Human in the loopGuard

Proactive triggers

Reach out first when the signals say it's time.

Humane watches each user's state and fires an event when a threshold is crossed: mood dropping, energy running low, a user going quiet, trust passing a milestone. Each trigger has a cooldown and carries a suggested action, and your app decides how to reach out.

Proactive triggers

  • mood_dropmood < 0.30cooldown 1h
  • energy_lowenergy < 0.25cooldown 2h
  • trust_milestonetrust > 0.70, 10+ turnscooldown 24h
  • user_inactivesilent for 48hcooldown 24h
  • engagement_dropquiet 24h, mood < 0.5cooldown 12h

→ proactive.mood_drop + proactive.triggered · suggested_action included

Illustration of proactive triggers and their defaults
Built-in triggers

Five triggers, plus your own

mood_drop

Mood below 0.30. Cooldown one hour. Suggests an empathetic check-in.

energy_low

Energy below 0.25. Cooldown two hours. Suggests fewer, simpler messages.

user_inactive

No messages for 48 hours. Cooldown one day. Suggests re-engaging from the last conversation.

engagement_drop

An established user quiet for a day with mood below 0.5. Cooldown 12 hours. Suggests following up on the last topic.

trust_milestone

Trust above 0.70 after more than ten interactions. Once a day at most. Suggests more personal recommendations.

Your own

Policy rules fire proactive.<name> events on any condition they can match.

Learn more
Delivery

Two events per firing, so you can subscribe either way

Each firing sends a specific event, such as proactive.mood_drop, and the umbrella proactive.triggered with the trigger name at the top level. Cooldowns are claimed atomically in the database, so two workers can't fire the same trigger for the same user twice.

  • Payload includes the user's state and a suggested action
  • Cooldowns survive restarts and work across workers
  • Proactive events also come back on process() responses
  • Slack and Discord connectors can post them to a channel

Alerts · live

  • safety.flaggedcritical
    user_88 · flags: self_harm→ Webhook · Slack
  • safety.holdhold
    user_42 · reply held for review→ Webhook
  • anomaly.detectedwarning
    user_17 · engagement_drop→ Webhook
  • proactive.mood_droptrigger
    user_05 · mood below 0.30→ Discord
Illustration of a risk alert feed
Handle it

One handler for every trigger

python
event = json.loads(raw)   # after verify_webhook_signature()

if event["event"] == "proactive.triggered":
    d = event["data"]
    print(d["trigger_name"], d["suggested_action"])
    schedule_check_in(d["end_user_id"], reason=d["trigger_name"])

Trigger questions

Something else? Ask us.

Every 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.