This folder is a flat legacy archive of benchmark summaries, traces, LocalMaxxing payloads/responses, and diagnostic output. Do not reorganize old files by moving them; many notes and patch records link to exact paths.
Model-specific result packets should link to these files rather than duplicate them. Start with ../results/README.md for promoted packets and ../docs/model-effort-index.md for the cross-model map.
Track compact artifacts that make a note reproducible:
*-summary-*.json: benchmark identity, metrics, and quality result.*-timing-decision-*.md and *-timing-decision-*.json: routing artifacts
used to choose the next target.*.jsonl traces when a note interprets a specific gate or state
transition.Usually leave these local unless a note needs them:
.log files.*-xpu-health.log dumps.For new Gemma 4 26B Q8 runs, prefer the existing run-directory shape produced
by scripts/run-gemma4-26b-first-baseline.sh /
scripts/run-gemma4-26b-mtp-candidate.sh:
data/gemma4-q8-gpu<N>-<variant>-<stamp>/summary.json
data/gemma4-q8-gpu<N>-<variant>-<stamp>/chat-canary.json
data/gemma4-q8-gpu<N>-<variant>-<stamp>/p512o512.json
For new Qwen runs, prefer:
qwen36-ablation-<label>-summary-<stamp>.json
qwen36-ablation-<label>-p<prompt>o<output>-<stamp>.json
qwen36-<purpose>-trace-<label>-<stamp>.jsonl
Keep the label stable across logs, summaries, and notes. If a run is invalid, say why in the note instead of deleting the artifact.
In mixed experiment worktrees, stage explicit paths. Do not use broad
git add -A or git add data/.
Good default:
git add -- data/*summary-*.json notes/<date>-<topic>.md patches/<patch>.patch