Save and review evidence

Use the retained data to review an experiment without rerunning it. Native suite evidence lives in ignored .zenbench/data/, or ZENBENCH_DATA_DIR when configured. It includes immutable hash-addressed gzip objects, process attempts and small references. Selected summaries are for review; raw packets are for reconstruction and inference.

Give a suite attempt a clear path

For a retained zenbench benchmark executable:

ZENBENCH_RESULT_PATH="$PWD/candidate-01.json.gz" ./builds/candidate/my_bench --group=sort
zenbench data read candidate-01.json.gz
zenbench data read candidate-01.json.gz --format json

Use a fresh output path for every attempt. data read accepts JSON, gzip JSON or an evidence reference and formats it as llm, json, csv, md or suite html. It does not run the benchmark. The same reader understands retained external pair packets for the text formats.

In PowerShell, set $env:ZENBENCH_RESULT_PATH before the executable invocation; see retaining builds.

Export charts and reports

Charts for slides and reports covers SVG, fixed light/dark themes, vertical grouped bars and self-contained HTML. Export from retained native suite evidence with zenbench data read candidate-01.json.gz --save-charts figures; no new measurement is needed.

Choose the correct comparison

data compare cannot turn Iterate data into confirming evidence. Inspect exploratory packets with data read. Match workload, inventory, settings, placement and testbed; an identity mismatch is a reason to investigate the experiment rather than edit the stored evidence.

Export one stable summary for Git

For a selected native suite packet:

mkdir -p benchmarks
zenbench data snapshot candidate-01.json.gz --output benchmarks/baseline.csv

Commit this deliberate snapshot and a pointer to the raw evidence and build manifest. Avoid exporting every exploratory attempt into Git. Snapshots retain evidence identity and read/compare command guidance. CSV is not an input to confirmation; it does not contain all process/round samples.

External pair attempts already have selected.csv. Their Confirm decision is retained in the pair packet; pair-diff remains descriptive. Iterate does not become confirming data when read later. Copy the chosen rows with a packet pointer rather than using the suite snapshot command. Fleet uses zenbench fleet export study.jsonl ATTEMPT...; the fleet guide explains its receipt checks.

Archive and reanalyze without measuring again

Losslessly compact an older suite JSON packet into a new gzip file:

zenbench data archive candidate.json --output candidate-archive.json.gz

Replay a complete suite packet's frozen process analysis into a separate artifact:

zenbench data reanalyze candidate-confirm.json.gz --output analysis-01.json

Reanalysis references the immutable packet hash. It cannot promote exploratory evidence or repair missing samples. Keep the original packet and policy alongside the new analysis.

Keep the objects that references point to

Copy full gzip objects, child attempts, runtime assets and retained binaries to durable artifact storage. Verify hashes when moving them. A reference file or CSV alone cannot recreate the data. Preserve named baseline references and their objects across CI jobs; set ZENBENCH_DATA_DIR to a persisted location when needed. See portable storage for reference resolution and retained process lifecycle records.

Build-time identity remains your responsibility: runtime Git HEAD is not the source used to build a retained executable. Bind source revision, dirty patch, compiler, target, flags, dependencies and input hashes to the binary hash.

Keep invalid, cancelled and inconclusive attempts too. Distinguish a planned replicate from a retry; do not select only the fastest successful receipt. You have a reviewable result when a reviewer can identify the intended experiment, find the selected summary and follow its pointers back to complete evidence.