Mapika/decider-4b-GGUF, file decider-4b-v2.1-Q4_K_M.gguf, ran through decider-ai 1.6.0 and llama-cpp-python 0.3.35 with Metal on an M5 Pro. The model reads a shared question and numbered passages, then scores one noul question per passage in a single system_one(..., independent=False) call. The n_ctx setting was 4096. Source revision: 23579f7a; model snapshot: b79f09d9. This uses Decider's typed probability output; ordinary Ollama text generation would not score the model correctly.

The paired trial used HotpotQA distractor validation examples [400:700] with their native contexts. All scorers used Barber's public transform with keep=0.6 and the same chunking, pins, markers, and no-negative-savings guard. Per-question records contain token counts, supporting titles, latency, and scorer input usage.

Decider missed a supporting passage on five questions: 453, 537, 649, 663, and 667. Four overlap Kev's 13 missed-support questions; 537 is new. It kept the support passages on nine questions Kev missed. Relative to Kev, Decider saved only 3,223 more downstream prompt tokens across 300 examples while spending 145 more seconds on selection and reading 69,761 more scorer tokens. Total Decider selection time was 434.5 seconds, versus 289.4 for Kev and 27.7 for BGE. No net dollar savings can be claimed without the downstream input price and local GPU cost.

For an answer-level check, local qwen3.6:27b-mlx answered from Decider's selected context on all 36 previously checked questions plus new miss 537. Temperature was zero and thinking disabled. Full and BGE contexts were also checked for 537. Decider's answer became clearly wrong relative to both full context and BGE on 537 (Melbourne → insufficient information) and 667 (Wikimedia Foundation → insufficient information). It answered correctly on the other three missed-support questions and on three of Kev's four previously observed clear regressions (468, 514, 675). The remaining Kev regression, 667, persisted. Per-question answers are saved for inspection.

The targeted check found individual regressions, but cannot establish the overall quality difference. We therefore generated paired BGE and Decider answers for all 300 examples with the same local qwen3.6:27b-mlx model, temperature zero, and thinking disabled. Using HotpotQA-style lowercasing, punctuation/article removal, and token F1 against the reference answer:

Decider's paired F1 advantage was 0.0288; a 10,000-resample paired bootstrap gave a 95% interval of approximately +0.0018 to +0.0563. It won 34 questions, lost 14, and tied 252 by F1. Token F1 is imperfect for verbose answers: some large numeric differences were equivalent answers with different phrasing. Manual review of large differences found real Decider wins (for example 469, 496, 564, 606) and losses (411, 537, 667).

On these native contexts, Decider reduced prompt tokens and improved measured aggregate answer quality relative to BGE. It is now available as the explicit barber.embedders.decider() option. The zero-dependency lexical default remains appropriate for users who have not chosen to download a 2.7 GB model and run local inference. Medium and large contexts remain untested. Whether Decider reduces total dollars depends on the downstream input price and local GPU cost: against BGE it saved 100,997 additional prompt tokens but used 406.8 more seconds of scorer time over these 300 examples.