Reactive autoscalers compress workload state into a scalar, but bounded ceiling partitions jointly induced by signal and target determine replica recommendations. We formalize five signals and distinguish the positive target, bounded desired recommendation, and deployed workers. We prove exact recommendation equivalence, a tight charge-estimation error bound, and an expected pre-bound quantization premium for i.i.d. mean-matched charges. A 100-seed executable check verified theorem implementation, and 27,900 simulations mapped threshold–capacity trade-offs. Four paired Windows process-level faster-whisper studies tested near-equivalence, partition crossing, and target relocation. E4 provided a local adjacent-frontier comparison in 60 fresh three-arm blocks; count targets 1.00 and 0.90 and a mean-matched work target replayed the same 48-request manifest in each block. Count 0.90 versus 1.00 improved deadline attainment by +2.92 pp (95% CI +0.97 pp to +4.90 pp) while increasing capacity by +0.0322 worker-s/audio-s. Work differed from count 1.00 by +2.57 pp and from count 0.90 by −0.35 pp. An independent audit found 134/137 held-out bundles deadline-feasible in isolated service; conditioning preserved both the E2 work-minus-count effect (+2.56 pp) and the E4 target effect (+2.96 pp). Within the tested partitions, these results separate threshold relocation from signal information and motivate testing adjacent targets before adopting a richer estimator; they do not establish the global count-target frontier.

Distribution-shift alarms do not establish classifier harm. We evaluated the detectability, frozen-head relevance, and performance deterioration of two public DenseNet-121 chest-radiography classifiers under 30 controlled image conditions. Signals covered pixels, 64-component PCA representations, logits, probabilities, and calibrated high-error selection. An exact whole-model-null intervention changed representations while preserving all 18 released output slots to within 4.09 × 10⁻¹⁴; PCA monitoring detected all six interventions, five-logit monitoring detected none, and the paired Brier change was negligible. Across image shifts, PCA and logit alarm rates were 97% and 52%. In held-source/held-family prediction of primary Brier harm, probability mean displacement had the lowest error (MAE 0.00679); neither PCA signal outperformed the training-mean baseline after correction, whereas three output-facing signals beat both prespecified baselines. Secondary results were outcome-dependent: PCA RFF-MMD ranked best for macro-AUC and locked-F1 loss. An exploratory ConvNeXtV2-Atto/MobileViT-XS arm reproduced the construct separation: representation monitoring detected 57/60 conditions and all six task-null controls, while logit monitoring detected no task-null control; probability mean had the lowest held-architecture/held-family Brier-harm error ratio (0.404). The results show that detectability, task relevance, and metric-specific harm are distinct constructs and should be reported separately.