120 BPM8A0:00$0 of $40
KvairiInboxScoresFitOpsSongDemo data

Emotion Fit

The example night, with a simulated felt path. Ratings you save on this machine replace this picture. No other listener is in the prior.

Predicted fit

How close the next block should feel.

Mean fit across the night is 92 percent. Journey objective 75 percent. Duration weights the segment rewards.

Felt against the target

Solid line is the target. Dashed line is what was felt.

  • Nostalgia

    100 percent fit

    Intensity match 67 percent. Quality 82 percent.

    No skip. Reward 88 percent.

    Target 60 percent, felt 93 percent

  • Tension

    77 percent fit

    Intensity match 33 percent. Quality 48 percent.

    Skipped. Reward 40 percent.

    Target 80 percent, felt 13 percent

  • Joyful activation

    100 percent fit

    Intensity match 91 percent. Quality 88 percent.

    No skip. Reward 95 percent.

    Target 90 percent, felt 99 percent

Nostalgia at 100 percent fit. Keep this setting.

Tension at 77 percent fit. Ease the next block toward what you actually felt.

The system learned to raise intensity for the feelings you rate highly.

Reward weights: fit 0.55, intensity 0.25, quality 0.2, skip penalty 0.2.

Synthetic listener

Seed 7, 40 sessions, a nostalgia preference at valence 0.10, arousal 0.30, dominance -0.10. The bandit trains on fit, intensity match, quality, and skip. The 1 to 10 slider is that same listener's score. Amber is the bandit. Blue is a random policy. Synthetic listener. No real listener data.

Bandit fit moves from 82 to 88 percent. Regret moves from 17 to 11 percent. A random policy stays near 76 percent fit.

More synthetic listeners

Three made-up people besides the nostalgia curve. Each one wants a different setting early, mid, and late. Heart rate stays out. Fixed-best is one setting for the whole night: the median preference on tempo, intensity, and brightness, and the hierarchical prior on everything else. It knows the preference and still cannot change with the block. The bandit starts from that prior. Synthetic listener. No real listener data.

ListenerBandit fitFixed fitRandom fitBandit regretFixed regret
Synthetic nostalgia listener. Not a real person.8810076110
Synthetic tension listener. The preference climbs across the night. Not a real person.8984761015
Synthetic peacefulness listener. The preference thins out across the night. Not a real person.938681614
Synthetic power listener. The preference arrives late. Not a real person.8683761317