DATA AVAILABILITY — Paper 3
"A permutation-null artifact drives a conservation–divergence spectrum's signature trend, and phylogeny dominates a protein language model's per-position coupling signal"

All data underlying the findings are within the manuscript and its supplementary datasets.

Per-position score tables:
- three_detectives.csv — the three per-position scores (S_sdp, S_net, S_plm) and their ranks for all 212 GPCR positions
- decomposition.csv — phylogenetic decomposition of the PLM signal (random-split vs clade-aware balanced accuracy per position)
- null_sides.csv — permutation-null divergence, excess, and z-score per position
- comparable_scores.csv — cross-family-comparable net-divergence scores

Cross-family and anchoring:
- crossfamily_null_summary.csv — observed vs null conservation–divergence trend for four families (GPCR, opsin, NR-DBD, HLA)
- anchor_enrichment.csv — gold-standard interface enrichment for each detective
- interface_goldstandard.csv — per-position Gα-interface and class-discrimination gold standard
- layer_scan_summary.csv — layer/probe robustness scan (ESM-2 layers 6/15/21/30, linear and MLP probes)

Opsin negative control:
- opsin_plm_probe.csv — per-position PLM random-split and clade-aware balanced accuracy on 348 opsin positions (ESM-2 650M, layer 33)
- opsin_negative_control_summary.csv — top-list recall and enrichment statistics for both detectives

Source data: class A GPCR sequences and coupling assignments from GtoPdb/GPCRdb; opsin sequences and spectral classes as in companion preprint P2. ESM-2 embeddings computed with fair-esm (esm2_t30_150M_UR50D for GPCR; esm2_t33_650M_UR50D for the opsin control). Analysis code (per-position probes, permutation null, phylogenetic decomposition, gold-standard anchoring) is archived on Zenodo: https://doi.org/10.5281/zenodo.21449958 (shared with companion preprints P1 and P2, which use the same conservation-divergence spectrum instrument).

Companion preprints P1 and P2 (Research Square, 2026; DOIs to be added on posting).