English measurement pages have one data source: session files on the machine that ran the work. The path and the parsing method are stated on the page; it is not 'we ran it, trust us'.
Three layers of traceability
| Layer | Where | What it lets you verify |
|---|---|---|
| Raw events | ~/.dsh/sessions/<workspace>/<uuid>/session.jsonl.zstd | Every event carries a millisecond time and a seq; token usage sits in the usage chunk of assistant/chunk |
| Parsing scripts | tools/usage.py / cost.py / latency.py | How tokens, cost, and time-to-first-token are derived from events |
| Page figures | The raw-data tables on each /lab/ page | They correspond item by item to script output |
How the English and Chinese pages relate
- One dataset, two language versions — the English page does not quote figures out of the Chinese one.
- The Chinese page is the natively written expression, and governs when the wording is ambiguous.
- To recompute: take the scripts and your own session files and get your own numbers — more useful than checking ours.
Why the raw JSON is not published
Session files contain absolute paths, working directory names, and occasionally prompt fragments. They stay off public pages to avoid leaking details about the machine, not because the figures are uncheckable. What the page gives is the computed values, the script used, and the event type names. Run it against your own session and compare.
This is where 'measured' parts company with 'relayed' here: a relay gives you conclusions, a measurement gives you a reproduction path. English and Chinese pages do not differ on this — both give the path.