# Which bucket adapters increase their target? This report analyzes the existing 72-adapter listening grid. No new audio, judging, training or best-seed selection was performed. Each adapter is compared with its own no-new-adapter baseline at strengths 0.5, 0.75 and 1.0. Each strength has 20 paired observations: five languages, two neutral-wording three-sentence texts per language, and two seeds. The literal text, GENERAL/SCRIPT prompt, seed, fixed speaker reference and frame budget match across conditions. All recordings, including termination failures, are retained. The shared parent is Humaneness Voice Large 1.5 Alpha with its real-speech rank-4 adapter already merged. Strength scales only the new rank-32 adapter residual. CFG is 1. ## Primary effect and the displayed score The primary effect is `(adapter_raw - baseline_raw) / frozen_training_std`. For a low-pole target, its sign is reversed. Thus a positive effect always means more of the requested trait, including slower tempo or lower tension. These are changes in training-standard-deviation units, not probabilities and not effect sizes based on the variability of this small evaluation sample. For example, +0.2 z means a gain of one fifth of the frozen training standard deviation for that head; it does not mean 20% more perceived emotion. The original page score is also shown. It transforms each raw prediction with the frozen training mean/std and a Gaussian CDF; for low-pole targets it is inverted. It can saturate. Averages of this nonlinear transform can move in a different direction from raw averages, so the report exposes both summaries and flags disagreements. A tiny consistent scorer change need not be audible. ## Uncertainty and selection Both texts and both seeds are kept together within each language cluster. We enumerate all 3,125 bootstrap resamples of the five language clusters and report percentile intervals. The stricter displayed interval is 98.33%, an exploratory allowance for examining three strengths within one adapter. With only five clusters, this is not a strong population-level confidence claim. The operational label **clearer gain on this grid** requires a positive lower bound of this stricter interval and positive means in at least four of the five languages. It does not mean statistically established benefit across all 72 adapter comparisons. For the two style-contrast requests, the opposite head must also move down under the same stricter criterion; increasing only the primary head does not establish the whole contrast. The starting strength is the largest primary gain among clearer strengths if any qualify. Otherwise the largest positive mean is an uncertain trial. If no tested strength has a positive primary mean, the no-adapter base is suggested. This strength selection uses the same evaluation grid, not independent held-out validation. It is not a claim of a universally optimal strength. The overview highlights clearer gains of at least 0.1 training-standard-deviation units as promising starting points. This is an explicit display heuristic, not a calibrated perceptual threshold. Smaller effects remain in the full table. ## Quality costs and limitations Changes in Content Enjoyment, genuineness, word error rate and frame-cap rate are displayed separately; they do not contribute to the target-score ranking. WER changes use percentage points; positive WER changes are worse. Positive CE/genuineness changes are higher predicted scores. These are frozen Humaneness EARS predictions and the existing multilingual ASR scores on complete waveforms before MP3 encoding. They are not human opinions. EARS is also the model family used in data screening. There is one fixed voice reference per language and only two translated texts, so language and speaker effects cannot be separated and generalization to new voices or richer acting scenes is unproven. Listen to the paired players before adopting a setting. ## Reproduction and artifacts `steering-comparison.json` contains every adapter/strength aggregate and every per-language result. `steering-comparison.csv` provides a flat version. Input JSON hashes and the normalization/taxonomy hashes are in the report. The original per-bucket JSON files remain public beside each audio page. The public analysis source uses Python's standard library: ```bash python steering_analyze.py --site downloaded-space \ --normalization downloaded-space/steering-normalization.json \ --taxonomy downloaded-space/steering-taxonomy.json --out reproduced ``` `downloaded-space` must contain `manifest.json` and the existing per-bucket JSON pages from this Space, plus the two frozen metadata files linked here. The deliberately omitted low-warmth/cold-delivery bucket has no adapter and is not silently assigned a result.