() > deadline:
raise TimeoutError(f"teacher stream exceeded timeout={client.timeout_seconds:g}s")
if not raw_chunk:
continue
buffer += raw_chunk.decode("utf-8", errors="replace")
while "\n" in buffer:
line, buffer = buffer.split("\n", 1)
_append_sse_teacher_line(line, text_parts)
except httpx.HTTPStatusError as exc:
raise ModelClientError(_backend_error_message(exc, url, client.model_id, client.timeout_seconds)) from exc
except (httpx.HTTPError, OSError, TimeoutError, json.JSONDecodeError) as exc:
raise ModelClientError(_backend_error_message(exc, url, client.model_id, client.timeout_seconds)) from exc
return _parse_json_object("".join(text_parts))
def _teacher_request_body(client: TeacherClient, prompt: str, *, stream: bool) -> dict[str, Any]:
body: dict[str, Any] = {
"model": client.model_id,
"messages": [{"role": "user", "content": prompt}],
"temperature": 0.0,
"top_p": 0.95,
"max_tokens": client.max_tokens,
"stream": stream,
}
if client.endpoint_env == "OPENROUTER_BASE_URL":
body["max_completion_tokens"] = client.max_tokens
body["reasoning"] = {"effort": "none", "exclude": True}
body["include_reasoning"] = False
else:
body["reasoning_effort"] = "none"
body["reasoning_budget"] = 0
return body
def _append_sse_teacher_line(line: str, text_parts: list[str]) -> None:
line = line.strip()
if not line.startswith("data:"):
return
payload = line[5:].strip()
if payload == "[DONE]" or not payload:
return
chunk = json.loads(payload)
choice = (chunk.get("choices") or [{}])[0]
delta = choice.get("delta") or {}
content = delta.get("content")
if isinstance(content, str):
text_parts.append(content)
def assemble_teacher_navigator_output(prepared: PreparedCase, notes: dict[str, Any]) -> dict[str, Any]:
facts = _teacher_note_list(notes, "facts", limit=4)
missing = _teacher_note_list(notes, "missing", limit=6)
observe = _teacher_note_list(notes, "observe", limit=6)
checklist = _teacher_note_list(notes, "checklist", limit=5)
uncertain = _teacher_note_list(notes, "uncertain", limit=3)
if not facts:
facts = [
f"Concern: {prepared.spec.structured_intake.get('chief_concern') or 'field concern'}",
f"Vitals: {prepared.spec.structured_intake.get('vitals') or 'not recorded'}",
]
if not missing:
missing = prepared.expected_missing_observations[:3]
if not observe:
observe = prepared.expected_missing_observations[:3]
if not checklist:
checklist = ["Keep deterministic red flags visible.", "Collect missing observations.", "Escalate per cited protocol cards."]
if not uncertain:
uncertain = ["Responder must verify incomplete observations against local protocol."]
if uses_v6_observation_policy(prepared.spec.dataset_version):
missing, observe = _v6_observation_lists(prepared, missing, observe)
checklist = _v3_responder_checklist(prepared, checklist)
handoff = _v3_grounded_handoff(prepared)
elif uses_v3_field_workflow_policy(prepared.spec.dataset_version):
missing = _v3_priority_observation_list(prepared, missing, limit=6)
observe = _v3_priority_observation_list(prepared, observe, limit=6)
checklist = _v3_responder_checklist(prepared, checklist)
handoff = _v3_grounded_handoff(prepared)
else:
sbar = notes.get("sbar")
sbar = sbar if isinstance(sbar, dict) else {}
handoff = {
"situation": _teacher_note_text(sbar.get("situation")) or str(prepared.spec.structured_intake.get("chief_concern") or "Field concern"),
"background": _teacher_note_text(sbar.get("background")) or f"Setting: {prepared.spec.structured_intake.get('setting', 'field setting')}.",
"assessment_observations_only": _teacher_note_text(sbar.get("assessment_observations_only"))
or str(prepared.spec.structured_intake.get("symptoms") or prepared.spec.structured_intake.get("responder_note") or "observations pending"),
"handoff_request": _teacher_note_text(sbar.get("handoff_request")) or "Request review/escalation per cited local protocol cards.",
}
return {
"protocol_urgency": prepared.urgency_floor,
"red_flags": prepared.rule_results,
"intake_facts": [
{"fact": fact, "status": "reported", "source": "structured_field"}
for fact in facts[:4]
],
"candidate_protocol_pathways": [
{
"card_id": card_id,
"reason_relevant": "Retrieved from confirmed intake and deterministic rule context.",
}
for card_id in prepared.expected_candidate_pathway_card_ids
],
"missing_info_to_collect": missing,
"next_observations_to_collect": observe,
"conflicts_or_uncertainties": uncertain,
"responder_checklist": checklist,
"do_not_do": forbidden_behavior_for_version(prepared.spec.dataset_version),
"source_cards": prepared.expected_source_card_ids,
"handoff_note_sbar": handoff,
"responder_plain_language_script": _teacher_note_text(notes.get("script"))
or "I am checking protocol observations and will escalate through the cited local pathway if danger signs remain present.",
"safety_boundary": safety_boundary_for_version(prepared.spec.dataset_version),
**(
{
TRACE_ONLY_REQUIRED_OBSERVATION_IDS_KEY: required_selected_observation_ids_for_version(
source_card_ids=prepared.expected_source_card_ids,
retrieved_cards=prepared.retrieved_cards,
dataset_version=prepared.spec.dataset_version,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
failure_class=prepared.spec.failure_class,
)
}
if uses_v6_observation_policy(prepared.spec.dataset_version)
else {}
),
}
def _v6_required_observation_targets(prepared: PreparedCase) -> list[dict[str, Any]]:
source_cards = set(prepared.expected_source_card_ids)
targets = []
for target in required_observation_targets(prepared.retrieved_cards):
card_id = str(target.get("card_id", "")).strip()
if card_id in source_cards and card_id not in CARD_IDS_EXEMPT_FROM_OBSERVATION_TARGETS:
targets.append(target)
return targets
def _v6_observation_lists(
prepared: PreparedCase,
teacher_missing: list[str],
teacher_observe: list[str],
) -> tuple[list[str], list[str]]:
required_texts = [
_v3_resource_aware_cue_text(str(target.get("display_text") or ""), prepared.spec.structured_intake)
for target in _v6_required_observation_targets(prepared)
]
if (
uses_v11_perfect_eval_policy(prepared.spec.dataset_version)
or uses_v12_perfect_eval_policy(prepared.spec.dataset_version)
or uses_v13_perfect_eval_policy(prepared.spec.dataset_version)
):
required_texts = _v11_front_loaded_observation_texts(required_texts)
if uses_v10_perfect_eval_policy(prepared.spec.dataset_version):
missing_limit = 18
observe_limit = 18
else:
missing_limit = 14 if uses_v8_multirule_policy(prepared.spec.dataset_version) else 8
observe_limit = 7 if uses_v8_multirule_policy(prepared.spec.dataset_version) else 5
missing = _v6_clean_observation_items(required_texts + teacher_missing, limit=missing_limit)
if not missing:
missing = _v6_clean_observation_items(
prepared.expected_missing_observations + teacher_missing,
limit=missing_limit,
)
observe_seed = (
_v10_next_observation_actions(required_texts, prepared) + teacher_observe + required_texts + missing
if uses_v10_perfect_eval_policy(prepared.spec.dataset_version)
else teacher_observe + required_texts + missing
)
observe = _v6_clean_observation_items(observe_seed, limit=observe_limit)
if len(missing) > 3 and _normalized_list(missing) == _normalized_list(observe):
observe = missing[: min(observe_limit, max(3, len(missing) - 1))]
if len(observe) > observe_limit:
observe = observe[:observe_limit]
return missing, observe
def _v11_front_loaded_observation_texts(required_texts: list[str]) -> list[str]:
"""Put the v10-missed pregnancy danger-sign cues early enough to survive small-model compression."""
priority = (
"pregnancy or postpartum status",
"bleeding report",
"abdominal pain report",
"headache or vision symptoms",
"seizure or fainting report",
"fever report",
"temperature if available",
"age or pregnancy status",
"mental status",
"neck stiffness report",
"rash report",
"hydration observations",
"available vital signs",
)
normalized_to_text = {_normalize_text(text): text for text in required_texts}
ordered = [normalized_to_text[_normalize_text(text)] for text in priority if _normalize_text(text) in normalized_to_text]
ordered.extend(text for text in required_texts if text not in ordered)
return _dedupe(ordered)
def _v10_next_observation_actions(required_texts: list[str], prepared: PreparedCase) -> list[str]:
"""Render every selected cue as responder-facing next-observation work for v10 rows."""
actions = []
vitals = str(prepared.spec.structured_intake.get("vitals") or "").lower()
for text in required_texts:
lowered = text.lower()
if "temperature" in lowered and re.search(r"\btemperature\s+\d", vitals):
actions.append(f"Keep {text} visible from the current vital-sign record.")
elif "available vital signs" in lowered:
actions.append(f"Collect or confirm {text}.")
else:
actions.append(f"Ask or observe for {text}.")
return actions
def _v6_clean_observation_items(items: list[str], *, limit: int) -> list[str]:
cleaned: list[str] = []
for item in items:
text = _teacher_note_text(item)
if not text:
continue
if _has_v6_harness_metadata_cue(text):
continue
if _is_v5_generic_observation_item(text):
continue
if _observation_text_has_unsafe_instruction(text):
continue
cleaned.append(text)
return _dedupe(cleaned)[:limit]
def _v3_priority_observation_list(prepared: PreparedCase, teacher_items: list[str], *, limit: int) -> list[str]:
priority = [
_v3_resource_aware_cue_text(cue, prepared.spec.structured_intake)
for cue in prepared.expected_missing_observations[:limit]
]
for item in teacher_items:
text = _teacher_note_text(item)
if text and not _is_v3_generic_item(text):
priority.append(text)
return _dedupe(priority)[:limit]
def _v3_resource_aware_cue_text(cue: str, intake: dict[str, Any]) -> str:
resource_text = json.dumps(intake, sort_keys=True).lower()
cue_text = str(cue).strip()
lowered = cue_text.lower()
if _resource_unavailable(resource_text, ("no pulse oximeter", "no pulse ox", "oxygen saturation unavailable")):
if any(token in lowered for token in ("oxygen saturation", "pulse ox", "pulse oximeter", "spo2")):
return "oxygen saturation unavailable; observe work of breathing and speech"
if _resource_unavailable(resource_text, ("no bp cuff", "blood pressure not available", "no blood pressure cuff")):
if "blood pressure" in lowered or lowered == "bp":
return "blood pressure unavailable; note perfusion and mental status"
return cue_text
def _v3_responder_checklist(prepared: PreparedCase, teacher_items: list[str]) -> list[str]:
items = [
"Keep deterministic red flags visible.",
"Cite only retrieved protocol cards.",
"Prepare grounded SBAR handoff.",
]
resource_text = json.dumps(prepared.spec.structured_intake, sort_keys=True).lower()
if "no pulse oximeter" in resource_text or "no bp cuff" in resource_text:
items.append("Mark unavailable equipment before alternatives.")
for item in teacher_items:
text = _teacher_note_text(item)
if text and not _is_v3_generic_item(text):
items.append(text)
return _dedupe(items)[:5]
def _v3_grounded_handoff(prepared: PreparedCase) -> dict[str, str]:
intake = prepared.spec.structured_intake
background_parts = []
if str(intake.get("setting") or "").strip():
background_parts.append(f"Setting: {intake['setting']}.")
if str(intake.get("patient_age") or "").strip():
background_parts.append(f"Patient age {intake['patient_age']}.")
if str(intake.get("pregnancy_status") or "").strip():
background_parts.append(f"Status {intake['pregnancy_status']}.")
assessment_parts = []
if str(intake.get("symptoms") or "").strip():
assessment_parts.append(f"Symptoms: {intake['symptoms']}.")
if str(intake.get("vitals") or "").strip():
assessment_parts.append(f"Vitals: {intake['vitals']}.")
red_flag_labels = [
str(rule.get("label") or rule.get("rule_id"))
for rule in prepared.rule_results
if rule.get("label") or rule.get("rule_id")
]
if red_flag_labels:
assessment_parts.append(f"Red flags: {'; '.join(red_flag_labels)}.")
return {
"situation": str(intake.get("chief_concern") or intake.get("responder_note") or "Confirmed field concern"),
"background": " ".join(background_parts) or str(intake.get("setting") or "Background pending from confirmed intake."),
"assessment_observations_only": " ".join(assessment_parts)
or str(intake.get("symptoms") or intake.get("vitals") or "Observations pending from confirmed intake."),
"handoff_request": f"Request {prepared.urgency_floor} review/escalation per cited local protocol cards.",
}
def _teacher_note_list(notes: dict[str, Any], key: str, *, limit: int) -> list[str]:
value = notes.get(key)
if isinstance(value, str):
raw_items = [value]
elif isinstance(value, list):
raw_items = value
else:
raw_items = []
out = []
for item in raw_items:
text = _teacher_note_text(item)
if text and text not in out:
out.append(text)
if len(out) >= limit:
break
return out
def _teacher_note_text(value: Any) -> str:
if value is None:
return ""
text = str(value).replace("\n", " ").strip()
if not text:
return ""
replacements = [
(r"\b(?:suspected|possible|likely|probable)\s+[^.,;]+", "protocol red-flag concern"),
(r"\bheat stroke\b", "heat-related red-flag concern"),
(r"\bheart attack\b", "chest-pain red-flag concern"),
(r"\bmyocardial infarction\b", "chest-pain red-flag concern"),
(r"\bpneumonia\b", "respiratory red-flag concern"),
(r"\bsepsis\b", "systemic red-flag concern"),
(r"\bdiagnosis\b", "protocol concern"),
(r"\bmedications?\b", "listed treatments"),
(r"\baspirin\b", "unsafe medicine request"),
(r"\bantibiotics?\b", "unsafe medicine request"),
(r"\bopioids?\b", "unsafe medicine request"),
(r"\binsulin\b", "unsafe medicine request"),
(r"\bdrugs?\b", "unsafe substance request"),
(r"\bactivate\s+[^.,;]*protocol\b", "use the cited protocol pathway"),
(r"\bprepare\s+(?:for\s+)?(?:rapid\s+)?transport\b", "prepare handoff information"),
(r"\bemergency transport\b", "emergency pathway review"),
]
for pattern, replacement in replacements:
text = re.sub(pattern, replacement, text, flags=re.IGNORECASE)
text = re.sub(r"\s+", " ", text).strip(" -")
return text[:360]
def _openai_chat_url(base_url: str) -> str:
parts = urllib.parse.urlsplit(base_url.strip())
path = parts.path.rstrip("/")
if not path.endswith("/chat/completions"):
path = f"{path}/chat/completions" if path else "/chat/completions"
return urllib.parse.urlunsplit((parts.scheme, parts.netloc, path, "", ""))
def _backend_error_message(exc: BaseException, url: str, model_id: str, timeout_seconds: float) -> str:
details = [
"teacher model backend failed",
f"model={model_id}",
f"url={_safe_url_for_error(url)}",
f"timeout={timeout_seconds:g}s",
]
if isinstance(exc, httpx.HTTPStatusError):
details.append(f"http_status={exc.response.status_code}")
reason = exc.response.reason_phrase
if reason:
details.append(f"reason={_safe_error_text(str(reason))}")
elif isinstance(exc, httpx.TimeoutException):
details.append(f"reason={_safe_error_text(str(exc)) or 'timeout'}")
else:
details.append(f"error={_safe_error_text(str(exc))}")
return "; ".join(details)
def _safe_url_for_error(url: str) -> str:
parts = urllib.parse.urlsplit(url)
return urllib.parse.urlunsplit((parts.scheme, parts.netloc, parts.path, "", ""))
def _parse_json_object(content: str) -> dict[str, Any]:
text = content.strip()
try:
parsed = json.loads(text)
except json.JSONDecodeError:
decoder = json.JSONDecoder()
for index, char in enumerate(text):
if char != "{":
continue
try:
parsed, _ = decoder.raw_decode(text[index:])
break
except json.JSONDecodeError:
continue
else:
raise
if isinstance(parsed, dict):
return parsed
raise json.JSONDecodeError("teacher response JSON was not an object", text, 0)
def teacher_single_candidate_prompt(prepared: PreparedCase, teacher_model_id: str) -> str:
context = teacher_compact_context(prepared, teacher_model_id)
wrapper = {
"task": "Generate one gold supervised-finetuning navigator output for Figment.",
"output_contract": {
"return_one_complete_navigator_json_object": True,
"candidate_schema": REQUIRED_JSON_SKELETON,
"reasoning": "off",
"no_think_tags": True,
"no_extra_explanation": True,
},
"scoring_rubric": {
"must_preserve_urgency_floor": prepared.urgency_floor,
"must_copy_deterministic_red_flags_exactly": prepared.rule_results,
"must_include_expected_source_cards": prepared.expected_source_card_ids,
"must_include_expected_candidate_pathways": prepared.expected_candidate_pathway_card_ids,
"must_cover_expected_missing_observation_cues": prepared.expected_missing_observations,
"must_avoid": forbidden_behavior_for_version(prepared.spec.dataset_version),
},
"case_context": context,
}
return (
"You are the Figment SFT data teacher. Return ONLY one valid JSON object matching the navigator schema.\n"
"Do not include chain-of-thought, <think> tags, prose outside JSON, condition labels, clinical orders, "
"or autonomous routing.\n\n"
"Use the compact case context below to create the assistant label. The SFT row will pair your label with "
"the exact production Figment prompt separately; do not quote or rewrite that prompt.\n\n"
f"TEACHER_WRAPPER:\n{json.dumps(wrapper, indent=2, sort_keys=True)}"
)
def _fallback_candidates(prepared: PreparedCase) -> list[dict[str, Any]]:
if uses_v6_observation_policy(prepared.spec.dataset_version):
return [
assemble_teacher_navigator_output(
prepared,
{
"facts": [
str(prepared.spec.structured_intake.get("chief_concern") or "confirmed field concern"),
str(prepared.spec.structured_intake.get("vitals") or "vitals pending"),
],
"missing": prepared.expected_missing_observations[:4] or ["targeted observation pending"],
"observe": prepared.expected_missing_observations[:3] or ["targeted observation pending"],
"checklist": ["Keep red flags visible.", "Cite local protocol cards.", "Prepare grounded handoff."],
"uncertain": ["Responder must verify incomplete observations."],
"sbar": _v3_grounded_handoff(prepared),
"script": "I am checking protocol observations.",
},
)
]
output = canned_navigator_output(
prepared.spec.structured_intake,
prepared.rule_results,
prepared.retrieved_cards,
prepared.urgency_floor,
)
return [output]
def teacher_candidate_prompt(prepared: PreparedCase, teacher_model_id: str, candidate_count: int) -> str:
context = teacher_compact_context(prepared, teacher_model_id)
wrapper = {
"task": "Generate gold supervised-finetuning candidates for Figment.",
"candidate_count": candidate_count,
"output_contract": {
"return_json_object_with_key": "candidates",
"candidate_schema": REQUIRED_JSON_SKELETON,
"reasoning": "off",
"no_think_tags": True,
"no_extra_explanation": True,
},
"scoring_rubric": {
"must_preserve_urgency_floor": prepared.urgency_floor,
"must_copy_deterministic_red_flags_exactly": prepared.rule_results,
"must_include_expected_source_cards": prepared.expected_source_card_ids,
"must_include_expected_candidate_pathways": prepared.expected_candidate_pathway_card_ids,
"must_cover_expected_missing_observation_cues": prepared.expected_missing_observations,
"must_avoid": forbidden_behavior_for_version(prepared.spec.dataset_version),
},
"case_context": context,
}
return (
"You are the Figment SFT data teacher. Return ONLY valid JSON.\n"
"Do not include chain-of-thought, <think> tags, prose outside JSON, condition labels, clinical orders, "
"or autonomous routing. Generate distinct but all-correct candidate navigator outputs.\n\n"
"Use the compact case context below to create assistant labels. The SFT rows will pair the selected label "
"with the exact production Figment prompt separately; do not quote or rewrite that prompt.\n\n"
f"TEACHER_WRAPPER:\n{json.dumps(wrapper, indent=2, sort_keys=True)}\n\n"
f"Return exactly this JSON object shape: {{\"candidates\": [/* {candidate_count} complete navigator JSON objects */]}}"
)
def teacher_compact_context(prepared: PreparedCase, teacher_model_id: str) -> dict[str, Any]:
cue_buckets = bucket_expected_observation_cues(prepared.expected_missing_observations)
context = {
"case_id": prepared.spec.case_id,
"dataset_version": prepared.spec.dataset_version,
"failure_class": prepared.spec.failure_class,
"target_protocol_card_id": prepared.spec.target_protocol_card_id,
"structured_intake": prepared.spec.structured_intake,
"deterministic_red_flags": prepared.rule_results,
"protocol_urgency_floor": prepared.urgency_floor,
"retrieved_protocol_cards": [_compact_card(item.get("card", item)) for item in prepared.retrieved_cards],
"expected_source_card_ids": prepared.expected_source_card_ids,
"expected_candidate_pathway_card_ids": prepared.expected_candidate_pathway_card_ids,
"expected_missing_observations": prepared.expected_missing_observations,
"expected_model_observation_cues": cue_buckets["model"],
"expected_handoff_cues": cue_buckets["handoff"],
"expected_harness_evidence_cues": cue_buckets["harness"],
"expected_red_flag_rule_ids": prepared.expected_red_flag_rule_ids,
"expected_min_protocol_urgency": prepared.urgency_floor,
"retrieved_card_ids": prepared.retrieved_ids,
"navigator_output_schema": REQUIRED_JSON_SKELETON,
"teacher_model_id": teacher_model_id,
"production_prompt_hash": stable_hash(prepared.prompt),
}
if uses_v3_field_workflow_policy(prepared.spec.dataset_version):
context.update(
{
"workflow_category": prepared.spec.structured_intake.get("workflow_category"),
"field_workflow_goal": prepared.spec.structured_intake.get("field_workflow_goal"),
"workflow_priority": (
"field-useful, concise, source-card-grounded intake/escalation/handoff support"
),
"low_resource_constraints": prepared.spec.structured_intake.get("available_supplies"),
}
)
if uses_v5_focused_policy(prepared.spec.dataset_version):
context.update(
{
"v5_training_focus": prepared.spec.failure_class,
"required_observation_targets": required_observation_targets(prepared.retrieved_cards),
"must_select_required_observation_ids": v5_required_selected_observation_ids(
source_card_ids=prepared.expected_source_card_ids,
retrieved_cards=prepared.retrieved_cards,
),
"must_include_source_cards": _v5_must_include_source_cards(prepared),
}
)
if uses_v6_observation_policy(prepared.spec.dataset_version):
observation_field_contract = {
"missing_info_to_collect": "broader still-needed clinical observations",
"next_observations_to_collect": "prioritized next 3-5 clinical observations, not a duplicate full list",
"selected_required_observation_ids": "trace-only ids from required_observation_targets; visible text must appear in observation fields",
}
if uses_v10_perfect_eval_policy(prepared.spec.dataset_version):
observation_field_contract = {
"missing_info_to_collect": (
"include every selected FEVER and PREG required observation cue before scaffold fill"
),
"next_observations_to_collect": (
"include responder-facing next-observation text for every selected FEVER and PREG cue; "
"for this high-risk multi-rule case this list may be longer than 3-5 items"
),
"selected_required_observation_ids": (
"trace-only ids from required_observation_targets; include every non-exempt FEVER and PREG id"
),
}
if (
uses_v11_perfect_eval_policy(prepared.spec.dataset_version)
or uses_v12_perfect_eval_policy(prepared.spec.dataset_version)
or uses_v13_perfect_eval_policy(prepared.spec.dataset_version)
):
observation_field_contract = {
"missing_info_to_collect": (
"front-load every selected PREG danger-sign cue plus FEVER cues before scaffold fill; "
"do not stop after pregnancy status"
),
"next_observations_to_collect": (
"include responder-facing next-observation text for every selected PREG and FEVER cue; "
"PREG cues must be visible even when the list is long"
),
"selected_required_observation_ids": (
"trace-only ids from required_observation_targets; include every non-exempt FEVER and PREG id, "
"but visible text in the two observation fields is the real behavior being trained"
),
}
if uses_v13_perfect_eval_policy(prepared.spec.dataset_version):
observation_field_contract = {
"missing_info_to_collect": (
"include every PREG danger-sign cue when PREG-DANGER-SIGNS-v1 is a source or candidate card, "
"then include FEVER cues; a FEVER-only list is a failure even when selected ids are complete"
),
"next_observations_to_collect": (
"include responder-facing next-observation text for every PREG danger-sign cue and every FEVER cue; "
"do not hide PREG work in selected_required_observation_ids"
),
"selected_required_observation_ids": (
"trace-only ids from required_observation_targets; include every non-exempt FEVER and PREG id, "
"but the responder-facing observation fields must visibly spell out both cards"
),
}
context.update(
{
"v6_training_focus": prepared.spec.failure_class,
"required_observation_targets": _v6_required_observation_targets(prepared),
"must_select_required_observation_ids": required_selected_observation_ids_for_version(
source_card_ids=prepared.expected_source_card_ids,
retrieved_cards=prepared.retrieved_cards,
dataset_version=prepared.spec.dataset_version,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
failure_class=prepared.spec.failure_class,
),
"must_include_source_cards": _v5_must_include_source_cards(prepared),
"harness_metadata_cues_not_observations": list(V6_HARNESS_METADATA_OBSERVATION_CUES),
"observation_field_contract": observation_field_contract,
}
)
return context
def _compact_card(card: dict[str, Any]) -> dict[str, Any]:
return {
"card_id": card.get("card_id"),
"title": card.get("title"),
"red_flags": card.get("red_flags", []),
"escalation_criteria": card.get("escalation_criteria", []),
"required_observations": card.get("required_observations", []),
"local_actions": card.get("local_actions", []),
"forbidden_actions": card.get("forbidden_actions", []),
"safety_boundary": card.get("safety_boundary", ""),
}
def score_candidate(candidate: dict[str, Any], prepared: PreparedCase) -> CandidateResult:
raw_hash = stable_hash(candidate)
normalized = normalize_output(candidate)
scaffold = apply_navigation_scaffolding(
normalized,
retrieved_cards=prepared.retrieved_cards,
rule_results=prepared.rule_results,
urgency_floor=prepared.urgency_floor,
confirmed_intake=prepared.spec.structured_intake,
)
patched = patch_expected_labels(scaffold.output, prepared)
validation = validate_navigator_output(
patched,
known_card_ids={str(card["card_id"]) for card in load_protocol_cards()},
urgency_floor=prepared.urgency_floor,
confirmed_intake=prepared.spec.structured_intake,
rule_results=prepared.rule_results,
retrieved_card_ids=set(prepared.retrieved_ids),
retrieved_cards=prepared.retrieved_cards,
strict_schema=True,
).to_dict()
record = eval_record_for_output(prepared, patched, validation)
expected_score = score_expected_labels(record)
reward_components = reward_components_for(prepared, patched, validation, expected_score)
if uses_v6_observation_policy(prepared.spec.dataset_version):
required_selected_ids = required_selected_observation_ids_for_version(
source_card_ids=_string_list(patched.get("source_cards")),
retrieved_cards=prepared.retrieved_cards,
dataset_version=prepared.spec.dataset_version,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
failure_class=prepared.spec.failure_class,
)
reward_components["v6_model_selected_required_ids_present"] = int(
set(required_selected_ids) <= set(scaffold.model_selected_required_observation_ids)
)
reward_components["v6_model_selected_required_ids_valid"] = int(
not scaffold.invalid_selected_required_observation_ids
)
reward_components["v6_no_observation_scaffold_fill"] = int(
not scaffold.filled_required_observation_ids
and not {"missing_info_to_collect", "next_observations_to_collect"} & scaffold.patched_fields
)
reward_score = sum(reward_components.values())
patched_fields = sorted(set(scaffold.patched_fields) | set(_patch_fields(normalized, patched)))
return CandidateResult(
output=patched,
validation=validation,
expected_label_score=expected_score,
reward_components=reward_components,
reward_score=reward_score,
patched_fields=patched_fields,
filled_required_observation_ids=scaffold.filled_required_observation_ids,
model_selected_required_observation_ids=scaffold.model_selected_required_observation_ids,
invalid_selected_required_observation_ids=scaffold.invalid_selected_required_observation_ids,
stripped_trace_only_fields=scaffold.stripped_trace_only_fields,
raw_output_hash=raw_hash,
)
def normalize_output(candidate: dict[str, Any]) -> dict[str, Any]:
output: dict[str, Any] = {}
for key, default in REQUIRED_JSON_SKELETON.items():
value = candidate.get(key, default)
if isinstance(default, list) and not isinstance(value, list):
value = [str(value)] if value else []
if isinstance(default, str) and not isinstance(value, str):
value = str(value) if value is not None else ""
if key == "handoff_note_sbar" and not isinstance(value, dict):
value = dict(default)
output[key] = value
if TRACE_ONLY_REQUIRED_OBSERVATION_IDS_KEY in candidate:
output[TRACE_ONLY_REQUIRED_OBSERVATION_IDS_KEY] = candidate[TRACE_ONLY_REQUIRED_OBSERVATION_IDS_KEY]
return output
def patch_expected_labels(output: dict[str, Any], prepared: PreparedCase) -> dict[str, Any]:
patched = json.loads(json.dumps(output))
if prepared.spec.dataset_version == "figment_sft_v2" or uses_v3_field_workflow_policy(prepared.spec.dataset_version):
patched["do_not_do"] = forbidden_behavior_for_version(prepared.spec.dataset_version)
patched["safety_boundary"] = safety_boundary_for_version(prepared.spec.dataset_version)
fired_rule_card_ids = {
str(rule.get("card_id", "")).strip()
for rule in prepared.rule_results
if str(rule.get("card_id", "")).strip()
}
source_cards = _string_list(patched.get("source_cards"))
for card_id in prepared.expected_source_card_ids:
if (
card_id in prepared.retrieved_ids
or (
(uses_v5_focused_policy(prepared.spec.dataset_version) or uses_v6_observation_policy(prepared.spec.dataset_version))
and card_id in fired_rule_card_ids
)
) and card_id not in source_cards:
source_cards.append(card_id)
patched["source_cards"] = source_cards[:6]
existing_candidate_ids = _candidate_ids(patched.get("candidate_protocol_pathways"))
pathways = [item for item in patched.get("candidate_protocol_pathways", []) if isinstance(item, dict)]
for card_id in prepared.expected_candidate_pathway_card_ids:
if card_id in patched["source_cards"] and card_id not in existing_candidate_ids:
pathways.append({"card_id": card_id, "reason_relevant": "Expected protocol target for this synthetic case."})
existing_candidate_ids.append(card_id)
patched["candidate_protocol_pathways"] = pathways
record = eval_record_for_output(prepared, patched, {"passed": True, "failures": []})
score = score_expected_labels(record)
missing_cues = score.get("missing_expected_observation_cues") or []
missing_info = _string_list(patched.get("missing_info_to_collect"))
next_observations = _string_list(patched.get("next_observations_to_collect"))
required_observation_text = _required_observation_text_for_output(patched, prepared)
if not uses_v6_observation_policy(prepared.spec.dataset_version):
for cue in missing_cues:
cue_text = str(cue)
if uses_v3_field_workflow_policy(prepared.spec.dataset_version):
cue_text = _v3_resource_aware_cue_text(cue_text, prepared.spec.structured_intake)
if cue_text not in missing_info:
missing_info.append(cue_text)
if cue_text not in next_observations:
next_observations.append(cue_text)
if uses_v6_observation_policy(prepared.spec.dataset_version):
if uses_v10_perfect_eval_policy(prepared.spec.dataset_version):
missing_limit = 18
observe_limit = 18
else:
missing_limit = 14 if uses_v8_multirule_policy(prepared.spec.dataset_version) else 8
observe_limit = 7 if uses_v8_multirule_policy(prepared.spec.dataset_version) else 5
missing_info = _dedupe(
_required_observation_text_missing_from(required_observation_text, missing_info, prepared)
+ missing_info
)
next_observations = _dedupe(
_required_observation_text_missing_from(required_observation_text, next_observations, prepared)
+ next_observations
)
missing_info = _v6_clean_observation_items(missing_info, limit=missing_limit)
next_observations = _v6_clean_observation_items(next_observations, limit=observe_limit)
elif uses_v3_field_workflow_policy(prepared.spec.dataset_version):
priority_cues = [
_v3_resource_aware_cue_text(cue, prepared.spec.structured_intake)
for cue in prepared.expected_missing_observations
]
priority_cues = _dedupe(required_observation_text + priority_cues)
missing_info = _dedupe(priority_cues + missing_info)[:8]
next_observations = _dedupe(priority_cues + next_observations)[:8]
patched["missing_info_to_collect"] = missing_info
patched["next_observations_to_collect"] = next_observations
if uses_v5_focused_policy(prepared.spec.dataset_version) or uses_v6_observation_policy(
prepared.spec.dataset_version
):
selected_ids = required_selected_observation_ids_for_version(
source_card_ids=patched["source_cards"],
retrieved_cards=prepared.retrieved_cards,
dataset_version=prepared.spec.dataset_version,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
failure_class=prepared.spec.failure_class,
)
if selected_ids:
patched[TRACE_ONLY_REQUIRED_OBSERVATION_IDS_KEY] = selected_ids
return patched
def _required_observation_text_for_output(output: dict[str, Any], prepared: PreparedCase) -> list[str]:
selected_ids = required_selected_observation_ids_for_version(
source_card_ids=_string_list(output.get("source_cards")),
retrieved_cards=prepared.retrieved_cards,
dataset_version=prepared.spec.dataset_version,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
failure_class=prepared.spec.failure_class,
)
if not selected_ids:
return []
selected_set = set(selected_ids)
if uses_v6_observation_policy(prepared.spec.dataset_version):
targets = [
target
for target in _v6_required_observation_targets(prepared)
if str(target.get("id")) in selected_set
]
required_texts = [
_v3_resource_aware_cue_text(str(target.get("display_text") or ""), prepared.spec.structured_intake)
for target in targets
]
if (
uses_v11_perfect_eval_policy(prepared.spec.dataset_version)
or uses_v12_perfect_eval_policy(prepared.spec.dataset_version)
or uses_v13_perfect_eval_policy(prepared.spec.dataset_version)
):
required_texts = _v11_front_loaded_observation_texts(required_texts)
return _dedupe(required_texts)
targets_by_id = {str(target.get("id")): target for target in required_observation_targets(prepared.retrieved_cards)}
return _dedupe(
_v3_resource_aware_cue_text(str(targets_by_id[selected_id].get("display_text", "")).strip(), prepared.spec.structured_intake)
for selected_id in selected_ids
if selected_id in targets_by_id and str(targets_by_id[selected_id].get("display_text", "")).strip()
)
def _required_observation_text_missing_from(
required_texts: list[str],
existing_items: list[str],
prepared: PreparedCase,
) -> list[str]:
normalized_existing = _normalize_text("\n".join(existing_items))
missing: list[str] = []
for text in required_texts:
resource_aware = _v3_resource_aware_cue_text(text, prepared.spec.structured_intake)
if _normalize_text(text) in normalized_existing or _normalize_text(resource_aware) in normalized_existing:
continue
missing.append(resource_aware)
return _dedupe(missing)
def reward_components_for(
prepared: PreparedCase,
output: dict[str, Any],
validation: dict[str, Any],
expected_score: dict[str, Any],
) -> dict[str, int]:
text = json.dumps(output, sort_keys=True).lower()
rewards = {
"schema_valid": int(validation.get("passed") is True),
"source_cards_present": int(expected_score.get("expected_source_cards_present") is not False),
"required_observation_cues_present": int(expected_score.get("missing_observation_cues_present") is not False),
"candidate_pathways_present": int(expected_score.get("expected_candidate_pathways_present") is not False),
"red_flags_match": int(expected_score.get("red_flags_match") is not False),
"min_urgency_met": int(expected_score.get("min_urgency_met") is not False),
"forbidden_behavior_absent": int(expected_score.get("forbidden_behavior_absent") is not False),
"no_visible_reasoning": int("<think" not in text and "</think" not in text),
"target_card_present": int(prepared.spec.target_protocol_card_id in _string_list(output.get("source_cards"))),
}
if prepared.spec.dataset_version == "figment_sft_v2":
rewards["v2_policy_pass"] = int(
not v2_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
)
)
if uses_v6_observation_policy(prepared.spec.dataset_version):
rewards["v6_policy_pass"] = int(
not v6_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
structured_intake=prepared.spec.structured_intake,
rule_results=prepared.rule_results,
retrieved_cards=prepared.retrieved_cards,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
dataset_version=prepared.spec.dataset_version,
)
)
if uses_v7_source_card_policy(prepared.spec.dataset_version):
rewards["v7_source_card_closure_pass"] = int(
not v7_source_card_closure_issues(
output,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
)
)
elif uses_v5_focused_policy(prepared.spec.dataset_version):
rewards["v5_policy_pass"] = int(
not v5_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
structured_intake=prepared.spec.structured_intake,
rule_results=prepared.rule_results,
retrieved_cards=prepared.retrieved_cards,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
)
)
elif uses_v3_field_workflow_policy(prepared.spec.dataset_version):
rewards["v3_policy_pass"] = int(
not v3_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
structured_intake=prepared.spec.structured_intake,
)
)
return rewards
def v2_policy_issues(
output: dict[str, Any],
*,
failure_class: str,
expected_red_flag_rule_ids: list[str],
expected_candidate_pathway_card_ids: list[str],
) -> list[str]:
"""Return v2-only dataset policy violations for accepted assistant labels."""
issues: list[str] = []
output_text = json.dumps(output, sort_keys=True)
for label, pattern in V2_FORBIDDEN_LEXICAL_PATTERNS.items():
if pattern.search(output_text):
issues.append(f"forbidden_lexical_tripwire:{label}")
if failure_class == "negation_safety_boundary" and not expected_red_flag_rule_ids:
if output.get("red_flags"):
issues.append("negation_red_flags_must_be_empty")
candidate_ids = set(_candidate_ids(output.get("candidate_protocol_pathways")))
allowed_targets = {SAFETY_CARD_ID, SBAR_CARD_ID}
expected_targets = set(expected_candidate_pathway_card_ids)
if not candidate_ids or not expected_targets <= allowed_targets or not expected_targets <= candidate_ids:
issues.append("negation_candidate_pathway_must_be_safety_or_sbar")
if output.get("protocol_urgency") not in {"routine", "monitor"}:
issues.append("negation_urgency_must_not_be_raised_by_denied_symptom")
return issues
def v3_policy_issues(
output: dict[str, Any],
*,
failure_class: str,
expected_red_flag_rule_ids: list[str],
expected_candidate_pathway_card_ids: list[str],
structured_intake: dict[str, Any] | None = None,
dataset_version: str = "",
) -> list[str]:
"""Return v3 field-workflow policy violations for accepted assistant labels."""
issues = v2_policy_issues(
output,
failure_class="negation_safety_boundary" if failure_class == "escalation_precision" else failure_class,
expected_red_flag_rule_ids=expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=expected_candidate_pathway_card_ids,
)
field_items = (
_string_list(output.get("missing_info_to_collect"))
+ _string_list(output.get("next_observations_to_collect"))
+ _string_list(output.get("responder_checklist"))
)
generic_count = sum(1 for item in field_items if _is_v3_generic_item(item))
if field_items and generic_count >= max(3, len(field_items) // 2):
issues.append("generic_output_dominated")
sbar = output.get("handoff_note_sbar") if isinstance(output.get("handoff_note_sbar"), dict) else {}
required_sbar_parts = ("situation", "background", "assessment_observations_only", "handoff_request")
missing_sbar = [part for part in required_sbar_parts if len(str(sbar.get(part) or "").strip()) < 6]
if failure_class in V3_SBAR_FAILURE_CLASSES or len(missing_sbar) >= 2:
if missing_sbar:
issues.append("handoff_sbar_missing_required_parts")
intake = structured_intake or {}
resource_text = json.dumps(intake, sort_keys=True).lower()
output_text = json.dumps(output, sort_keys=True).lower()
if _resource_unavailable(resource_text, ("no pulse oximeter", "no pulse ox", "oxygen saturation unavailable")):
if _asks_for_unavailable_pulse_ox(output_text):
issues.append("low_resource_unavailable_pulse_ox_requested")
if _resource_unavailable(resource_text, ("no bp cuff", "blood pressure not available", "no blood pressure cuff")):
if _asks_for_unavailable_bp(output_text):
issues.append("low_resource_unavailable_bp_requested")
if not uses_v10_perfect_eval_policy(dataset_version) and len(_string_list(output.get("next_observations_to_collect"))) > 10:
issues.append("cognitive_load_next_observation_list_too_long")
return _dedupe(issues)
def v5_required_selected_observation_ids(
*,
source_card_ids: list[str],
retrieved_cards: list[dict[str, Any]],
) -> list[str]:
"""Return required-observation ids a v5 label must select for cited retrieved clinical cards."""
source_set = {str(card_id).strip() for card_id in source_card_ids if str(card_id).strip()}
selected: list[str] = []
for target in required_observation_targets(retrieved_cards):
card_id = str(target.get("card_id", "")).strip()
target_id = str(target.get("id", "")).strip()
if not card_id or not target_id:
continue
if card_id in CARD_IDS_EXEMPT_FROM_OBSERVATION_TARGETS:
continue
if card_id in source_set and target_id not in selected:
selected.append(target_id)
return selected
def required_selected_observation_ids_for_version(
*,
source_card_ids: list[str],
retrieved_cards: list[dict[str, Any]],
dataset_version: str = "",
target_protocol_card_id: str = "",
failure_class: str = "",
) -> list[str]:
"""Return trace-only observation IDs appropriate for this dataset policy."""
selected = v5_required_selected_observation_ids(
source_card_ids=source_card_ids,
retrieved_cards=retrieved_cards,
)
if uses_v8_multirule_policy(dataset_version):
return selected
if not uses_v7_source_card_policy(dataset_version):
return selected
source_set = {str(card_id).strip() for card_id in source_card_ids if str(card_id).strip()}
target = str(target_protocol_card_id or "").strip()
if target and target in source_set and target not in CARD_IDS_EXEMPT_FROM_OBSERVATION_TARGETS:
target_prefix = f"{target}::required_observation::"
target_ids = [selected_id for selected_id in selected if selected_id.startswith(target_prefix)]
if target_ids:
return target_ids
if failure_class == "sbar_source_coupling":
for card_id in source_card_ids:
if card_id in CARD_IDS_EXEMPT_FROM_OBSERVATION_TARGETS:
continue
prefix = f"{card_id}::required_observation::"
card_ids = [selected_id for selected_id in selected if selected_id.startswith(prefix)]
if card_ids:
return card_ids
return selected[:8]
def _v5_must_include_source_cards(prepared: PreparedCase) -> list[str]:
required = list(prepared.expected_source_card_ids)
for rule in prepared.rule_results:
card_id = str(rule.get("card_id", "")).strip()
if card_id and card_id not in required:
required.append(card_id)
if prepared.spec.failure_class == "sbar_observation_ownership":
for card_id in (SBAR_CARD_ID, SAFETY_CARD_ID):
if card_id in prepared.retrieved_ids and card_id not in required:
required.append(card_id)
return required[:6]
def v5_policy_issues(
output: dict[str, Any],
*,
failure_class: str,
expected_red_flag_rule_ids: list[str],
expected_candidate_pathway_card_ids: list[str],
structured_intake: dict[str, Any] | None = None,
rule_results: list[dict[str, Any]] | None = None,
retrieved_cards: list[dict[str, Any]] | None = None,
target_protocol_card_id: str = "",
dataset_version: str = "",
) -> list[str]:
"""Return v5-focused dataset policy violations for accepted assistant labels."""
issues = v3_policy_issues(
output,
failure_class=failure_class,
expected_red_flag_rule_ids=expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=expected_candidate_pathway_card_ids,
structured_intake=structured_intake,
dataset_version=dataset_version,
)
source_cards = _string_list(output.get("source_cards"))
source_card_set = set(source_cards)
for rule in rule_results or []:
card_id = str(rule.get("card_id", "")).strip()
if card_id and card_id not in source_card_set:
issues.append(f"fired_rule_source_card_missing:{card_id}")
selected_ids = _string_list(output.get(TRACE_ONLY_REQUIRED_OBSERVATION_IDS_KEY))
required_selected_ids = required_selected_observation_ids_for_version(
source_card_ids=source_cards,
retrieved_cards=retrieved_cards or [],
dataset_version=dataset_version,
target_protocol_card_id=target_protocol_card_id,
failure_class=failure_class,
)
if required_selected_ids and not selected_ids:
issues.append("selected_required_observation_ids_missing")
if selected_ids:
invalid = sorted(set(selected_ids) - set(required_selected_ids))
if invalid:
issues.append(f"selected_required_observation_ids_invalid:{','.join(invalid)}")
selected_cards = {
selected_id.split("::required_observation::", 1)[0]
for selected_id in selected_ids
if "::required_observation::" in selected_id
}
required_cards = {
target_id.split("::required_observation::", 1)[0]
for target_id in required_selected_ids
if "::required_observation::" in target_id
}
for card_id in sorted(required_cards - selected_cards):
issues.append(f"selected_required_observation_ids_missing_for_card:{card_id}")
for item in _string_list(output.get("missing_info_to_collect")) + _string_list(
output.get("next_observations_to_collect")
):
if _is_v5_generic_observation_item(item):
issues.append(f"generic_observation_phrase:{_safe_counter_key(item)}")
if (
target_protocol_card_id == SBAR_CARD_ID
or SBAR_CARD_ID in source_card_set
or failure_class in {"sbar_observation_ownership", *V3_SBAR_FAILURE_CLASSES}
):
sbar = output.get("handoff_note_sbar") if isinstance(output.get("handoff_note_sbar"), dict) else {}
missing_sbar = [
part
for part in ("situation", "background", "assessment_observations_only", "handoff_request")
if len(str(sbar.get(part) or "").strip()) < 6
]
if missing_sbar:
issues.append("handoff_sbar_missing_required_parts")
return _dedupe(issues)
def v6_policy_issues(
output: dict[str, Any],
*,
failure_class: str,
expected_red_flag_rule_ids: list[str],
expected_candidate_pathway_card_ids: list[str],
structured_intake: dict[str, Any] | None = None,
rule_results: list[dict[str, Any]] | None = None,
retrieved_cards: list[dict[str, Any]] | None = None,
target_protocol_card_id: str = "",
dataset_version: str = "",
) -> list[str]:
"""Return v6 observation-ownership policy violations for assistant labels."""
issues = v5_policy_issues(
output,
failure_class=failure_class,
expected_red_flag_rule_ids=expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=expected_candidate_pathway_card_ids,
structured_intake=structured_intake,
rule_results=rule_results,
retrieved_cards=retrieved_cards,
target_protocol_card_id=target_protocol_card_id,
dataset_version=dataset_version,
)
missing = _string_list(output.get("missing_info_to_collect"))
next_observations = _string_list(output.get("next_observations_to_collect"))
if missing and len(missing) > 3 and _normalized_list(missing) == _normalized_list(next_observations):
issues.append("duplicate_long_missing_and_next_observations")
for item in missing + next_observations:
if _has_v6_harness_metadata_cue(item):
issues.append(f"harness_metadata_observation:{_safe_counter_key(item)}")
if _observation_text_has_unsafe_instruction(item):
issues.append(f"unsafe_observation_instruction:{_safe_counter_key(item)}")
selected_ids = _string_list(output.get(TRACE_ONLY_REQUIRED_OBSERVATION_IDS_KEY))
required_selected_ids = required_selected_observation_ids_for_version(
source_card_ids=_string_list(output.get("source_cards")),
retrieved_cards=retrieved_cards or [],
dataset_version=dataset_version,
target_protocol_card_id=target_protocol_card_id,
failure_class=failure_class,
)
if required_selected_ids and not selected_ids:
issues.append("selected_required_observation_ids_missing")
invalid_ids = sorted(set(selected_ids) - {str(target.get("id")) for target in required_observation_targets(retrieved_cards or [])})
if invalid_ids:
issues.append(f"selected_required_observation_ids_invalid:{','.join(invalid_ids)}")
missing_required = sorted(set(required_selected_ids) - set(selected_ids))
if missing_required:
issues.append(f"selected_required_observation_ids_missing_required:{','.join(missing_required)}")
visible_text = "\n".join(missing + next_observations)
visible_tokens = set(re.findall(r"[a-z0-9]+", visible_text.lower()))
targets_by_id = {str(target.get("id")): target for target in required_observation_targets(retrieved_cards or [])}
for selected_id in selected_ids:
target = targets_by_id.get(selected_id)
if not target:
continue
if not _required_observation_target_visible(
target,
visible_text,
visible_tokens,
structured_intake=structured_intake,
):
issues.append(f"selected_required_observation_id_not_visible:{selected_id}")
if uses_v10_perfect_eval_policy(dataset_version):
issues.extend(
_v10_dual_field_observation_issues(
output,
required_selected_ids=required_selected_ids,
retrieved_cards=retrieved_cards or [],
structured_intake=structured_intake,
)
)
return _dedupe(issues)
def _v10_dual_field_observation_issues(
output: dict[str, Any],
*,
required_selected_ids: list[str],
retrieved_cards: list[dict[str, Any]],
structured_intake: dict[str, Any] | None = None,
) -> list[str]:
"""Require v10 labels to own selected observations in both observation fields."""
issues: list[str] = []
targets_by_id = {str(target.get("id")): target for target in required_observation_targets(retrieved_cards)}
field_values = {
"missing_info_to_collect": "\n".join(_string_list(output.get("missing_info_to_collect"))),
"next_observations_to_collect": "\n".join(_string_list(output.get("next_observations_to_collect"))),
}
for field, text in field_values.items():
tokens = set(re.findall(r"[a-z0-9]+", text.lower()))
for required_id in required_selected_ids:
target = targets_by_id.get(required_id)
if not target:
continue
if not _required_observation_target_visible(
target,
text,
tokens,
structured_intake=structured_intake,
):
issues.append(f"v10_{field}_missing_required:{required_id}")
return _dedupe(issues)
def v7_source_card_closure_issues(
output: dict[str, Any],
*,
target_protocol_card_id: str | None = None,
) -> list[str]:
"""Return v7 source-card closure policy violations for assistant labels."""
issues: list[str] = []
source_cards = set(_string_list(output.get("source_cards")))
if target_protocol_card_id and target_protocol_card_id not in source_cards:
issues.append(f"missing_target_source_card:{target_protocol_card_id}")
if output.get("handoff_note_sbar") and SBAR_CARD_ID not in source_cards:
issues.append("missing_referral_sbar_source_card")
safety_text = json.dumps(
{
"safety_boundary": output.get("safety_boundary"),
"do_not_do": output.get("do_not_do"),
"responder_plain_language_script": output.get("responder_plain_language_script"),
},
sort_keys=True,
).lower()
safety_terms = (
"local protocol",
"do not diagnose",
"do not provide clinical orders",
"do not provide treatment instructions",
"safety boundary",
)
if any(term in safety_text for term in safety_terms) and SAFETY_CARD_ID not in source_cards:
issues.append("missing_safety_boundaries_source_card")
return _dedupe(issues)
def _policy_issues_for_prepared(output: dict[str, Any], prepared: PreparedCase) -> list[str]:
if uses_v7_source_card_policy(prepared.spec.dataset_version):
return _dedupe(
v6_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
structured_intake=prepared.spec.structured_intake,
rule_results=prepared.rule_results,
retrieved_cards=prepared.retrieved_cards,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
dataset_version=prepared.spec.dataset_version,
)
+ v7_source_card_closure_issues(
output,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
)
)
if uses_v6_observation_policy(prepared.spec.dataset_version):
return v6_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
structured_intake=prepared.spec.structured_intake,
rule_results=prepared.rule_results,
retrieved_cards=prepared.retrieved_cards,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
dataset_version=prepared.spec.dataset_version,
)
if uses_v5_focused_policy(prepared.spec.dataset_version):
return v5_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
structured_intake=prepared.spec.structured_intake,
rule_results=prepared.rule_results,
retrieved_cards=prepared.retrieved_cards,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
)
if uses_v3_field_workflow_policy(prepared.spec.dataset_version):
return v3_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
structured_intake=prepared.spec.structured_intake,
)
if prepared.spec.dataset_version == "figment_sft_v2":
return v2_policy_issues(
output,
failure_class=prepared.spec.failure_class,
expected_red_flag_rule_ids=prepared.expected_red_flag_rule_ids,
expected_candidate_pathway_card_ids=prepared.expected_candidate_pathway_card_ids,
)
return []
def _is_v3_generic_item(value: str) -> bool:
text = value.strip()
if not text:
return False
return any(pattern.search(text) for pattern in V3_GENERIC_OUTPUT_PATTERNS)
def _is_v5_generic_observation_item(value: str) -> bool:
text = value.strip()
if not text:
return False
return any(pattern.search(text) for pattern in V5_GENERIC_OBSERVATION_PATTERNS)
def _normalized_list(values: list[str]) -> list[str]:
return [_normalize_text(value) for value in values if _normalize_text(value)]
def _has_v6_harness_metadata_cue(value: str) -> bool:
normalized = _normalize_text(value)
return any(_normalize_text(cue) in normalized for cue in V6_HARNESS_METADATA_OBSERVATION_CUES)
def _observation_text_has_unsafe_instruction(value: str) -> bool:
return any(pattern.search(value) for pattern in V2_FORBIDDEN_LEXICAL_PATTERNS.values())
def _required_observation_target_visible(
target: dict[str, Any],
text: str,
tokens: set[str],
*,
structured_intake: dict[str, Any] | None = None,
) -> bool:
cue_tokens = {str(token) for token in target.get("cue_tokens", []) if str(token)}
if cue_tokens and cue_tokens <= tokens:
return True
display_text = str(target.get("display_text") or "")
normalized_display = _normalize_text(display_text)
normalized_text = _normalize_text(text)
if normalized_display and normalized_display in normalized_text:
return True
resource_aware_text = _v3_resource_aware_cue_text(display_text, structured_intake or {})
normalized_resource_aware = _normalize_text(resource_aware_text)
return bool(normalized_resource_aware and normalized_resource_aware in normalized_text)
def _resource_unavailable(resource_text: str, phrases: tuple[str, ...]) -> bool:
return any(phrase in resource_text for phrase in phrases)
def _asks_for_unavailable_pulse_ox(output_text: str) -> bool:
if "unavailable" in output_text and any(
phrase in output_text for phrase in ("pulse ox", "pulse oximeter", "oxygen saturation", "spo2")
):
return False
return any(phrase in output_text for phrase in ("pulse ox", "pulse oximeter", "oxygen saturation", "spo2", "repeat vitals"))
def _asks_for_unavailable_bp(output_text: str) -> bool:
if "unavailable" in output_text and any(phrase in output_text for phrase in ("blood pressure", "bp cuff", "bp")):
return False
return any(phrase in output_text for phrase in ("blood pressure", "bp cuff", "repeat vitals"))
def eval_record_for_output(
prepared: PreparedCase,
output: dict[str, Any],
validation: dict[str, Any],
) -> dict[str, Any]:
cue_buckets = bucket_expected_observation_cues(prepared.expected_missing_observations)
harness_evidence = build_harness_evidence(
confirmed_intake=prepared.spec.structured_intake,
retrieved_card_ids=prepared.retrieved_ids,
rule_results=prepared.rule_results,
urgency_floor=prepared.urgency_floor,
validator_result=validation,
final_output=output,
)
return {
"case_id": prepared.spec.case_id,
"structured_intake": prepared.spec.structured_intake,
"target_protocol_card_id": prepared.spec.target_protocol_card_id,
"expected_min_protocol_urgency": prepared.urgency_floor,
"expected_red_flag_rule_ids": prepared.expected_red_flag_rule_ids,
"expected_source_card_ids": prepared.expected_source_card_ids,
"expected_candidate_pathway_card_ids": prepared.expected_candidate_pathway_card_ids,
"expected_missing_observations": prepared.expected_missing_observations,
"expected_model_observation_cues": cue_buckets["model"],
"expected_handoff_cues": cue_buckets["handoff"],
"expected_harness_evidence_cues": cue_buckets["harness"],
"forbidden_behavior": forbidden_behavior_for_version(prepared.spec.dataset_version),
"actual_red_flag_rule_ids": [str(rule["rule_id"]) for rule in prepared.rule_results],
"actual_protocol_urgency": output.get("protocol_urgency"),
"actual_source_card_ids": _string_list(output.get("source_cards")),
"actual_candidate_pathway_card_ids": _candidate_ids(output.get("candidate_protocol_pathways")),
"retrieved_card_ids": prepared.retrieved_ids,
"harness_evidence": harness_evidence,
"final_output": output,
"final_validation": validation,
}
def build_sft_row(
*,
prepared: PreparedCase,
result: CandidateResult,
teacher_model_id: str,
candidate_total: int,
candidate_passed: int,
) -> dict[str, Any]:
workflow_category = str(prepared.spec.structured_intake.get("workflow_category") or "")
cue_buckets = bucket_expected_observation_cues(prepared.expected_missing_observations)
metadata = {
"teacher_model_id": teacher_model_id,
"critic_model_id": teacher_model_id,
"teacher_label_mode": "streamed_ultra_semantic_notes_harness_prompt",
"teacher_base_url_env": _endpoint_env_name(teacher_model_id),
"teacher_api_key_env": _api_key_env_name(teacher_model_id),
"failure_class": prepared.spec.failure_class,
"dataset_version": prepared.spec.dataset_version,
"expected_action": {
"target_card": prepared.spec.target_protocol_card_id,
"source_cards": prepared.expected_source_card_ids,
"candidate_pathway_card_ids": prepared.expected_candidate_pathway_card_ids,
"required_observation_cues": prepared.expected_missing_observations,
"model_observation_cues": cue_buckets["model"],
"handoff_cues": cue_buckets["handoff"],
"harness_evidence_cues": cue_buckets["harness"],
"red_flag_rule_ids": prepared.expected_red_flag_rule_ids,
"min_protocol_urgency": prepared.urgency_floor,
},
"reward_components": result.reward_components,
"pass_rate_total": candidate_total,
"pass_rate_passed": candidate_passed,
"dedupe_hash": dedupe_hash(prepared),
"input_hash": stable_hash(prepared.spec.structured_intake),
"prompt_hash": stable_hash(prepared.prompt),
"prompt_template_hash": prepared.prompt_hash,
"raw_teacher_output_hash": result.raw_output_hash,
"deterministic_scaffold_patched_fields": result.patched_fields,
"filled_required_observation_ids": result.filled_required_observation_ids,
"model_selected_required_observation_ids": result.model_selected_required_observation_ids,
"invalid_selected_required_observation_ids": result.invalid_selected_required_observation_ids,
"stripped_trace_only_fields": result.stripped_trace_only_fields,
"validation_result": result.validation,
"expected_label_score": result.expected_label_score,
"retrieved_card_ids": prepared.retrieved_ids,
"recipe_sources": [
"nvidia/Nemotron-Post-Training-Dataset-v2",
"nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1",
"nvidia/Nemotron-CC-v2",
],
"validator_passed": True,
"license_review": "synthetic_figment_row_no_nvidia_rows_copied",
"phi_status": "synthetic_deidentified_no_phi",
"generated_at": datetime.now(UTC).isoformat(),
}
if workflow_category:
metadata.update(
{
"workflow_category": workflow_category,
"field_workflow_goal": prepared.spec.structured_intake.get("field_workflow_goal"),
"workflow_priority_observations": prepared.expected_missing_observations[:5],
"v3_workflow_validator_version": 1,
"anti_overfit_policy": {
"locked_eval_copying_allowed": False,
"holdout_copying_allowed": False,
"primary_success_surface": "field_workflow_holdout_v1",
},
}
)
if uses_v5_focused_policy(prepared.spec.dataset_version):
must_include_selected = v5_required_selected_observation_ids(
source_card_ids=_string_list(result.output.get("source_cards")),
retrieved_cards=prepared.retrieved_cards,
)
metadata.update(
{
"training_focus": prepared.spec.failure_class,
"v5_training_policy_version": 1,
"excluded_eval_case_ids": list(V5_EXCLUDED_EVAL_CASE_IDS),
"must_include_source_cards": _v5_must_include_source_cards(prepared),
"must_include_selected_required_observation_ids": must_include_selected,
}
)
if uses_v6_observation_policy(prepared.spec.dataset_version):
must_include_selected = required_selected_observation_ids_for_version(
source_card_ids=_string_list(result.output.get("source_cards")),
retrieved_cards=prepared.retrieved_cards,
dataset_version=prepared.spec.dataset_version,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
failure_class=prepared.spec.failure_class,
)
metadata.update(
{
"training_focus": prepared.spec.failure_class,
"v6_training_policy_version": 1,
"must_include_source_cards": _v5_must_include_source_cards(prepared),
"required_observation_targets": _v6_required_observation_targets(prepared),
"must_include_selected_required_observation_ids": must_include_selected,
"harness_metadata_cues_not_observations": list(V6_HARNESS_METADATA_OBSERVATION_CUES),
"observation_field_contract": {
"missing_info_to_collect": "broader still-needed clinical observations",
"next_observations_to_collect": "prioritized next 3-5 clinical observations",
"selected_required_observation_ids": "trace-only ids; runtime strips from user-visible output",
},
}
)
if uses_v7_source_card_policy(prepared.spec.dataset_version):
metadata.update(
{
"v7_training_policy_version": 1,
"source_card_closure_contract": {
"target_protocol_card_id": prepared.spec.target_protocol_card_id,
"must_include_source_cards": _v5_must_include_source_cards(prepared),
"safety_card_required_when_safety_text_present": SAFETY_CARD_ID,
"sbar_card_required_when_handoff_present": SBAR_CARD_ID,
},
}
)
return {
"case_id": prepared.spec.case_id,
"uuid": prepared.spec.case_id,
"license": "synthetic internal training data",
"generator": teacher_model_id,
"version": prepared.spec.dataset_version,
"category": prepared.spec.failure_class,
"reasoning": "off",
"messages": [
{"role": "user", "content": prepared.prompt},
{"role": "assistant", "content": json.dumps(result.output, sort_keys=True)},
],
"tags": prepared.spec.tags,
"metadata": metadata,
}
def case_spec_record(prepared: PreparedCase) -> dict[str, Any]:
cue_buckets = bucket_expected_observation_cues(prepared.expected_missing_observations)
record = {
"case_id": prepared.spec.case_id,
"dataset_version": prepared.spec.dataset_version,
"failure_class": prepared.spec.failure_class,
"target_protocol_card_id": prepared.spec.target_protocol_card_id,
"structured_intake": prepared.spec.structured_intake,
"expected_red_flag_rule_ids": prepared.expected_red_flag_rule_ids,
"expected_min_protocol_urgency": prepared.urgency_floor,
"expected_source_card_ids": prepared.expected_source_card_ids,
"expected_candidate_pathway_card_ids": prepared.expected_candidate_pathway_card_ids,
"expected_missing_observations": prepared.expected_missing_observations,
"expected_model_observation_cues": cue_buckets["model"],
"expected_handoff_cues": cue_buckets["handoff"],
"expected_harness_evidence_cues": cue_buckets["harness"],
"retrieved_card_ids": prepared.retrieved_ids,
"tags": prepared.spec.tags,
}
workflow_category = str(prepared.spec.structured_intake.get("workflow_category") or "")
if workflow_category:
record.update(
{
"workflow_category": workflow_category,
"field_workflow_goal": prepared.spec.structured_intake.get("field_workflow_goal"),
"workflow_priority_observations": prepared.expected_missing_observations[:5],
"field_workflow_holdout_relevant": True,
}
)
if uses_v6_observation_policy(prepared.spec.dataset_version):
record.update(
{
"required_observation_targets": _v6_required_observation_targets(prepared),
"must_include_selected_required_observation_ids": required_selected_observation_ids_for_version(
source_card_ids=prepared.expected_source_card_ids,
retrieved_cards=prepared.retrieved_cards,
dataset_version=prepared.spec.dataset_version,
target_protocol_card_id=prepared.spec.target_protocol_card_id,
failure_class=prepared.spec.failure_class,
),
"harness_metadata_cues_not_observations": list(V6_HARNESS_METADATA_OBSERVATION_CUES),
}
)
if uses_v7_source_card_policy(prepared.spec.dataset_version):
record.update(
{
"source_card_closure_contract": {
"target_protocol_card_id": prepared.spec.target_protocol_card_id,
"must_include_source_cards": _v5_must_include_source_cards(prepared),
},
}
)
return record
def build_manifest(
*,
output_path: Path,
case_specs_path: Path,
dataset_version: str,
rows: list[dict[str, Any]],
started_at: datetime,
teacher_model_id: str,
dry_run: bool,
attempts: int,
start_index: int,
index_stride: int,
counters: Counter[str],
candidate_totals: Counter[str],
rejection_reasons: Counter[str],
events: list[dict[str, Any]],
exclusion_paths: list[Path] | None = None,
exclusion_signature_count: int = 0,
) -> dict[str, Any]:
finished_at = datetime.now(UTC)
return {
"dataset_version": dataset_version,
"row_count": len(rows),
"output_path": str(output_path),
"case_specs_path": str(case_specs_path),
"output_sha256": _file_sha256(output_path) if output_path.exists() else None,
"case_specs_sha256": _file_sha256(case_specs_path) if case_specs_path.exists() else None,
"started_at": started_at.isoformat(),
"finished_at": finished_at.isoformat(),
"elapsed_seconds": round((finished_at - started_at).total_seconds(), 3),
"teacher_model_id": teacher_model_id,
"teacher_endpoint_env": _endpoint_env_name(teacher_model_id),
"teacher_api_key_env": _api_key_env_name(teacher_model_id),
"dry_run": dry_run,
"attempts": attempts,
"start_index": start_index,
"index_stride": index_stride,
"accepted_by_failure_class": {
key: value for key, value in sorted(counters.items()) if not key.startswith("tag:")
},
"accepted_by_tag": {
key[4:]: value for key, value in sorted(counters.items()) if key.startswith("tag:")
},
"candidate_totals": dict(candidate_totals),
"rejection_reasons": dict(rejection_reasons),
"anti_overfit_exclusions": {
"enabled": bool(exclusion_paths),
"eval_paths": [str(path) for path in exclusion_paths or []],
"signature_count": exclusion_signature_count,
"policies": [
"exact clinical-intake hash rejection",
"same target/workflow high-token-overlap rejection",
],
},
"source_recipe_links": [
"nvidia/Nemotron-Post-Training-Dataset-v2",
"nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1",
"nvidia/Nemotron-CC-v2",
],
"license_phi_assertions": {
"synthetic_only": True,
"no_phi": True,
"no_locked_eval_rows_copied": True,
"no_nvidia_dataset_rows_copied": True,
"requires_license_review_before_distribution": True,
},
"prompt_template_hash": stable_hash(SYSTEM_PROMPT),
"event_sample": events[-50:],
}
def _load_existing_rows(path: Path) -> list[dict[str, Any]]:
if not path.exists():
return []
rows = []
for line in path.read_text(encoding="utf-8").splitlines():
if line.strip():
rows.append(json.loads(line))
return rows
def _rejection_key(result: CandidateResult) -> str:
if result.validation.get("passed") is not True:
failures = result.validation.get("failures") or ["validation_failed"]
return _safe_counter_key(str(failures[0]))
if result.expected_label_score.get("all_expected_labels_passed") is not True:
for key, value in result.expected_label_score.items():
if value is False:
return f"expected_label_{key}"
return "expected_label_failed"
failed_rewards = [key for key, value in result.reward_components.items() if not value]
return f"reward_{failed_rewards[0]}" if failed_rewards else "unknown"
def _patch_fields(before: dict[str, Any], after: dict[str, Any]) -> list[str]:
return sorted(field for field in REQUIRED_JSON_SKELETON if before.get(field) != after.get(field))
def _candidate_ids(value: Any) -> list[str]:
if not isinstance(value, list):
return []
ids: list[str] = []
for item in value:
card_id = item.get("card_id") if isinstance(item, dict) else item
if card_id:
ids.append(str(card_id))
return ids
def _string_list(value: Any) -> list[str]:
if isinstance(value, list):
return [str(item) for item in value if str(item)]
if isinstance(value, str) and value.strip():
return [value.strip()]
return []
def _dedupe(values: list[str]) -> list[str]:
out: list[str] = []
for value in values:
text = str(value).strip()
if text and text not in out:
out.append(text)
return out
def _pick(values: list[str], index: int) -> str:
return values[index % len(values)]
def _tag_for_card(card_id: str) -> str:
return card_id.lower().replace("-v1", "").replace("-", "_")
def dedupe_hash(prepared: PreparedCase) -> str:
payload = {
"normalized_intake": _normalize_text(json.dumps(prepared.spec.structured_intake, sort_keys=True)),
"target": prepared.spec.target_protocol_card_id,
"expected_source": prepared.expected_source_card_ids,
"expected_missing": prepared.expected_missing_observations,
}
return "sha256:" + hashlib.sha256(json.dumps(payload, sort_keys=True).encode("utf-8")).hexdigest()
def _normalize_text(value: str) -> str:
return " ".join(re.findall(r"[a-z0-9]+", value.lower()))
def _endpoint_env_name(teacher_model_id: str | None = None) -> str:
# Record only the variable name, never the resolved endpoint or secret.
if teacher_model_id and _openrouter_config_for_teacher_model(teacher_model_id):
return "OPENROUTER_BASE_URL"
if os.getenv("OMNI_ENDPOINT_URL", "").strip():
return "OMNI_ENDPOINT_URL"
if os.getenv("HF_ENDPOINT_URL", "").strip():
return "HF_ENDPOINT_URL"
if os.getenv("NVIDIA_BASE_URL", "").strip():
return "NVIDIA_BASE_URL"
return f"NVIDIA_BASE_URL(default:{NVIDIA_API_BASE_URL})"
def _api_key_env_name(teacher_model_id: str | None = None) -> str:
# Record only the variable name, never the resolved secret.
if teacher_model_id and _openrouter_config_for_teacher_model(teacher_model_id):
return "OPENROUTER_API_KEY"
if os.getenv("OMNI_ENDPOINT_URL", "").strip() or os.getenv("HF_ENDPOINT_URL", "").strip():
return "HF_TOKEN" if os.getenv("HF_TOKEN", "").strip() else ""
if os.getenv("NVIDIA_API_KEY", "").strip():
return "NVIDIA_API_KEY"
return ""
def _file_sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as file:
for chunk in iter(lambda: file.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _print_progress_event(event: dict[str, Any]) -> None:
print(json.dumps(event, sort_keys=True), flush=True)
def _safe_counter_key(value: str) -> str:
return re.sub(r"[^a-z0-9_.:-]+", "_", value.lower()).strip("_")[:120] or "unknown"
def _safe_error_text(value: str) -> str:
text = re.sub(r"(?i)bearer\s+[^\s]+", "Bearer [redacted]", value)
text = re.sub(r"(?i)(api_key|token|authorization)=([^&\s]+)", r"\1=[redacted]", text)
return text[:500]
if __name__ == "__main__":
raise SystemExit(main())