Early detection of crisis signals in therapy: how we assist the clinician without replacing them

Before anything else: CauceOS does not diagnose. It assists. This distinction is not semantic — it is ethical and clinical, and we repeat it in every product surface, every disclaimer, and every conversation with the professionals already testing the system.

That said, there is one thing we can do well: help the therapist not miss a signal that matters.

The problem we are solving

In a 50-minute session, a therapist holds several loads at once. They follow content. They model the alliance. They manage transference. They make framing decisions. They watch body language on a small screen. And at the same time, they need to listen for signals that can change the course of the session: suicidal ideation, self-harm plans, descriptions of domestic violence, dissociative episodes, indicators of acute risk.

The reality is that human attention has limits. A patient can say, in the middle of a long sentence about their work week, "sometimes I think if I weren't here no one would notice" — and if the therapist was tracking a different line, the sentence passes. It should not pass.

What the co-pilot does

CauceOS listens to the live transcript and compares it against patterns of language associated with risk. We are not looking for an isolated keyword — we are looking for contextualized patterns.

The categories we monitor in the MVP:

• Suicidal ideation — direct and indirect verbalizations, expressions of burden ("everyone would weigh less without me"), plan or method mentioned, lethality factors (access to means).
• Self-harm — descriptions of active or recent past self-injurious behavior, language of emotional regulation through physical harm.
• Domestic or gender-based violence — descriptions of coercion, control, isolation, threats, typical minimizing language ("he got angry but it was nothing").
• Clinically significant dissociation — described time loss, depersonalization, derealization.
• Substance abuse in acute crisis — recent escalation, intoxication during the session, prior overdoses.

When the system identifies a signal, it does not interrupt the session. A discreet notification appears in the therapist's panel with three elements:

1. The exact quote that triggered the alert, with a precise timestamp.
2. The category (for example: "Possible suicidal ideation — direct verbalization").
3. A fixed disclaimer: "This is algorithmic assistance, not a diagnosis. Clinical interpretation is yours."

The therapist decides what to do with that information. They can ignore it because they had already noticed. They can use it to ask a follow-up question they were going to skip. They can flag it for supervision. The decision is always human.

Why latency matters

An alert that reaches the therapist 90 seconds after the sentence is useless. The clinical moment has passed. The patient has already changed topics. The chance to ask the right question is gone.

That is why we optimize every part of the pipeline so that the time between "the sentence was said" and "the therapist sees it on screen" is under 2 seconds in normal network conditions. It is one of the most expensive architectural commitments of the product, and one of the most important.

Compare this with the dominant market model: transcribe the full session, generate a post-session summary, deliver it the next day. If that summary contains a phrase of suicidal ideation the therapist did not catch live, the information arrives 24 hours late. The session where intervention was possible already ended.

What we do not do

To remove any ambiguity:

• We do not diagnose. We do not output "this patient is in crisis". We output "this sentence matches a pattern associated with X, review".
• We do not call emergency services automatically. That decision belongs to the professional and the protocols of their practice.
• We do not replace risk assessment protocols. A validated suicide ideation assessment scale remains the clinical standard. We are a layer of early alerting, not an assessment instrument.
• We do not log alerts to the patient file without your consent. You decide what is documented and how.

How we trained it

The detections are not simple keyword matches. We use specialized models trained specifically to identify crisis language in conversational context, in neutral Spanish and English, including the indirect and euphemistic forms that predominate in Spanish-speaking cultures.

We validate the patterns with clinical professionals in consultation — practicing therapists who review de-identified transcripts and tell us which signals we correctly detected, which we missed, and which were false positives. Each iteration improves the precision-recall curve. We will share concrete metrics later.

A note on false positives

In crisis detection, a false positive is preferable to a false negative, within reasonable limits. If the system flags a sentence you already had clear, you lost 2 seconds. If the system fails to flag a sentence you did not hear, you lost a clinical opportunity.

We calibrate the system biased toward sensitivity by default, with the option for each professional to adjust the threshold to their practice. A therapist specializing in trauma may want high sensitivity. An executive coach may prefer only the strongest signals.

If you work with high-risk populations

If your practice includes suicide prevention, support for violence survivors, work with adolescents, or any population where early detection is critical, we want to talk to you. We are refining these capabilities specifically with feedback from professionals in these areas, and early users get priority access to the advanced detection configuration.

CauceOS will not replace your training, your intuition, or your clinical supervision. That is how it should be. What it can do is stay attentive to what your attention, in the moment, cannot catch.

That is the promise. That is what we are building.