Module media_integrity

Module media_integrity 

Source
Expand description

2026-09-26: Cross-request and cross-path integrity checks, used by the vision and video benchmarks. They target machinery a single well-formed request cannot reach, whose failure is a fluent answer to the wrong input rather than an error:

  • heterogeneous_concurrency: several different requests in flight at once, each required to get its own answer.
  • cache_leak: the same prompt text with a different image, back to back.
  • long_prompt_path: the same image and question with filler prepended.
  • media_in_history: media in an earlier turn or a tool result.
  • stream_parity: the same request streamed and blocking.
  • multi_choice: n > 1 with an image.
  • responses_parity and thinking_parity (the responses submodule).

Owner: bench, media integrity. Invariants: a check whose request failed never returns Cell::Pass.

Enums§

Cell
2026-09-26: Outcome of one integrity check.

Functions§

cache_leak
2026-09-26: The same prompt text with a different image, back to back.
heterogeneous_concurrency
2026-09-26: Several different requests at once, each scored against its own expected answer.
image_request
2026-09-26: One image part plus a prompt, as a streaming chat body with thinking off.
long_prompt_path
2026-09-26: The same image and question, short and with a large block of filler text prepended; both replies must satisfy want.
media_in_history
2026-09-26: Media that is not in the message asking the question: in an earlier turn or in a tool result. The server collects media from every message, tool results included (api/chat/msg_entry.rs, collect_message_media); a regression that scanned only the final message would pass the legs that put the image and the question in one turn.
multi_choice
2026-09-26: n > 1 with an image, on the blocking path.
responses_image_request
2026-09-26: A blocking Responses-API request carrying one image.
responses_parity
2026-09-26: The Responses API sees the image, and sizes it the same.
stream_parity
2026-09-26: The same request streamed and not streamed.
thinking_image_request
2026-09-26: The same image request with thinking on.
thinking_parity
2026-09-26: Vision with thinking on, through chat completions.

Type Aliases§

Subject
2026-09-26: One concurrent subject: the request, a predicate scoring its own reply, and a label for the failure message.