Signals that actually help
Message cadence, topic diversity, and willingness to confirm plans modestly predict ghosting risk—but not long-term satisfaction. Better models weight values statements you verify through prompts, not just swipe direction history.
Known limits
No API reads “chemistry.” Algorithms can reduce bad pairings; they cannot manufacture consent or emotional labor. Treat percentage scores as triage, not prophecy.
Human checkpoints
Weekly reflection—“did I feel more seen?”—anchors tech outputs to somatic truth. Pair app usage with offline hobbies so self-worth is not entirely algorithm-conditioned.