The 48-profile action space the meta-controller selects from is described in the paper (§3) but isn't yet documented in the repo. We should add an explicit table mapping each profile index to its (ASR config, LLM config) tuple.
Acceptance criteria
Why
Without this, anyone subclassing BaseRouter or interpreting PretrainedPAVORouter outputs has to read paper §3 to know what each profile index means. The 4-profile collapse the public API uses is documented; the underlying 48-profile space is not.
The mapping is in the paper's Appendix B; this is mostly a transcription job.
The 48-profile action space the meta-controller selects from is described in the paper (§3) but isn't yet documented in the repo. We should add an explicit table mapping each profile index to its (ASR config, LLM config) tuple.
Acceptance criteria
docs/PROFILES.mdwith a 48-row table:| Profile idx | ASR | ASR config | LLM | LLM config | Expected latency | Expected quality |
pavo_bench/_profile_costs.pyif any profile metadata changes.Why
Without this, anyone subclassing
BaseRouteror interpretingPretrainedPAVORouteroutputs has to read paper §3 to know what each profile index means. The 4-profile collapse the public API uses is documented; the underlying 48-profile space is not.The mapping is in the paper's Appendix B; this is mostly a transcription job.