[!NOTE] Working Paper Draft v1.0. The working paper, pipeline, and full numbers trace are public in the repository.
Overview
Do language models know what warm sounds like? Six production LLMs are asked to translate semantic audio descriptors (“warm”, “bright”) into 5-band parametric EQ curves, graded against SAFE-DB, a crowdsourced corpus of real audio-engineer settings collected in-plugin a decade before LLMs could have shaped it. There is no LLM judge anywhere in the pipeline; the gold standard is numeric human behavior in rendered frequency-response space.
Key Findings
- Models are hyper-central, not wrong. Median Human-Relative Percentile is 3.0 where a typical engineer sits at 50; models show 5.9x less dispersion than real engineers.
- RAG makes it worse. Grounding with five individual human exemplars degrades centrality (HRP 3.0 to 10.5), while grounding with the consensus centroid snaps the model onto the anchor. Models defer to aggregates and are destabilized by instances, the exportable design finding for numeric RAG.
- Human practice is multimodal and models miss the mixture. The data contains three warm and four bright “schools”; models concentrate into single modes (22 of 24 goodness-of-fit tests significant).
- On rare descriptors, nothing beats a flat curve. Deployed text-to-EQ tools serving open vocabulary are, on this evidence, decorating a no-op for most requests.
- 37 of 53 tests survive consolidated Benjamini-Hochberg FDR correction; the 16 failures are reported, not hidden.