Guiding Principle β Empowers the Caregiver, Never Replaces Them
The tool empowers the caregiver β it never replaces them. The system executes monitoring rules the clinician defined and continues their decisions; it never makes a clinical decision.
Note: section numbers below (Β§9, Β§12.2, Β§12.2.1, Β§12.2.2, Β§12.3) are preserved from the source specification for traceability.
Β§12.2 β The Uncertainty-Mapping Engine
Rewritten from a verdict engine. Instead of telling the caregiver what a message means, the engine maps what is certain, what is possible, and what only the caregiver can know.
Fixed Three-Part Output
- 1 What is literally written β the literal reading only, with no interpretation added.
- 2 A possible alternative reading β "if this is ironic / understated / exaggerated, the meaning flips to X." Always phrased as a hypothesis, never a determination.
- 3 What would decide β phrased as a question to the caregiver, drawing on extra-textual information only they hold: patient history, personal style, the content of the last session. E.g. "Did the last session deal with Y?", "Is this a departure from his usual style?"
Removed from the Spec
- β Numeric sentiment scores β no more "sadness: 0.82"
- β Unambiguous emotion classifications β no verdicts on what the patient feels
- β Autonomous "concerning thoughts" alert β replaced by caregiver-controlled alerts (see Β§12.2.2)
Example β the Uncertainty Card as It Appears in the App ("Readings for your review")
"It's all fine, don't worry about me."
WhatsApp Β· yesterday 23:10
What was said, literally
Reports that things are fine and reassures you not to worry.
A possible alternative reading
Read through the "understatement" style you noted, "fine" may be minimising, and "don't worry" can close a topic before it opens.
What would decide it β only you know
Did S8 touch the work review he was dreading? Is closing-off a change from how he usually writes to you?
Β§12.2.1 β Signal Taxonomy
Not all signals are equal. The engine distinguishes three types β each handled by a different pathway.
π§
A Β· Caregiver Priors
Per-patient communication-style tags recorded by the caregiver: irony, understatement, exaggeration, avoidance of detail. Every tag carries a timestamp and a source β the session in which it was recorded.
π
B Β· Observable Triggers
Low-inference, measurable signals: message frequency, message length, specificity, sending hours, channel switching, repeated superficial or generic statements. These may alert directly. Every such alert includes the rule the caregiver defined, the raw data behind it, and the deviation from the patient's personal baseline.
π
C Β· Interpretive Readings
High-inference readings: irony, hidden distress, subtext. Never surfaced as an unambiguous alert β always through the Β§12.2 three-part uncertainty-mapping format.
Example β "Triggers from rules you set" in the App
Measurable changes β each names your rule and the raw data behind it
π Message frequency down
Rule you set
Daily messages drop > 50% vs 14-day baseline
What was measured
4.2 / day β 1.1 / day
βοΈ Messages getting shorter
Rule you set
Average length < 40% of baseline
What was measured
22 words β 6 words
π Unusual send hours
Rule you set
Messages sent 01:00β05:00
What was measured
3 nights Β· 02:40, 03:15, 04:02
The Temporal Model of Priors β Decay Without Deletion
- β A prior's weight decays with the age of the assessment β a recent assessment outweighs an old one
- β Nothing is ever deleted. Patient behaviour is non-linear and relapses happen; the full history is preserved, and the historical trajectory stays visible to both the engine and the caregiver
- β Changes in the treatment or in the patient receive temporally boosted weight
Example β the Monitoring Profile in the App (priors + temporal model)
Communication style you've noted
Sherpii weighs its readings against these Β· Edit any time
π· Understatement
Said "fine" the week PHQ-9 rose
Noted S4 Β· 15 May
π· Irony / deflection
Deflects hard topics with humour
Noted S7 Β· 22 May
π· Detail-avoidance
Prefers brief written replies
Noted S4 Β· 17 Apr
Your assessment over time
Newer notes carry more weight β the full path stays visible
02 Jun Β· S8 Weighted new
More open; engaged with restructuring. Less deflection than S7.
22 May Β· S7
Guarded about the upcoming work review; deflected with humour.
15 May Β· S6
Family context surfaced; some minimising of low mood.
01 May Β· Intake
Prior MDD (F32.1). Baseline reserved, brief written replies.
Β§12.2.2 β Caregiver-Controlled Alert Sensitivity
The alert that replaced the autonomous "concerning thoughts" alert. The threshold belongs to the caregiver β the system only executes it.
The Sensitivity Parameter
- β Alerts are defined on clear expressions of distress β where the threshold ("what counts as clear") is a sensitivity parameter the caregiver sets per patient
- β Sensitivity is tunable at every session β e.g., deterioration in the patient or their environment β temporarily heightened sensitivity
- β Every change is logged: date, previous value, new value, and a reason note
Mandated Phrasing
Throughout the product and this document, the system "executes the monitoring rule that was defined" β it never "detects distress". This is not cosmetic: it is the boundary between continuing a clinical decision and making one.
Example β the Alert-Sensitivity Control in the App
Alert sensitivity
LowBalancedElevatedHigh
π Last changed: 22 May Β· S7 β Reason: Work performance review coming up β raised sensitivity to clear distress markers for the next two weeks.
Β§12.3 β Prosody Extended to the Bridge Channel
Previously limited to session audio β now extended to voice messages and between-session calls on the Bridge.
An Independent Signal
Voice messages and between-session calls are treated as an independent prosodic signal on the Bridge β the only channel that does not share the biases of text analysis. How something is said can carry information the words themselves do not.
ContentβTone Gap Analysis
When the text reads positive but the prosody is flat, the mismatch is presented as a gap for the caregiver's judgment in the Β§12.2 three-part structure β never as a conclusion.
Example β the Voice Note in the App
βΆ
Content
"All good, I'm fine β let's talk after the weekend."
Tone Independent signal
Flat affect Β· slow tempo (β30% vs baseline) Β· low pitch variation Β· long pauses
Content and tone don't line up
The words say fine; the voice reads flat and slow. The two channels don't match. Worth a gentle check at your next contact?
Β§9 β Provenance of the Cumulative Background
The cumulative background contains AI outputs β it is therefore not an independent verification layer. An earlier misreading can contaminate later readings.
Source Tagging Requirements
- β Every context item carries a source tag: caregiver-entered or AI-output (with a link to the original analysis)
- β When the Β§12.2 engine uses context from Β§9, it marks which context items are AI-sourced
- β An AI summary is never displayed at the same visual weight as a caregiver entry β mockups show an "auto-generated" tag plus a source link
Anchoring Warning
A rich background biases toward reading neutral text as distress. The protection: the output always separates "what is in the text" from "what the context adds" β so the caregiver can see where an impression is coming from.
Example β Source Tags on Context Items
Caregiver entry
Prior MDD (F32.1). Baseline reserved, brief written replies.
Auto-generated Β· view source
S8 summary: more open; engaged with restructuring.
Cross-Cutting Principle β Empowers the Caregiver, Never Replaces Them
Every system output is phrased as a continuation of a caregiver decision: "the rule you defined fired", "per the sensitivity you set", "for your review".
Acceptance Criterion β Every Alert Must Answer
- a Which caregiver-defined rule fired?
- b Based on what data?
- c What is the system missing in order to decide?
Forbidden Phrasings
- β Diagnostic phrasing β "the system diagnosed", "the patient is depressed"
- β Opaque risk scores β "Risk level: HIGH"
- β Autonomous treatment recommendations
Open Item β Regulation
β OPEN β not closed in this update
The regulatory classification of the alerting component β under the Israeli Privacy Law Amendment 13 and medical-device software rules β is still open. Note: phrasing the feature as "executing clinician-defined monitoring rules" plus "uncertainty mapping" (rather than "diagnosis" or "autonomous clinical alerting") may affect the classification. Review with a regulatory expert is required before this item can be closed.
Mockup Mapping
The spec is the source of truth for behaviour; the mockup for visual layout. Explore the live app mockup.
- Β§12.2 Patient card β "Between sessions" tab β three-part uncertainty cards under "Readings for your review" ("What was said, literally" / "A possible alternative reading" / "What would decide it β only you know")
- Β§12.2.1 Patient card β "Monitoring" tab β style tags with timestamps, assessment timeline, sensitivity slider with change log
- Β§12.2.2 Patient card β "Between sessions" tab β "Triggers from rules you set" visually separates observable triggers (rule + raw data + baseline deviation, e.g. "Daily messages drop > 50% vs 14-day baseline") from interpretive readings
- Β§12.3 Voice-note view in the "Between sessions" tab β prosody shown beside text as an independent channel
Version History
- v1.0 Original definition β Β§12.2 as a verdict engine: sentiment score, sadness/anxiety level, autonomous alert on "concerning thoughts".
- v1.1 (July 2026) Β§12.2 rewritten verdict β uncertainty-mapping; Β§12.2.1 signal taxonomy added; Β§12.2.2 caregiver-controlled sensitivity added; Β§12.3 extended to the Bridge channel; Β§9 provenance requirements added; cross-cutting principle added; regulation open item added.