Intended use
MoodSpan is a research prototype with strict source-support gates for mental-health education. It is not a diagnostic, treatment, triage, or clinician-replacement system.
Research methods
How MoodSpan measures a source-supported education prototype.
The current work is evaluation infrastructure: retrieval coverage, source support, citation behavior, safety routing, and explicit failure reporting. The system is not presented as clinically validated.
Current release gate
0.0%
Visible unsupported-claim rate against a 25.0% release target. The current blocker is corpus provenance and human review, not the stored strict gate alone.
Retrieval first
Kira searches retrieved source material before answering clinical questions. Retrieved chunks, source IDs, citation checks, and grounding metrics are retained for evaluation.
Claim discipline
Clinical factual sentences must stay tied to retrieved support. Shorter answers and abstention are preferred when the retrieved evidence is thin.
Safety boundary
Crisis and self-harm signals route through deterministic safety checks before ordinary education content. Safety responses do not replace emergency or professional care.
Evaluation protocol
The system is evaluated before it is trusted.
Retrieval evaluation on the internal held-out query set
Response-quality judging with unsupported-claim measurement
Citation coverage and invalid-citation checks
Grounded-answer contract with pass, warn, and refuse decisions
Failure taxonomy for unsupported examples
Human groundedness review packet, built but not yet completed
What this does not claim
Clinically validated
Safe for clinical use
Diagnosis or treatment recommendation
Clinician-reviewed content corpus
User-outcome evidence