Per-Channel Quantile Fullgrid Probing for Small Language Model Observation

We describe a dense activation probing protocol for small to mid-sized language models that operates within a fixed cell budget while maintaining high observation coverage. The protocol combines per-channel quantile binning, an expanded calibration probe corpus (P=1200), and an additive shard merge step that allows heterogeneous compute resources to contribute. Across thirteen runs spanning seven model families and parameter counts from 124M to 7.6B, the protocol reaches at least 90.6 percent coverage on every run, with a median of 94.95 percent and a maximum of 97.05 percent. Per-channel binning combined with a larger probe budget was associated with higher coverage compared to earlier pooled-bin, smaller-corpus configurations on the same models, but the isolated contribution of each factor remains untested in this draft. A controlled 2x2 ablation is the recommended next experiment. The pipeline is organized for reproducibility and can be released alongside the run artifacts. It runs on commodity hardware.