Runner documentation

Every page under docs/ in the repository, by file name. Pages that exist on this site link here; the rest link to GitHub.

file title
adaptation-engine.md The adaptation engine: scoring, adapters, and training in the serving binary
afmoe-cert-goal-2026-08-05.md Goal: certify afmoe (Trinity-Nano) on this box — 2026-08-05 evening
afmoe-cert-report-2026-08-05.md afmoe (Trinity-Nano) certification report — 2026-08-05
afmoe-divergence-triage-2026-08-05.md afmoe greedy-identity failure: root cause found — 2026-08-05 evening
afmoe-sensitivity-floor-2026-08-05.md afmoe (Trinity-Nano) sensitivity-floor run — 2026-08-05
agent-compatibility.md Coding-agent compatibility evidence
agent-profile.md Xyntetik agent profile metadata
agent-torture.md The agent torture suite
atem-tool-calling-plan.md Build plan — atem constrained tool calling
bench-2026-08-01-3070.md Cross-engine benchmarks — RTX 3070 box, 2026-08-01
benchmarks.md GPU benchmarks — Runner vs llama.cpp (CUDA)
blackwell-results-2026-08-17.md Blackwell Q3_K rung results — 2026-08-17
cert-matrix-2026-08-05.md Cert-matrix detailed report — GPT-OSS x Gemma 4 derivative ecosystems
cert-matrix-goal-2026-08-05.md Goal: artifact certification matrix — GPT-OSS × Gemma 4 derivative ecosystems
cert-matrix-status.md Cert-matrix status — GPT-OSS x Gemma 4 derivative ecosystems
compatibility-program.md Compatibility program
context-drafts.md Context-grounded drafts: --draft-lookup
cuda-gptoss-divergence-2026-08-19.md The remaining gpt-oss CPU/CUDA divergence is chaotic amplification, not a wrong op
cuda-gptoss-router-bias-2026-08-18.md The CUDA fused MoE path routed gpt-oss without its router bias
cuda-microbatch-identity-2026-08-18.md The CUDA decode microbatch is not bit-identical on quantized models
determinism-scope.md What the determinism claim covers, exactly
gemma4-31b-metal-m5max-2026-09-01.md gemma-4-31B dense Metal validation — M5 Max, 2026-09-01
gpt-oss-120b-metal-m5max-2026-08-31.md gpt-oss-120B fully resident on Metal — M5 Max, 2026-08-31
gpt-oss-harmony-2026-08-14.md gpt-oss Harmony chat — before and after, 2026-08-14
granite-cert-2026-08-11.md granite (IBM Granite 4.1 dense) certification — 2026-08-11
idle-coexistence-120b-m5max-2026-09-01.md Idle coexistence at 120B — M5 Max, 2026-09-01
llama33-70b-metal-m5max-2026-08-31.md Llama 3.3 70B dense Metal validation — M5 Max, 2026-08-31
metal-bigfile-limits-m5max-2026-09-01.md Big-file limits exercised for real — M5 Max, 2026-09-01
metal-decode-dispatch-budget-2026-09-01.md Where the 120B’s decode milliseconds go — M5 Max, 2026-09-01
metal-decode-kv-traffic-2026-08-15.md What the long-context decode loss actually is: the KV read, at 1.5 GB/s
metal-dispatch-census-2026-08-13.md Metal dispatch census: where the batch dimension actually collapses
metal-fallback.md Metal runtime fallback ownership
metal-full-router-trace-2026-08-31.md Full Metal router-logit trace
metal-gemma4-moe-divergence-2026-08-31.md gemma-4-26B-A4B generated token 0 on Metal: an unclamped GELU in the routed-expert kernel
metal-long-context-decode-2026-08-14.md Metal decode at long context: the chunked path already captures it
metal-microbatch-decode-2026-09-01.md Metal microbatch decode — M5 Max, 2026-09-01
metal-moe-grouped-mma-2026-09-01.md Grouped-MMA MoE prefill: the win, and the instrument that judges it — M5 Max, 2026-09-01
metal4-tensor-measurement-2026-08-31.md Metal 4 tensor-op GEMM: first M5 measurement, and what it does to the premise
model-scope.md Model scope — geographic focus: Europe & US (standing policy, 2026-07-29)
moe-support.md Sparse-MoE support — implementation and test report
muse-atem-cert-2026-08-11.md Muse Glimmer native atem certification — 2026-08-11
muse-glimmer-cert-2026-08-11.md muse-glimmer (Meta Muse Glimmer 30B) certification — 2026-08-11
negative-result-expert-cache.md Negative result: an expert-residency cache tier for MoE models larger than RAM
negative-result-harmony-analysis-bound.md Negative result: lifting the Harmony analysis bound on the auto branch
negative-result-metal-gemm-occupancy.md Negative result: Metal prefill GEMM occupancy and tile shape
negative-result-metal-moe-expert-major.md Negative result: expert-major MoE prefill kernels — M5 Max, 2026-09-01
negative-result-metal-multirow-matvec.md Negative result: multi-row-per-simdgroup Metal decode matvec
negative-result-muse-glimmer-selective-2026-08-20.md Negative result: Runner-built selective quants of Muse Glimmer 30B
negative-result-nemotron-lightning-artifacts-2026-08-20.md Negative result: three of the four planned Nemotron-3.5-Lightning artifacts
ornith-reference.md Ornith 1.0 9B reference gate
performance.md Performance: closing the CPU/GPU gap
quant-fidelity.md Quant-vs-tool-call-fidelity harness
qwen3-235b-metal-m5max-2026-09-01.md Qwen3 235B-A22B Q2_K-mix Metal validation — M5 Max, 2026-09-01
qwen3-30b-a3b-metal-m5max-2026-08-31.md Qwen3 30B-A3B Q8_0 Metal validation — M5 Max, 2026-08-31
qwen35-precision-2026-08-21.md Qwen3.5-4B precision-vs-fidelity — BF16 vs upstream Q4_K_M (2026-08-21)
reproducible-lora-training-receipts.md Reproducible LoRA training with receipts
review-2026-08-12.md Bug-hunt and simplification review — 2026-08-12
review-2026-08-13.md External-evaluation bug sweep — 2026-08-13
sensitivity-floors-m5max-2026-09-01.md Self-sensitivity floors: 235B and gpt-oss — M5 Max, 2026-09-01
sublayer-removal.md Sublayer removal: --remove-sublayer
tool-choice-boundary-lane.md Tool-choice decision-boundary lane (unlabeled)
train-lora-on-quantized-gguf.md Train a LoRA directly on the quantized GGUF you serve
training-floor-m5max-2026-09-01.md The local training floor, re-measured — M5 Max, 2026-09-01
tray-controller.md Tray / menu-bar controller
tray-v1-report.md Tray controller v1 — validation report (branch tray-controller)
truncation-benchmark.md Truncation-recovery benchmark
truncation-safe-tool-calling.md Tool calls that survive the token limit
windows-remote-checks.md Windows remote-check protocol (learned the hard way, 2026-08-10)
write-side-stall.md Real write-side stall experiment