mirror of
https://github.com/block/buzz.git
synced 2026-11-07 18:01:06 +00:00
5c0fc2e78a
Provenance was always engine-owned and deterministic — shown_ids and the coverage window come from the plan, never from model output. The citation-subset refusal was a second layer on top whose failure mode (refuse the whole paid run) fired on model formatting slips: rewriting prior Log entries non-verbatim, or citing an event id quoted inside a message's text. Both hit in dogfood. validate_output now checks only the H1 section structure, which the append-splice and the reader depend on. Citations stay in the prompt as guidance and render as verifiable links; a wrong id is a visible dead chip instead of a lost run. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Riley Crane <rileycrane@squareup.com>