Skip to the content.

Hardware bench plan — 20260903T053000Z

Auto-generated by tools/bench_plan.py --md. Do not hand-edit; re-run the tool (or the FleetBenchPlanDoc task) to refresh.

PLAN ONLY – no benchmark was run; every figure is a literal catalog fact and all ages are computed against –now.

Generated from catalog.json (last_updated 2026-09-03T05:30:00Z): 7 bench-nodes, 70 feasible cells (23 empty coverage holes, 3 with a recorded baseline), 7 infeasible cells, 95 runs on record.

Coverage matrix (bench-node × workload-kind)

. = empty feasible cell (a coverage hole) · - = infeasible on this hardware (e.g. a CUDA kind on the Apple mac) · a number = runs on record.

machine model gpu qwen36 radix session fan agent-live turn-tax parity livecodebench concept-bench
a100 . . . . . . . . . - .
cpu-server-a . - . . . . . . . - .
gcp-a3-high-h100-1g 1 1 1 1 1 1 1 1 1 - 1
gcp-g2-l4 1 1 1 1 1 1 1 1 1 - 1
gcp-g2-l4-32 2 1 1 1 1 1 1 1 1 - 1
node-macos-a 3 - 5 1 1 . 1 . 1 . .
workstation-a 5 1 9 2 1 2 5 2 4 2 1

Next test per machine

machine next workload model / precision intent score why
a100 gpu-benchmark qwen2.5-3b / Q8_0 learn-collect 0.8356 a100 has 0 runs (last_run=None); the gpu-benchmark cell has never run – a first-ever measurement on this node.
cpu-server-a model-benchmark SmolLM2-135M-Instruct / q8_0 coverage 0.7326 cpu-server-a has run other workloads but the model-benchmark cell is empty – a coverage gap to fill (SmolLM2-135M-Instruct).
gcp-a3-high-h100-1g model-benchmark SmolLM2-135M-Instruct / q8_0 benchmark 0.3607 gcp-a3-high-h100-1g/model-benchmark ran 1x, last 50.24d ago (re-check 7d) but recorded no number – re-run to capture one.
gcp-g2-l4 model-benchmark SmolLM2-135M-Instruct / q8_0 benchmark 0.3583 gcp-g2-l4/model-benchmark ran 1x, last 50.24d ago (re-check 7d) but recorded no number – re-run to capture one.
gcp-g2-l4-32 qwen36 qwen3.6-27b / q8 benchmark 0.3594 gcp-g2-l4-32/qwen36 ran 1x, last 50.19d ago (re-check 14d) but recorded no number – re-run to capture one.
node-macos-a livecodebench qwen3.6-27b / official coverage 0.6287 node-macos-a has run other workloads but the livecodebench cell is empty – a coverage gap to fill (qwen3.6-27b).
workstation-a model-benchmark SmolLM2-135M / q8_0 regression 0.4307 workstation-a/model-benchmark has a recorded 31.02 tok/s baseline aged 50.2d (re-check every 7d) – re-measure to catch drift.

Do next — global ranked plan

# machine workload intent score suggested command (hint)
1 a100 gpu-benchmark learn-collect 0.8356 go run -tags cuda ./cmd/gpucheck
2 a100 model-benchmark learn-collect 0.8326 go run ./cmd/modelbench -quant
3 a100 radix-benchmark learn-collect 0.8326 go run ./cmd/radixbench
4 a100 qwen36 learn-collect 0.812 fak serve + fak agent (qwen3.6-27b via gateway)
5 a100 agent-live learn-collect 0.77 go run ./cmd/fak agent --task <task>
6 a100 concept-benchmark learn-collect 0.77 go run ./cmd/conceptbench --replay cmd/conceptbench/testdata/replay
7 a100 fan-benchmark learn-collect 0.77 go run ./cmd/fanbench
8 a100 parity learn-collect 0.77 go run ./cmd/paritybench
9 a100 session-benchmark learn-collect 0.77 go run ./cmd/sessionbench
10 a100 turn-tax learn-collect 0.77 go run ./cmd/fak turntax --suite turntax-airline

By intent

The planner guarantees all four operator intents are surfaced.

Benchmark perf

Learn / collect new data

Prevent regression

Fill coverage gaps

Frontier & honesty call-outs

Excluded (not a bench target)

Methodology & honesty

Each feasible (machine × workload-kind) cell is scored on five dimensions, each in [0,1]:

dimension weight intent reads
coverage_gap 0.34 coverage run count in the cell (empty = 1.0)
machine_novelty 0.20 learn-collect machine total runs / last_run
staleness_overdue 0.16 regression newest cell timestamp vs –now, per-kind interval
baseline_drift 0.16 regression a recorded peak_tok_per_sec aging past its interval
model_diversity 0.14 coverage anti-monoculture: how rare the proposed model is

Cells sort empty-tier-first (coverage dominance), then by weighted score. Feasibility (no CUDA on the mac) is applied before scoring. --now is injected, never read from the wall clock, so a fixed stamp yields identical output. marginal-information-gain (novelty / diversity) is a heuristic proxy, not a measured quantity — the ranking is a guide, not a verdict. This tool only PLANS; a run is a later action on the remote bench-node.