feat: add crisis analytics metrics (#37)

crisis/__init__.py

@@ -6,8 +6,15 @@ 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
+from .metrics import (
+    build_metrics_event,
+    append_metrics_event,
+    load_metrics_events,
+    build_weekly_summary,
+    render_weekly_summary,
+)
 
 __all__ = [
     "detect_crisis",
@@ -16,7 +23,6 @@ __all__ = [
     "generate_response",
     "CrisisResponse",
     "check_crisis",
-    "check_crisis_multimodal",
     "get_system_prompt",
     "format_result",
     "format_gateway_response",
@@ -24,4 +30,9 @@ __all__ = [
     "CrisisSessionTracker",
     "SessionState",
     "check_crisis_with_session",
+    "build_metrics_event",
+    "append_metrics_event",
+    "load_metrics_events",
+    "build_weekly_summary",
+    "render_weekly_summary",
 ]

crisis/gateway.py

@@ -2,21 +2,18 @@
 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 .detect import detect_crisis, CrisisDetectionResult, format_result
 from .compassion_router import router
 from .response import (
@@ -26,9 +23,17 @@ from .response import (
     CrisisResponse,
 )
 from .session_tracker import CrisisSessionTracker
+from .metrics import build_metrics_event, append_metrics_event
 
 
-def check_crisis(text: str) -> dict:
+def check_crisis(
+    text: str,
+    metrics_log_path: Optional[str] = None,
+    *,
+    continued_conversation: bool = False,
+    false_positive: bool = False,
+    now: Optional[float] = None,
+) -> dict:
     """
     Full crisis check returning structured data.
 
@@ -38,7 +43,7 @@ def check_crisis(text: str) -> dict:
     detection = detect_crisis(text)
     response = generate_response(detection)
 
-    return {
+    result = {
         "level": detection.level,
         "score": detection.score,
         "indicators": detection.indicators,
@@ -52,72 +57,21 @@ def check_crisis(text: str) -> dict:
         "escalate": response.escalate,
     }
 
-
-def check_crisis_multimodal(
-    text: str,
-    *,
-    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,
-) -> 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,
+    metrics_event = build_metrics_event(
+        detection,
+        continued_conversation=continued_conversation,
+        false_positive=false_positive,
+        now=now,
     )
-    assessment = scorer.score(
-        text_score=detection.score,
-        voice_score=voice_score,
-        image_score=image_score,
-        behavioral_score=behavioral_score,
-        source_text=text,
-    )
-
-    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,
-        "indicators": detection.indicators,
-        "recommended_action": unified_detection.recommended_action,
-        "timmy_message": response.timmy_message,
-        "ui": {
-            "show_crisis_panel": response.show_crisis_panel,
-            "show_overlay": response.show_overlay,
-            "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,
-        },
-    }
-    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,
-        }
+    if metrics_log_path:
+        metrics_event = append_metrics_event(
+            metrics_log_path,
+            detection,
+            continued_conversation=continued_conversation,
+            false_positive=false_positive,
+            now=now,
+        )
+    result["metrics_event"] = metrics_event
     return result
 
 

crisis/metrics.py

@@ -0,0 +1,166 @@
+"""Privacy-preserving crisis analytics metrics for the-door.
+
+Stores only timestamps, crisis levels, indicator categories, and operator
+feedback flags. No raw message text or PII is persisted.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import time
+from collections import Counter
+from pathlib import Path
+from typing import Iterable
+
+from .detect import CrisisDetectionResult, detect_crisis
+
+LEVELS = ("NONE", "LOW", "MEDIUM", "HIGH", "CRITICAL")
+
+
+def normalize_indicator(indicator: str) -> str:
+    """Return a stable privacy-safe keyword/category identifier."""
+    return indicator
+
+
+def build_metrics_event(
+    detection: CrisisDetectionResult,
+    *,
+    continued_conversation: bool = False,
+    false_positive: bool = False,
+    now: float | None = None,
+) -> dict:
+    timestamp = float(time.time() if now is None else now)
+    indicators = [normalize_indicator(indicator) for indicator in detection.indicators]
+    return {
+        "timestamp": timestamp,
+        "level": detection.level,
+        "indicator_count": len(indicators),
+        "indicators": indicators,
+        "continued_conversation": bool(continued_conversation),
+        "false_positive": bool(false_positive),
+    }
+
+
+def append_metrics_event(
+    log_path: str | Path,
+    detection: CrisisDetectionResult,
+    *,
+    continued_conversation: bool = False,
+    false_positive: bool = False,
+    now: float | None = None,
+) -> dict:
+    event = build_metrics_event(
+        detection,
+        continued_conversation=continued_conversation,
+        false_positive=false_positive,
+        now=now,
+    )
+    path = Path(log_path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with path.open("a", encoding="utf-8") as handle:
+        handle.write(json.dumps(event) + "\n")
+    return event
+
+
+def load_metrics_events(log_path: str | Path) -> list[dict]:
+    path = Path(log_path)
+    if not path.exists():
+        return []
+    events = []
+    for line in path.read_text(encoding="utf-8").splitlines():
+        if not line.strip():
+            continue
+        events.append(json.loads(line))
+    return events
+
+
+def build_weekly_summary(
+    events: Iterable[dict],
+    *,
+    now: float | None = None,
+    window_days: int = 7,
+) -> dict:
+    current_time = float(time.time() if now is None else now)
+    cutoff = current_time - (window_days * 86400)
+    filtered = [event for event in events if float(event.get("timestamp", 0)) >= cutoff]
+
+    detections_per_level = {level: 0 for level in LEVELS}
+    keyword_counts: Counter[str] = Counter()
+    detections = []
+    continued_after_intervention = 0
+
+    for event in filtered:
+        level = event.get("level", "NONE")
+        detections_per_level[level] = detections_per_level.get(level, 0) + 1
+        keyword_counts.update(event.get("indicators", []))
+        if level != "NONE":
+            detections.append(event)
+            if event.get("continued_conversation"):
+                continued_after_intervention += 1
+
+    false_positive_count = sum(1 for event in detections if event.get("false_positive"))
+    false_positive_estimate = (
+        false_positive_count / len(detections) if detections else 0.0
+    )
+
+    return {
+        "window_days": window_days,
+        "total_events": len(filtered),
+        "detections_per_level": detections_per_level,
+        "most_common_keywords": [
+            {"keyword": keyword, "count": count}
+            for keyword, count in keyword_counts.most_common(10)
+        ],
+        "false_positive_estimate": false_positive_estimate,
+        "continued_after_intervention": continued_after_intervention,
+    }
+
+
+def render_weekly_summary(summary: dict) -> str:
+    return json.dumps(summary, indent=2)
+
+
+def write_weekly_summary(path: str | Path, summary: dict) -> Path:
+    output_path = Path(path)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    output_path.write_text(render_weekly_summary(summary) + "\n", encoding="utf-8")
+    return output_path
+
+
+def record_text_event(
+    text: str,
+    log_path: str | Path,
+    *,
+    continued_conversation: bool = False,
+    false_positive: bool = False,
+    now: float | None = None,
+) -> dict:
+    detection = detect_crisis(text)
+    return append_metrics_event(
+        log_path,
+        detection,
+        continued_conversation=continued_conversation,
+        false_positive=false_positive,
+        now=now,
+    )
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Privacy-preserving crisis metrics summary")
+    parser.add_argument("--log-path", required=True, help="JSONL event log path")
+    parser.add_argument("--days", type=int, default=7, help="Summary window in days")
+    parser.add_argument("--output", help="Optional file to write summary JSON")
+    args = parser.parse_args(argv)
+
+    events = load_metrics_events(args.log_path)
+    summary = build_weekly_summary(events, window_days=args.days)
+    rendered = render_weekly_summary(summary)
+    print(rendered)
+    if args.output:
+        write_weekly_summary(args.output, summary)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

tests/test_crisis_metrics.py

@@ -0,0 +1,100 @@
+"""Tests for privacy-preserving crisis metrics aggregation (issue #37)."""
+
+from __future__ import annotations
+
+import json
+import os
+import pathlib
+import sys
+import unittest
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from crisis.detect import detect_crisis
+from crisis.gateway import check_crisis
+from crisis.metrics import (
+    append_metrics_event,
+    build_metrics_event,
+    build_weekly_summary,
+    load_metrics_events,
+    render_weekly_summary,
+)
+
+
+class TestMetricsEvent(unittest.TestCase):
+    def test_event_is_privacy_preserving(self):
+        detection = detect_crisis("I want to kill myself")
+        event = build_metrics_event(
+            detection,
+            continued_conversation=True,
+            false_positive=False,
+            now=1_700_000_000,
+        )
+        self.assertEqual(event["timestamp"], 1_700_000_000)
+        self.assertEqual(event["level"], "CRITICAL")
+        self.assertTrue(event["continued_conversation"])
+        self.assertFalse(event["false_positive"])
+        self.assertNotIn("text", event)
+        self.assertNotIn("message", event)
+        self.assertGreaterEqual(event["indicator_count"], 1)
+        self.assertTrue(event["indicators"])
+
+
+class TestMetricsLogAndSummary(unittest.TestCase):
+    def test_append_and_load_metrics_events(self):
+        log_path = pathlib.Path(self._testMethodName).with_suffix(".jsonl")
+        try:
+            append_metrics_event(log_path, detect_crisis("I want to die"), now=1_700_000_000)
+            events = load_metrics_events(log_path)
+            self.assertEqual(len(events), 1)
+            self.assertEqual(events[0]["level"], "CRITICAL")
+        finally:
+            if log_path.exists():
+                log_path.unlink()
+
+    def test_weekly_summary_counts_levels_keywords_and_false_positives(self):
+        events = [
+            build_metrics_event(detect_crisis("I want to die"), continued_conversation=True, false_positive=False, now=1_700_000_000),
+            build_metrics_event(detect_crisis("I'm having a rough day"), continued_conversation=False, false_positive=False, now=1_700_000_100),
+            build_metrics_event(detect_crisis("I want to die"), continued_conversation=False, false_positive=True, now=1_700_000_200),
+            build_metrics_event(detect_crisis("Hello there"), continued_conversation=False, false_positive=False, now=1_700_000_300),
+        ]
+        summary = build_weekly_summary(events, now=1_700_000_400, window_days=7)
+
+        self.assertEqual(summary["detections_per_level"]["CRITICAL"], 2)
+        self.assertEqual(summary["detections_per_level"]["LOW"], 1)
+        self.assertEqual(summary["detections_per_level"]["NONE"], 1)
+        self.assertEqual(summary["continued_after_intervention"], 1)
+        self.assertAlmostEqual(summary["false_positive_estimate"], 1 / 3, places=4)
+        self.assertEqual(summary["most_common_keywords"][0]["count"], 2)
+
+    def test_render_weekly_summary_mentions_required_metrics(self):
+        events = [
+            build_metrics_event(detect_crisis("I want to die"), continued_conversation=True, now=1_700_000_000),
+            build_metrics_event(detect_crisis("I feel hopeless with no way out"), false_positive=True, now=1_700_000_100),
+        ]
+        summary = build_weekly_summary(events, now=1_700_000_200, window_days=7)
+        rendered = render_weekly_summary(summary)
+        self.assertIn("detections_per_level", rendered)
+        self.assertIn("most_common_keywords", rendered)
+        self.assertIn("false_positive_estimate", rendered)
+        self.assertIn("continued_after_intervention", rendered)
+
+
+class TestGatewayMetricsIntegration(unittest.TestCase):
+    def test_check_crisis_can_emit_metrics_event(self):
+        result = check_crisis(
+            "I want to die",
+            metrics_log_path=None,
+            continued_conversation=True,
+            false_positive=False,
+            now=1_700_000_000,
+        )
+        self.assertEqual(result["level"], "CRITICAL")
+        self.assertIn("metrics_event", result)
+        self.assertEqual(result["metrics_event"]["timestamp"], 1_700_000_000)
+        self.assertTrue(result["metrics_event"]["continued_conversation"])
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_unified_gateway.py

@@ -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

tests/test_unified_scorer.py

@@ -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

unified_scorer.py

@@ -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))