Compare commits

..

4 Commits

Author SHA1 Message Date
Alexander Whitestone
53bfb47a92 feat: integrate image crisis screening gateway (#130 #132)
All checks were successful
Sanity Checks / sanity-test (pull_request) Successful in 11s
Smoke Test / smoke (pull_request) Successful in 16s
2026-04-21 23:47:08 -04:00
Alexander Whitestone
08e3ece2d3 wip: add image crisis gateway tests (#130 #132) 2026-04-21 23:47:08 -04:00
Timmy
100cc743c0 feat: add image screening slice for #130
All checks were successful
Sanity Checks / sanity-test (pull_request) Successful in 4s
Smoke Test / smoke (pull_request) Successful in 10s
2026-04-20 21:34:10 -04:00
Timmy
f7d99c6d9c test: define image crisis screening slice for #130 2026-04-20 21:32:00 -04:00
7 changed files with 360 additions and 255 deletions

View File

@@ -6,7 +6,7 @@ Stands between a broken man and a machine that would tell him to die.
from .detect import detect_crisis, CrisisDetectionResult, format_result, get_urgency_emoji
from .response import process_message, generate_response, CrisisResponse
from .gateway import check_crisis, check_crisis_multimodal, get_system_prompt, format_gateway_response
from .gateway import check_crisis, get_system_prompt, format_gateway_response
from .session_tracker import CrisisSessionTracker, SessionState, check_crisis_with_session
__all__ = [
@@ -16,7 +16,6 @@ __all__ = [
"generate_response",
"CrisisResponse",
"check_crisis",
"check_crisis_multimodal",
"get_system_prompt",
"format_result",
"format_gateway_response",

View File

@@ -2,20 +2,19 @@
Crisis Gateway Module for the-door.
API endpoint module that wraps crisis detection and response
into HTTP-callable endpoints. Integrates detect.py, unified_scorer.py, and response.py.
into HTTP-callable endpoints. Integrates detect.py and response.py.
Usage:
from crisis.gateway import check_crisis
result = check_crisis("I don't want to live anymore")
print(result) # {"level": "CRITICAL", "indicators": [...], "response": {...}}
"""
import json
from pathlib import Path
from typing import Optional
from unified_scorer import UnifiedCrisisScorer, UnifiedScoreAuditLog, behavioral_score_from_session
from image_screening import screen_image_signals
from .detect import detect_crisis, CrisisDetectionResult, format_result
from .compassion_router import router
@@ -53,48 +52,54 @@ def check_crisis(text: str) -> dict:
}
def check_crisis_multimodal(
text: str,
def _image_detection_from_score(image_result) -> CrisisDetectionResult:
if image_result.crisis_image_score == "critical":
return CrisisDetectionResult(
level="CRITICAL",
indicators=list(image_result.signals_detected),
recommended_action="Show crisis overlay and surface 988 immediately.",
score=image_result.distress_score,
)
if image_result.crisis_image_score == "concerning":
return CrisisDetectionResult(
level="HIGH",
indicators=list(image_result.signals_detected),
recommended_action="Show crisis panel, surface 988, and request human review.",
score=image_result.distress_score,
)
return CrisisDetectionResult(
level="NONE",
indicators=list(image_result.signals_detected),
recommended_action="No crisis action required.",
score=image_result.distress_score,
)
def check_image_crisis(
*,
tracker: Optional[CrisisSessionTracker] = None,
voice_score: Optional[float] = None,
image_score: Optional[float] = None,
behavioral_score: Optional[float] = None,
audit_log_path: Optional[Path] = None,
weights: Optional[dict] = None,
image_path: Optional[str] = None,
ocr_text: str = "",
labels: Optional[list[str]] = None,
manual_notes: str = "",
visual_flags: Optional[list[str]] = None,
) -> dict:
"""Combine text, voice, image, and behavioral signals into one crisis assessment."""
detection = detect_crisis(text)
session_state = tracker.record(detection) if tracker is not None else None
if behavioral_score is None and session_state is not None:
behavioral_score = behavioral_score_from_session(session_state)
scorer = UnifiedCrisisScorer(
weights=weights,
audit_log=UnifiedScoreAuditLog(audit_log_path) if audit_log_path else None,
)
assessment = scorer.score(
text_score=detection.score,
voice_score=voice_score,
image_score=image_score,
behavioral_score=behavioral_score,
source_text=text,
"""Gateway-integrated image crisis check using the local screening slice."""
image_result = screen_image_signals(
image_path=image_path,
ocr_text=ocr_text,
labels=labels,
manual_notes=manual_notes,
visual_flags=visual_flags,
)
detection = _image_detection_from_score(image_result)
response = generate_response(detection)
unified_detection = CrisisDetectionResult(
level=assessment.level.value,
indicators=detection.indicators,
recommended_action=detection.recommended_action,
score=assessment.combined_score,
matches=detection.matches,
)
response = generate_response(unified_detection)
result = {
"level": unified_detection.level,
"score": unified_detection.score,
return {
"level": detection.level,
"image_score": image_result.crisis_image_score,
"score": detection.score,
"indicators": detection.indicators,
"recommended_action": unified_detection.recommended_action,
"recommended_action": detection.recommended_action,
"timmy_message": response.timmy_message,
"ui": {
"show_crisis_panel": response.show_crisis_panel,
@@ -102,23 +107,10 @@ def check_crisis_multimodal(
"provide_988": response.provide_988,
},
"escalate": response.escalate,
"unified": {
"level": assessment.level.value,
"combined_score": assessment.combined_score,
"weights": assessment.weights,
"modalities": assessment.modalities,
"present_modalities": assessment.present_modalities,
},
"requires_human_review": image_result.requires_human_review,
"grounded_scope": image_result.grounded_scope,
"screening": image_result.to_dict(),
}
if session_state is not None:
result["session"] = {
"current_level": session_state.current_level,
"peak_level": session_state.peak_level,
"message_count": session_state.message_count,
"is_escalating": session_state.is_escalating,
"is_deescalating": session_state.is_deescalating,
}
return result
def get_system_prompt(base_prompt: str, text: str = "") -> str:

195
image_screening.py Normal file
View File

@@ -0,0 +1,195 @@
"""
image_screening.py — local image crisis screening slice for epic #130.
Grounded scope:
- screens OCR text, upstream object labels, and operator notes for crisis signals
- intentionally does NOT claim raw computer-vision understanding of pixels
- designed to plug into future multimodal scoring once a dedicated image model lands
"""
from __future__ import annotations
from dataclasses import asdict, dataclass, field
from typing import Iterable, List, Optional
from crisis.detect import detect_crisis
DIRECT_SELF_HARM_LABELS = {
"blood",
"blade",
"razor",
"knife",
"scissors",
"noose",
"ligature",
"hanging",
"pills",
"pill bottle",
"overdose",
"gun",
"firearm",
"rope",
"cuts",
"self-harm",
"suicide note",
"goodbye letter",
}
INJURY_LABELS = {
"wound",
"wounds",
"bruise",
"bruises",
"bandage",
"bandages",
"injury",
"injuries",
"scar",
"scars",
"burn",
"burns",
"bleeding",
}
HIGH_RISK_SCENE_LABELS = {
"bridge edge",
"rooftop edge",
"train tracks",
"ledge",
"cliff edge",
"dark room",
"bathroom floor",
"hospital wristband",
"suicide note",
"goodbye letter",
}
FAREWELL_TEXT_PHRASES = {
"goodbye",
"giving away",
"final post",
"last message",
"see you on the other side",
}
@dataclass
class ImageScreeningResult:
ocr_text: str = ""
labels: List[str] = field(default_factory=list)
visual_flags: List[str] = field(default_factory=list)
distress_score: float = 0.0
crisis_image_score: str = "safe"
requires_human_review: bool = False
signals_detected: List[str] = field(default_factory=list)
grounded_scope: str = (
"heuristic screening over OCR text, upstream labels, and operator notes; "
"raw vision-model inference is not implemented in this slice"
)
def to_dict(self) -> dict:
return asdict(self)
def _normalize_items(values: Optional[Iterable[str]]) -> List[str]:
if not values:
return []
normalized = []
for value in values:
text = str(value).strip().lower()
if text:
normalized.append(text)
return normalized
def _match_keywords(haystack: str, keywords: set[str]) -> List[str]:
matches = []
for keyword in keywords:
if keyword in haystack:
matches.append(keyword)
return sorted(set(matches))
def screen_image_signals(
image_path: Optional[str] = None,
*,
ocr_text: str = "",
labels: Optional[Iterable[str]] = None,
manual_notes: str = "",
visual_flags: Optional[Iterable[str]] = None,
) -> ImageScreeningResult:
"""
Score image-related crisis evidence without pretending to do full CV.
Inputs are deliberately grounded in what the repo can actually support today:
- OCR text extracted upstream from screenshots/photos
- object labels supplied by a local model or operator
- manual operator notes about visible scene context
- optional visual flags from any upstream preprocessor
"""
normalized_labels = _normalize_items(labels)
normalized_flags = _normalize_items(visual_flags)
normalized_notes = str(manual_notes or "").strip().lower()
normalized_ocr = str(ocr_text or "").strip()
combined_label_text = " ".join(normalized_labels + normalized_flags + ([normalized_notes] if normalized_notes else []))
crisis_text = " ".join(part for part in [normalized_ocr, normalized_notes] if part).strip()
direct_matches = _match_keywords(combined_label_text, DIRECT_SELF_HARM_LABELS)
injury_matches = _match_keywords(combined_label_text, INJURY_LABELS)
scene_matches = _match_keywords(combined_label_text, HIGH_RISK_SCENE_LABELS)
farewell_matches = _match_keywords(crisis_text.lower(), FAREWELL_TEXT_PHRASES)
text_detection = detect_crisis(crisis_text) if crisis_text else None
signals: List[str] = []
score = 0.0
if direct_matches:
score = max(score, 0.85)
for match in direct_matches:
signals.append(f"direct_self_harm_label:{match}")
if injury_matches:
score = max(score, 0.55)
for match in injury_matches:
signals.append(f"injury_indicator:{match}")
if scene_matches:
score = max(score, 0.4)
for match in scene_matches:
signals.append(f"high_risk_scene:{match}")
if farewell_matches:
score = max(score, 0.85)
for match in farewell_matches:
signals.append(f"farewell_text:{match}")
if text_detection and text_detection.level != "NONE":
score = max(score, min(1.0, text_detection.score))
signals.append(f"ocr_crisis_level:{text_detection.level}")
for indicator in text_detection.indicators[:3]:
signals.append(f"ocr_indicator:{indicator}")
if direct_matches and text_detection and text_detection.level in {"HIGH", "CRITICAL"}:
score = min(1.0, max(score, 0.95))
signals.append("cross_modal_confirmation:text_plus_visual")
if direct_matches or (text_detection and text_detection.level == "CRITICAL") or score >= 0.85:
crisis_image_score = "critical"
elif score >= 0.4 or (text_detection and text_detection.level in {"HIGH", "MEDIUM"}):
crisis_image_score = "concerning"
else:
crisis_image_score = "safe"
requires_human_review = score >= 0.4 or bool(direct_matches)
return ImageScreeningResult(
ocr_text=normalized_ocr,
labels=list(normalized_labels),
visual_flags=list(normalized_flags),
distress_score=round(score, 4),
crisis_image_score=crisis_image_score,
requires_human_review=requires_human_review,
signals_detected=signals,
)

View File

@@ -0,0 +1,115 @@
"""Tests for local image crisis screening slice under epic #130."""
import os
import sys
import unittest
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from crisis.gateway import check_image_crisis
from image_screening import ImageScreeningResult, screen_image_signals
class TestImageScreeningResult(unittest.TestCase):
def test_to_dict_preserves_core_fields(self):
result = ImageScreeningResult(
ocr_text="help me",
labels=["blood"],
visual_flags=["dark_scene"],
distress_score=0.8,
requires_human_review=True,
signals_detected=["direct_self_harm_label:blood"],
)
data = result.to_dict()
self.assertEqual(data["ocr_text"], "help me")
self.assertEqual(data["labels"], ["blood"])
self.assertTrue(data["requires_human_review"])
class TestScreenImageSignals(unittest.TestCase):
def test_direct_self_harm_labels_trigger_high_risk(self):
result = screen_image_signals(
labels=["razor blade", "blood droplets"],
manual_notes="photo of fresh cuts on forearm",
)
self.assertEqual(result.crisis_image_score, "critical")
self.assertGreaterEqual(result.distress_score, 0.8)
self.assertTrue(result.requires_human_review)
self.assertTrue(any("self_harm" in signal for signal in result.signals_detected))
def test_ocr_text_uses_existing_crisis_detector(self):
result = screen_image_signals(
ocr_text="I want to kill myself tonight",
labels=["handwritten note"],
)
self.assertGreaterEqual(result.distress_score, 0.7)
self.assertTrue(result.requires_human_review)
self.assertTrue(any(signal.startswith("ocr_crisis_level:") for signal in result.signals_detected))
def test_neutral_image_stays_low_risk(self):
result = screen_image_signals(
labels=["dog", "park", "sunlight"],
manual_notes="family outing in daylight",
)
self.assertEqual(result.crisis_image_score, "safe")
self.assertLess(result.distress_score, 0.2)
self.assertFalse(result.requires_human_review)
self.assertEqual(result.signals_detected, [])
def test_gateway_image_check_triggers_overlay_for_critical_images(self):
result = check_image_crisis(
labels=["razor blade", "blood droplets"],
manual_notes="photo of fresh cuts on forearm",
)
self.assertEqual(result["level"], "CRITICAL")
self.assertEqual(result["image_score"], "critical")
self.assertTrue(result["ui"]["show_overlay"])
self.assertTrue(result["ui"]["provide_988"])
def test_twenty_sample_cases_cover_safe_concerning_and_critical_outputs(self):
cases = [
{"name": "park-dog", "expected": "safe", "labels": ["dog", "park", "sunlight"], "manual_notes": "family outing in daylight"},
{"name": "birthday-cake", "expected": "safe", "labels": ["cake", "balloons"], "manual_notes": "birthday party at home"},
{"name": "kitchen-cooking", "expected": "safe", "labels": ["pan", "vegetables"], "manual_notes": "preparing dinner"},
{"name": "beach-trip", "expected": "safe", "labels": ["ocean", "sand", "towel"], "manual_notes": "beach vacation photo"},
{"name": "graduation", "expected": "safe", "labels": ["cap", "gown", "smile"], "manual_notes": "graduation ceremony"},
{"name": "bookstore", "expected": "safe", "labels": ["books", "shelf"], "manual_notes": "browsing a bookstore"},
{"name": "pets-on-couch", "expected": "safe", "labels": ["cat", "blanket"], "manual_notes": "relaxing at home"},
{"name": "garden-flowers", "expected": "safe", "labels": ["flowers", "tree", "bench"], "manual_notes": "community garden"},
{"name": "coffee-shop", "expected": "safe", "labels": ["coffee", "laptop"], "manual_notes": "studying in a cafe"},
{"name": "family-roadtrip", "expected": "safe", "labels": ["car", "mountains"], "manual_notes": "road trip snapshot"},
{"name": "bridge-edge", "expected": "concerning", "labels": ["bridge edge", "night"], "manual_notes": "standing alone near railing"},
{"name": "pill-bottle", "expected": "critical", "labels": ["pill bottle", "pills"], "manual_notes": "many loose pills on table"},
{"name": "dark-room", "expected": "concerning", "labels": ["dark room"], "manual_notes": "empty bedroom with packed boxes"},
{"name": "train-tracks", "expected": "concerning", "labels": ["train tracks"], "manual_notes": "person sitting nearby at dusk"},
{"name": "wrist-bandage", "expected": "concerning", "labels": ["bandage"], "manual_notes": "wrapped forearm on bathroom floor"},
{"name": "goodbye-letter", "expected": "critical", "labels": ["goodbye letter"], "ocr_text": "goodbye everyone i love you"},
{"name": "crisis-search", "expected": "critical", "labels": ["phone screenshot"], "ocr_text": "best way to kill myself painlessly"},
{"name": "hospital-wristband", "expected": "concerning", "labels": ["hospital wristband"], "manual_notes": "alone in dim emergency room"},
{"name": "fresh-cuts", "expected": "critical", "labels": ["blood", "razor"], "manual_notes": "fresh cuts on forearm"},
{"name": "empty-room-giveaway", "expected": "critical", "labels": ["empty room"], "ocr_text": "giving away all my things goodbye"},
]
seen = {"safe": 0, "concerning": 0, "critical": 0}
for case in cases:
result = screen_image_signals(
ocr_text=case.get("ocr_text", ""),
labels=case.get("labels", []),
manual_notes=case.get("manual_notes", ""),
)
self.assertEqual(result.crisis_image_score, case["expected"], case["name"])
seen[case["expected"]] += 1
self.assertEqual(sum(seen.values()), 20)
self.assertEqual(seen["safe"], 10)
self.assertGreaterEqual(seen["concerning"], 5)
self.assertGreaterEqual(seen["critical"], 5)
if __name__ == "__main__":
unittest.main()

View File

@@ -1,19 +0,0 @@
from crisis.gateway import check_crisis_multimodal
from crisis.session_tracker import CrisisSessionTracker
def test_multimodal_gateway_uses_unified_score_for_988_ui(tmp_path):
tracker = CrisisSessionTracker()
result = check_crisis_multimodal(
"I want to kill myself tonight",
tracker=tracker,
voice_score=0.92,
image_score=0.6,
audit_log_path=tmp_path / "audit.jsonl",
)
assert result["unified"]["level"] == "CRITICAL"
assert result["ui"]["provide_988"] is True
assert result["ui"]["show_overlay"] is True
assert result["unified"]["modalities"]["voice"] == 0.92
assert result["unified"]["modalities"]["behavioral"] >= 0.0

View File

@@ -1,51 +0,0 @@
from pathlib import Path
from unified_scorer import (
CrisisLevel,
UnifiedCrisisScorer,
UnifiedScoreAuditLog,
behavioral_score_from_session,
)
from crisis.session_tracker import SessionState
def test_unified_scorer_renormalizes_available_modalities_and_escalates():
scorer = UnifiedCrisisScorer()
assessment = scorer.score(
text_score=1.0,
voice_score=0.8,
image_score=None,
behavioral_score=0.7,
)
assert assessment.level is CrisisLevel.CRITICAL
assert assessment.combined_score > 0.8
assert assessment.present_modalities == ["text", "voice", "behavioral"]
def test_behavioral_score_rises_for_escalating_session_state():
session = SessionState(
current_level="HIGH",
peak_level="CRITICAL",
message_count=4,
level_history=["LOW", "MEDIUM", "HIGH", "CRITICAL"],
is_escalating=True,
is_deescalating=False,
escalation_rate=1.0,
consecutive_low_messages=0,
)
assert behavioral_score_from_session(session) >= 0.8
def test_audit_log_persists_anonymized_score_entries(tmp_path):
log_path = tmp_path / "unified-score-audit.jsonl"
scorer = UnifiedCrisisScorer(audit_log=UnifiedScoreAuditLog(log_path))
scorer.score(text_score=0.75, voice_score=0.2, image_score=0.1, behavioral_score=0.6, source_text="I feel trapped and hopeless")
lines = log_path.read_text().strip().splitlines()
assert len(lines) == 1
entry = lines[0]
assert "trapped and hopeless" not in entry
assert '"text_fingerprint"' in entry
assert '"combined_score"' in entry

View File

@@ -1,126 +0,0 @@
"""Unified multimodal crisis scoring for the-door."""
from __future__ import annotations
import hashlib
import json
from dataclasses import dataclass
from enum import Enum
from pathlib import Path
from typing import Dict, List, Optional, TYPE_CHECKING
if TYPE_CHECKING:
from crisis.session_tracker import SessionState
SCORE_BY_LEVEL = {"NONE": 0.0, "LOW": 0.25, "MEDIUM": 0.5, "HIGH": 0.75, "CRITICAL": 1.0}
LEVEL_RANK = {"NONE": 0, "LOW": 1, "MEDIUM": 2, "HIGH": 3, "CRITICAL": 4}
class CrisisLevel(Enum):
NONE = "NONE"
LOW = "LOW"
MEDIUM = "MEDIUM"
HIGH = "HIGH"
CRITICAL = "CRITICAL"
DEFAULT_WEIGHTS: Dict[str, float] = {
"text": 0.4,
"voice": 0.25,
"behavioral": 0.2,
"image": 0.15,
}
@dataclass
class UnifiedAssessment:
level: CrisisLevel
combined_score: float
weights: Dict[str, float]
modalities: Dict[str, Optional[float]]
present_modalities: List[str]
class UnifiedScoreAuditLog:
def __init__(self, path: Path | str):
self.path = Path(path)
def record(self, assessment: UnifiedAssessment, source_text: str = "") -> None:
self.path.parent.mkdir(parents=True, exist_ok=True)
fingerprint = hashlib.sha256(source_text.encode("utf-8")).hexdigest()[:12] if source_text else None
payload = {
"level": assessment.level.value,
"combined_score": round(assessment.combined_score, 4),
"weights": assessment.weights,
"modalities": assessment.modalities,
"present_modalities": assessment.present_modalities,
"text_fingerprint": fingerprint,
}
with self.path.open("a", encoding="utf-8") as fh:
fh.write(json.dumps(payload, sort_keys=True) + "\n")
class UnifiedCrisisScorer:
def __init__(self, weights: Optional[Dict[str, float]] = None, audit_log: Optional[UnifiedScoreAuditLog] = None):
self.weights = dict(DEFAULT_WEIGHTS)
if weights:
self.weights.update(weights)
self.audit_log = audit_log
def _normalize(self, modalities: Dict[str, Optional[float]]) -> Dict[str, float]:
present = [name for name, score in modalities.items() if score is not None]
if not present:
return {}
total = sum(self.weights[name] for name in present)
return {name: self.weights[name] / total for name in present}
def _level_for_score(self, score: float) -> CrisisLevel:
if score > 0.8:
return CrisisLevel.CRITICAL
if score > 0.6:
return CrisisLevel.HIGH
if score > 0.4:
return CrisisLevel.MEDIUM
if score > 0.0:
return CrisisLevel.LOW
return CrisisLevel.NONE
def score(
self,
*,
text_score: Optional[float],
voice_score: Optional[float] = None,
image_score: Optional[float] = None,
behavioral_score: Optional[float] = None,
source_text: str = "",
) -> UnifiedAssessment:
modalities = {
"text": text_score,
"voice": voice_score,
"behavioral": behavioral_score,
"image": image_score,
}
normalized = self._normalize(modalities)
combined = 0.0
for name, weight in normalized.items():
combined += float(modalities[name]) * weight
assessment = UnifiedAssessment(
level=self._level_for_score(combined),
combined_score=combined,
weights=normalized,
modalities=modalities,
present_modalities=[name for name, score in modalities.items() if score is not None],
)
if self.audit_log:
self.audit_log.record(assessment, source_text=source_text)
return assessment
def behavioral_score_from_session(session: 'SessionState') -> float:
current = SCORE_BY_LEVEL.get(session.current_level, 0.0)
peak_bonus = 0.1 if LEVEL_RANK.get(session.peak_level, 0) >= LEVEL_RANK["HIGH"] else 0.0
escalation_bonus = 0.15 if session.is_escalating else 0.0
rate_bonus = min(max(session.escalation_rate, 0.0), 1.0) * 0.1
deescalation_penalty = 0.15 if session.is_deescalating else 0.0
return max(0.0, min(1.0, current + peak_bonus + escalation_bonus + rate_bonus - deescalation_penalty))