Prosodia — an audio Jev

A System One-shaped decision model that takes speech as its state: audio is encoded once, several typed questions branch over that single encoding and are answered in parallel as calibrated probability distributions. No transcript, no speech recognition, no generated text — the Whisper decoder never runs.

Across the projects the README surveys, vision is covered — PlayJev reads a game frame in one forward pass — but every other audio entry routes ASR to text and then to Jev, so the state is a transcript and the audio never reaches the model. Prosodia feeds audio features directly, which skips ASR latency and its error propagation on the current turn.

Its readout scores each option as an independent inner product, so p(a)/p(b) cannot change when a third option is added: it satisfies independence of irrelevant alternatives exactly. Jev demonstrably violates that — appending an irrelevant option shifts log-odds between two untouched options by roughly -0.28 across ten randomised blocks, every one decreasing — so Prosodia is not yet in Jev's functional class.

A train-fitted constant predictor scores ECE 0.0097, and ranking the 12 arms by ECE versus Brier gives near-independent orders (Spearman +0.147, -0.294, -0.266). The Whisper control encoder beat WavLM everywhere. The prosody effect did not survive its control, and class collapse had two causes: joy recovered under reweighting while sadness recall never exceeded 2.9% under a 6.9x upweight.

Apparently yes: pitch_direction reaches about 0.51 against a 0.36 majority, which the README reads as an ASR-trained encoder carrying contour information that transcription never needed.

The README links a demo Space but describes it as private, so treat it as unavailable unless you are granted access. The repository itself, its findings tables and the Phase 2 design notes are the public material.

Everything is one seed and the README says to treat single numbers as approximate. The architecture was constant across all twelve arms, so nothing there bears on whether the Jev shape helps, and closing the IIA gap needs options that attend to each other instead of independent frozen embeddings.