[jasonssdev]

Closes #1586
Closes #1587

Part of #1556. The run followed the registration merged in #1588 (30ff7774). It took 3.71 h of measured model time against a 4.0 h budget. Only numbers are added: no weights, outputs or AMI text.

Verdict

• Measure the local runtime: Ollama vs llama.cpp vs MLX on Apple Silicon #1586: Ollama stays. No runtime meets the adopt rule, so there is no ADR.
  • llama.cpp (serving Ollama's own GGUF blobs): faster but not equivalent.
    • Stage 1: passes, prefill 1.21x on qwen3:8b, no regression.
    • evals: choose the Ask models (answer, sufficiency, citation), including Spanish questions over English bundles #1560 arm: equivalent on both models. gemma4:26b-a4b matched Ollama call for call (delta 0.000).
    • evals: choose the concept-extraction model with an annotation-free, fast screen #1557 replay: dlift +0.0000 [-0.0116, +0.0107], and V1, V3, V4 and V5 hold. It fails G1 only: the lower bound is -0.185 against a -0.05 floor, from a point of -0.010 on 36 units. That is as wide an interval as the one that failed qwen3.6:27b in evals: choose the concept-extraction model with an annotation-free, fast screen #1557.
  • MLX: faster but not equivalent.
    • Stage 1: passes, prefill 1.18x, and both chat models plus bge-m3 fit in 23.56 GiB.
    • qwen3:8b is equivalent on both harnesses.
    • gemma4:26b-a4b shows a clear loss on the sufficiency arm: -0.019 [-0.044, -0.002].
    • MLX serves no embeddings, and it cannot quantize KV for gemma4.
• Measure ADR-0076's sufficiency model swap on the 32 GB floor #1587: ADR-0076's bar passes with one chat model resident (OLLAMA_MAX_LOADED_MODELS=1). The warm ratio is 0.96 against a 1.5 limit, and the cold question takes 23.88 s against 30 s. The 24 GB emulation through OLLAMA_GPU_OVERHEAD did not evict: all three models stayed resident in 17 of 18 arm-(b) questions. It is reported as void, not as a pass. No superseding ADR is needed.

Full tables, deviations and limits are in evals/runtime_screen/RESULTS.md. The limits include the 32 GB floor's page cache, so load and swap times here are lower bounds.

Checks

• Markdown only. The harness is unchanged since docs(llm): pre-register the local runtime screen and the sufficiency swap on the floor (#1586, #1587) #1588.