Loading recorded reports…
Chatterbox T3 · acoustic-code evidence
Follow one speech code—then replay a causal edit.
Three frozen TTS runs compare fitted acoustic-code readouts. The bridge report also connects depth, local text diagnostics, the actual head, and a recorded residual intervention.
Fitted layers and the actual HEAD answer different questions. Hover or focus each slice to inspect exact candidate rank and readout probability.
READING BOUNDARY
Three measurements, three claims.
Fitted readout: which acoustic-code direction is readable from an intermediate residual.
Text diagnostics: which prompt tokens receive local gradient sensitivity or within-text attention.
Residual intervention: whether a small recorded edit changed the emitted code and its autoregressive continuation.