2.88.1

2.88.1

Changed

• The blocked-agent list moved from the Authenticated Feeds card to Security.
  The gate runs whether or not feed auth is enabled, so the old placement
  implied a dependency that does not exist.

Fixed

• Held verification conflicts appeared twice in the review queue. 2.88.0 added
  each one to both the held list and the UI list, so the marker saved twice and
  rendered as duplicate cards that shared playback state. They now go to the
  held list only, which is what the function has always documented, and the
  merge that concatenates those two lists drops repeated spans so no future
  path can persist the same pass-2 marker twice.
• deduplicate_patterns picked the surviving row by confirmation count alone,
  so an auto-learned pattern could beat and delete one an operator created by
  hand. Tier now outranks confirmation count, and merged stats still carry the
  group total.

2.88.0

Added

• A Settings list of user agents that must never trigger just-in-time
  processing. A matching request gets a 302 to the origin audio instead of
  a queued transcription and detection run. Matching is case-insensitive
  and looks for the pattern anywhere in the agent string; start a pattern
  with ^ to anchor it to the beginning, which short strings need so they
  cannot match mid-agent. The list is empty by default, so upgrading
  changes nothing until an operator adds a pattern. Closes #645.

Tooling (benchmark; not in runtime image)

• Six models added to the sweep roster: bytedance-seed/seed-2-1-turbo,
  deepseek/deepseek-v4-pro-0813, qwen/qwen3.8-2.4t-a95b, x-ai/grok-4.6,
  nvidia/nemotron-3.5-lightning, and meta/muse-glimmer-30b. Slugs verified
  against the live OpenRouter catalog.
• prompts/2026-08.txt: the 2026-08 campaign's system prompt, frozen. The
  cross_promo wording change in 2.86.2 altered DEFAULT_SYSTEM_PROMPT, and
  since prompt_hash is part of the work-unit key, every one of the campaign's
  64,125 rows would otherwise be treated as incomplete. Pinning the run to this
  file adds new models to an existing campaign without re-running the old ones.
• Fresh pricing snapshot covering the added models.
• Regenerated report and charts: 81 active models, 69,255 work units, 69,125
  scored. x-ai/grok-4.6 joins the top statistical tier at F0.5 0.760; the
  other five land between 0.052 and 0.650 and none displaces an existing
  recommendation. docs/llm-providers.md refreshed to match.

Changed

• huggingface-hub 1.26.1 to 1.27.0, which pulls in hf-xet 1.5.1 to 1.6.0.
• The report's "Errors resolved by retry" section is gone. Every row in it had
  already succeeded, so none affected a score, and its contents were either
  duplicated elsewhere or noise: the largest entry restated what the JSON mode
  column already says, and thirteen of fifteen rows covered under 1% of a
  model's work units. Unresolved failures keep their own tables, which is where
  a provider refusal or a rate-limit ceiling actually belongs.
• Provider-policy blocks no longer classify as Unknown model (404). OpenRouter
  answers 404 when an account blocks a provider, so a correct slug read as a
  wrong one; these now bucket as Account gating (provider policy).

Fixed

• A verification finding that contradicts a kept pass-1 span is now held for
  review instead of being discarded. The keep still stands, so pass 2 never
  cuts through an operator's segment-action choice, but the disagreement is
  visible and one approval away from a cut. Previously it was dropped with
  only a debug line to show for it, so an ad the second pass had caught at
  high confidence vanished silently.
• The "re-cutting pass 1 output" log fired before the filters that decide
  whether anything gets re-cut, so it announced work that often never
  happened. It now reports the actual count, after the gate.
• benchmark run --dry-run ignored --retry-errors, reporting errored units as
  skipped and under-counting the real queue. The preview now takes the same
  errored-key set the run does.