Independent confirmation + a control-arm twist: residual codes look like sampling freedom, not recoverable audio information
Hi — I've been working on an independent verification effort around the same wall you documented (no RVQ quantizer in dav.pth), taking the workaround path: predict c0 only from audio, let the public Local LM re-sample c1-c7. Sharing a few results that intersect directly with your FINDINGS, including one place where I think I can strengthen a negative result of yours.
- On "Native-state Stage-1 residual prediction ❌ (mean CE 6.834)"
We hit the same wall, but with a control-arm setup that separates "our predictor is too weak" from "the information isn't in the audio":
- arm A: audio (DAV 128ch) + Global LM state + lower codes
- arm B: audio zeroed, context only
- arm C: A evaluated with deranged audio (same prompt, different clip)
Result over 120 clips / 24k frames: CE_B − CE_A (delta_audio) is negative at all 7 depths (−0.03 to −0.05 nat), far from a +0.10 nat threshold; delta_shuffle is weakly positive — the model uses audio a bit, but it doesn't help prediction (overfit direction). Unigram baseline sits at ~6.90–6.95 ≈ ln(1024), and the native Local LM's own policy NLL on the same frames is 2.48–3.34. So at least in our regime (same-prompt 8s clips), the residual codes look like sampling freedom, not recoverable acoustic detail — the LM context nearly determines what you can know, and the rest is entropy.
Your 6.834 lands right in our unigram band, which suggests your result is the same phenomenon — but with a ctx-only arm you could claim the stronger version ("information absent") instead of "we failed to learn it." Worth adding if you revisit that experiment.
- On "Arbitrary-WAV continuation ❌"
One thing we can confirm: you don't need to recover all 8 streams. Forcing true c0 only (c1-c7 re-sampled by the Local LM) keeps source identity well above controls on objective metrics: log-mel 0.816 / chroma 0.565 vs. prompt-only natural generation at 0.552 / 0.041, CI-separated (prompt-cluster bootstrap, n=30). Full-code replay saturates at 0.938 mel (the fresh-noise resynthesis ceiling on our pipeline is ~0.946). So a learned conditioner that targets c0 alone has a real signal to hit — your repeat/roll baselines may have been competing against an unnecessarily hard full-token target.
- On token capture
We found bit-exact replay is achievable if you persist the full policy state (variable-length support IDs, guided logits, policy + untruncated NLL, Flow noise/scheduler/latent) under CUBLAS_WORKSPACE_CONFIG=:4096:8 + deterministic algorithms: 35/35 clips reproduce hidden/condition/latent/waveform with zero error, and 3 negative controls (1 priming code, 1 mid-frame code, Flow noise swap) all trip the corresponding tolerances.