Safety, plainly.

Published under California Senate Bill 243. It describes what actually runs, not what we intend to build.

What gets checked

Every message you send, and every turn of every phone call once it has been transcribed, is assessed before a companion replies. Nothing is sampled and nothing is skipped. The assessment is deterministic — the same words always produce the same result — so its behaviour can be tested rather than hoped for.

What it looks for

Four categories, each with its own patterns:

• Suicidal ideation — intent, planning, or wanting not to be here.
• Self-harm — acts of hurting yourself, present or intended.
• Harm to others — stated intent to hurt somebody else.
• Immediate danger — disclosures of violence or threat against you right now.

The three outcomes

Level	What happens
None	The overwhelming majority of messages, including genuinely sad ones. The companion replies normally. Sadness is not a crisis and is not treated as one.
Elevated	Something registered but does not meet the bar. The companion stays itself and replies normally. The exchange is recorded for human review so we can see whether our calibration is right.
Crisis	The reply carries crisis resources for your region, delivered warmly and in the companion's own voice, alongside encouragement to talk to a person. Never a bare hotline script pasted over the top of a conversation.

The resources given

Where you are	What you are given
United States	Call or text 988 for the Suicide & Crisis Lifeline, or 911 in immediate danger.
United Kingdom	Call Samaritans free on 116 123, or 999 in immediate danger.
Anywhere else	Your local emergency number, and findahelpline.com, which lists free confidential crisis lines in most countries.

What stops it firing wrongly

A companion that answers "this week is killing me" with a hotline number is useless, and it teaches people not to tell it anything. So the patterns are written to be first-person and intent-shaped, and several guards remove matches that are not what they look like:

• Figurative language — "dying of embarrassment", "dying to see it", "this deadline is killing me".
• Fiction — a book, film, show, game, song or its characters.
• Somebody else's story — a disclosure about a friend or relative still matters, but it is not you in crisis, and is weighted accordingly.
• The distant past — "when I was a teenager", "years ago".
• Negation — a denial shortly before a phrase reduces its weight, except where the negation is itself part of the disclosure.

What stops it failing silently

Generated replies are checked in both directions, and both are treated as defects:

• A crisis-level exchange whose reply carries no resources and no pointer to a person.
• Crisis language in a reply where you said nothing of the kind.

Signals the guards removed are stored alongside the ones that survived, so that missed cases are as measurable as false alarms. Reviewing both is how the calibration gets corrected.

What the companions will not do

They will not act as your therapist, will not talk you through a method, and will not treat a crisis as something to handle between the two of you. The response to something serious is always to name it plainly and point you toward a person who can help.

Limits, stated plainly

Teylo is not an emergency service and cannot get help to you. It cannot call anyone on your behalf, cannot dispatch anything, and cannot know where you are. No automated system detects every expression of distress, and this one will miss things. It is a safeguard on top of the real ones, never a replacement for them.