Skip to content

Errata

The campaign plan said finish reasons would be retained. Successful anvil-serving.benchmark/v1 timing rows retain first-output, first-visible, E2E, usage-derived token totals, chunk counts, and capture-truncation state. They do not retain finish reasons, terminal form, raw usage objects, or visible character count. The bundle therefore claims runner-validated visible output and terminal observation only, never stop validation or semantic correctness.