Alerts: confidence gate before any push, held alerts, release on corroboration (server-26#300)

No confidence field exists anywhere in scene/incident data, so the alerter
derives one: match weight (talkgroup 1.0, extracted tag 0.8, raw transcript
0.6) scaled down only when stt_confidence is under ALERT_STT_REFERENCE (0.3),
zero for not_speech. Below ALERT_MIN_CONFIDENCE, or high severity (term list,
or a major incident) without ALERT_CORROBORATION_MIN_CALLS calls or
ALERT_HIGH_SEVERITY_MIN_CONFIDENCE, the match is written to alert_holds
(no Firestore rule, admin-only GET /alerts/held) and logged instead of pushed.
A needs_corroboration hold is released when a later call brings its incident
to the call minimum; a rule pushed by that call supersedes its hold.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

drb-c2-core/app/internal/alerter.py

@@ -2,9 +2,19 @@
 Alert dispatch engine.
 
 Loads enabled alert rules from Firestore and checks each one against the call's
-talkgroup ID, tags, and transcript.  On a match:
+talkgroup ID, tags, and transcript.  On a match that passes the confidence
+gate (server-26#300, `_assess`):
   1. Creates an AlertEvent document in Firestore.
   2. Optionally POSTs a Discord webhook message if the rule has one configured.
+A match that fails the gate is written to `alert_holds` instead -- never
+pushed, readable by admins (GET /alerts/held) -- and a hold that was waiting
+only on corroboration is released when a later call brings its incident to
+alert_corroboration_min_calls calls.
+
+WHY A GATE: CrimeRadar pushed "shots fired at an elementary school" after STT
+heard "fire alarm" as "firearm". A keyword hit in raw transcript text is the
+weakest evidence this pipeline has, and a push alert is the one surface that
+cannot be taken back.
 
 Raw transcript text is withheld from both by default -- see _snippet_allowed
 and server-26#85.
@@ -54,6 +64,8 @@ async def check_and_dispatch(
     # Loop-invariant: every rule here belongs to the same org, so the opt-in is
     # resolved once rather than per match.
     snippet_allowed = await _snippet_allowed(org_id)
+    evidence = await _evidence(call_doc or {})
+    pushed_rules: set[str] = set()
 
     for rule in rules:
         matched_keywords = _match_rule(rule, talkgroup_id, tags, transcript)
@@ -63,6 +75,12 @@ async def check_and_dispatch(
         alert_id = str(uuid.uuid4())
         snippet = _snippet(transcript) if snippet_allowed else None
         now = datetime.now(timezone.utc).isoformat()
+        verdict = _assess(matched_keywords, tags, evidence)
+        if not verdict["push"]:
+            await _hold(alert_id, rule, org_id, call_id, node_id, talkgroup_id,
+                        talkgroup_name, matched_keywords, snippet, now, verdict, evidence)
+            continue
+        pushed_rules.add(rule.get("rule_id", ""))
         event = {
             "alert_id": alert_id,
             "org_id": org_id,
@@ -76,6 +94,9 @@ async def check_and_dispatch(
             "transcript_snippet": snippet,
             "triggered_at": now,
             "acknowledged": False,
+            "confidence": verdict["confidence"],
+            "high_severity": verdict["high_severity"],
+            "corroborating_calls": evidence["max_calls"],
         }
 
         try:
@@ -92,6 +113,192 @@ async def check_and_dispatch(
         if webhook_url:
             await _post_webhook(webhook_url, rule.get("name", ""), talkgroup_name, matched_keywords, snippet)
 
+    await _release_corroborated(rules, evidence, call_id, pushed_rules)
+
+
+# ── Confidence gate (server-26#300) ─────────────────────────────────────────
+
+# How a rule matched, as evidence. Talkgroup activity is metadata, not
+# recognised speech, so it cannot be misheard.
+_W_TALKGROUP = 1.0
+_W_TAG = 0.8          # the extraction model put it in tags
+_W_TRANSCRIPT = 0.6   # substring of raw STT text only (above the 0.5 default floor)
+
+
+def _norm_term(term: str) -> str:
+    return (term or "").strip().lower().replace("-", " ")
+
+
+def _is_high_severity_term(term: str) -> bool:
+    wanted = {_norm_term(t) for t in settings.alert_high_severity_terms}
+    return _norm_term(term) in wanted
+
+
+async def _evidence(call_doc: dict) -> dict:
+    """What the call doc and its incidents say about how much to trust it."""
+    incident_ids = list(call_doc.get("incident_ids") or [])
+    max_calls = 0
+    max_severity = call_doc.get("severity")
+    severity_rank = {"routine": 0, "minor": 1, "moderate": 2, "major": 3}
+    for iid in incident_ids:
+        try:
+            inc = await fstore.doc_get("incidents", iid) or {}
+        except Exception as e:
+            logger.warning(f"Alerter: could not read incident {iid}: {e}")
+            continue
+        max_calls = max(max_calls, len(inc.get("call_ids") or []))
+        if severity_rank.get(inc.get("severity"), -1) > severity_rank.get(max_severity, -1):
+            max_severity = inc.get("severity")
+    stt = call_doc.get("stt_confidence")
+    return {
+        "incident_ids": incident_ids,
+        "max_calls": max_calls,
+        "severity": max_severity,
+        "stt_confidence": stt if isinstance(stt, (int, float)) else None,
+        "not_speech": bool(call_doc.get("transcript_not_speech")),
+    }
+
+
+def _assess(matched: list[str], tags: list[str], evidence: dict) -> dict:
+    """
+    Push or hold, and why. Pure, so the gate is testable without Firestore.
+
+    confidence = match weight x STT factor. The STT factor is 1.0 unless
+    Whisper's own mean token probability is under alert_stt_reference, and
+    1.0 when there is none (a json-only model) rather than guessing.
+    """
+    content = [m for m in matched if not m.startswith("talkgroup:")]
+    if not content:
+        weight = _W_TALKGROUP
+    elif any(m in (tags or []) for m in content):
+        weight = _W_TAG
+    else:
+        weight = _W_TRANSCRIPT
+
+    stt = evidence.get("stt_confidence")
+    stt_factor = 1.0
+    if content and stt is not None and settings.alert_stt_reference > 0:
+        stt_factor = min(1.0, max(0.0, stt / settings.alert_stt_reference))
+    if content and evidence.get("not_speech"):
+        stt_factor = 0.0  # the corrector called this recogniser noise
+    confidence = round(weight * stt_factor, 3)
+
+    high = bool(content) and (
+        any(_is_high_severity_term(m) for m in content)
+        or any(_is_high_severity_term(t) for t in (tags or []))
+        or evidence.get("severity") == "major"
+    )
+    corroborated = evidence.get("max_calls", 0) >= settings.alert_corroboration_min_calls
+
+    if confidence < settings.alert_min_confidence:
+        return {"push": False, "confidence": confidence, "high_severity": high,
+                "hold_reason": "low_confidence"}
+    if high and not corroborated and confidence < settings.alert_high_severity_min_confidence:
+        return {"push": False, "confidence": confidence, "high_severity": high,
+                "hold_reason": "needs_corroboration"}
+    return {"push": True, "confidence": confidence, "high_severity": high, "hold_reason": None}
+
+
+async def _hold(alert_id, rule, org_id, call_id, node_id, talkgroup_id, talkgroup_name,
+                matched_keywords, snippet, now, verdict, evidence) -> None:
+    """Record a gated match where admins can see it, and push nothing."""
+    hold = {
+        "alert_id": alert_id,
+        "org_id": org_id,
+        "rule_id": rule.get("rule_id", ""),
+        "rule_name": rule.get("name", ""),
+        "call_id": call_id,
+        "node_id": node_id,
+        "talkgroup_id": talkgroup_id,
+        "talkgroup_name": talkgroup_name or "",
+        "matched_keywords": matched_keywords,
+        "transcript_snippet": snippet,
+        "triggered_at": now,
+        "status": "held",
+        "hold_reason": verdict["hold_reason"],
+        "confidence": verdict["confidence"],
+        "high_severity": verdict["high_severity"],
+        # One incident to wait on: the first the call linked to. A call that
+        # linked to none can never be corroborated, so it stays held.
+        "incident_id": (evidence["incident_ids"] or [None])[0],
+        "corroborating_calls": evidence["max_calls"],
+    }
+    try:
+        await fstore.doc_set("alert_holds", alert_id, hold, merge=False)
+    except Exception as e:
+        logger.warning(f"Alerter: could not save held alert: {e}")
+        return
+    logger.info(
+        f"Alert HELD ({verdict['hold_reason']}, confidence={verdict['confidence']}, "
+        f"high_severity={verdict['high_severity']}, calls={evidence['max_calls']}): "
+        f"rule='{rule.get('name')}' call={call_id} keywords={matched_keywords}"
+    )
+
+
+async def _release_corroborated(rules: list[dict], evidence: dict, call_id: str,
+                                pushed_rules: set[str]) -> None:
+    """
+    Push holds that were waiting only on corroboration once their incident
+    has enough calls. A low-confidence hold is never released this way: a
+    second call on the incident says the event is real, not that the first
+    call was heard right. A rule this call already pushed for supersedes its
+    hold instead of alerting twice.
+    """
+    if evidence["max_calls"] < settings.alert_corroboration_min_calls:
+        return
+    by_id = {r.get("rule_id", ""): r for r in rules}
+    for iid in evidence["incident_ids"]:
+        try:
+            holds = await fstore.collection_list("alert_holds", incident_id=iid, status="held")
+        except Exception as e:
+            logger.warning(f"Alerter: could not read holds for incident {iid}: {e}")
+            continue
+        for hold in holds:
+            if hold.get("hold_reason") != "needs_corroboration":
+                continue
+            rule = by_id.get(hold.get("rule_id", ""))
+            now = datetime.now(timezone.utc).isoformat()
+            if rule is None or hold.get("rule_id", "") in pushed_rules:
+                status = "superseded" if rule is not None else "rule_gone"
+                await _set_hold_status(hold, status, call_id, now)
+                continue
+            event = {k: hold.get(k) for k in (
+                "alert_id", "org_id", "rule_id", "rule_name", "call_id", "node_id",
+                "talkgroup_id", "talkgroup_name", "matched_keywords",
+                "transcript_snippet", "confidence", "high_severity",
+            )}
+            event.update({
+                "triggered_at": now,
+                "acknowledged": False,
+                "released_from_hold": True,
+                "corroborated_by_call_id": call_id,
+                "corroborating_calls": evidence["max_calls"],
+            })
+            try:
+                await fstore.doc_set("alert_events", hold["alert_id"], event, merge=False)
+            except Exception as e:
+                logger.warning(f"Alerter: could not release held alert {hold.get('alert_id')}: {e}")
+                continue
+            await _set_hold_status(hold, "released", call_id, now)
+            pushed_rules.add(hold.get("rule_id", ""))
+            logger.info(
+                f"Alert released from hold: rule='{hold.get('rule_name')}' incident={iid} "
+                f"corroborated by call {call_id} ({evidence['max_calls']} calls)"
+            )
+            webhook_url = rule.get("discord_webhook")
+            if webhook_url:
+                await _post_webhook(webhook_url, rule.get("name", ""), hold.get("talkgroup_name"),
+                                    hold.get("matched_keywords") or [], hold.get("transcript_snippet"))
+
+
+async def _set_hold_status(hold: dict, status: str, call_id: str, now: str) -> None:
+    try:
+        await fstore.doc_update("alert_holds", hold["alert_id"], {
+            "status": status, "resolved_at": now, "resolved_by_call_id": call_id,
+        })
+    except Exception as e:
+        logger.warning(f"Alerter: could not update hold {hold.get('alert_id')}: {e}")
+
 
 async def _snippet_allowed(org_id: Optional[str]) -> bool:
     """