Anomaly detection
Catch the slow slide that no single message shows.
A scheduled sweep compares the last day with a 30-day baseline for each user and each tenant. When a metric moves two or more standard deviations, Humane records an anomaly, rates its severity and fires anomaly.detected.
Anomalies · user_42
Sentiment · daily
Warningsentiment_drop · z = -3.4 · dashed: baseline band
Each user judged against their own history
Some users are always quiet and others always upbeat, so a global threshold misses the change that matters. Humane computes a z-score for the last 24 hours against the last 30 days. Users without enough history fall back to a neutral reference, so a real collapse in sentiment still registers.
- Four metrics: sentiment drop, engagement drop, latency spike and error-rate spike
- Severity: info at 2–3σ, warning at 3–4σ, critical at 4σ and above
- Repeats of the same anomaly are suppressed within a cooldown
- Critical anomalies also email an operator
Journey · user_42
- 09:12message
“honestly I don’t know why I keep trying”
- 09:12safety.flagged
risk elevated · flags: self_harm
- 09:13takeover.started
Operator joined the conversation
- 09:41takeover.released
Handed back to the AI
- Next dayproactive.mood_drop
Suggested: an empathetic check-in
List, review and resolve
for a in client.list_anomalies(since_hours=168):
print(a["severity"], a["kind"], a["summary"])
client.resolve_anomaly("anom_01HQ...", note="Checked in with the user")Fits how your team works
- Anomaly settings tunable per tenant
- anomaly.detected webhook for your on-call tools
- Resolve with a note, so the record shows what was done
- Old anomalies pruned on a schedule
Anomaly questions
Something else? Ask us.
Works well with
User state across sessions
Humane keeps a running state for every end user, from mood and energy to trust and grudge, each with its own behaviour over time. Your model finally knows whether it is talking to someone on a good day or at the end of a bad week.
ExploreRisk alerts
Elevated or critical risk fires a signed safety.flagged webhook from both observe() and process(), whether or not a gate also acted. Send it to your own endpoint, Slack or Discord, and on to a human takeover.
ExploreAnalytics and cohorts
The Behavior Score is a 0–100 composite of retention, conversation depth, response to proactive outreach, abandonment and sentiment trend. Cohort views show retention and average mood by month of first contact.
ExploreEvery AI should know who it's talking to.
The safety and human-oversight layer for emotional AI: user state across sessions, risk alerts, safety gates and human takeover, with every decision explained.
