Skip to content

v0.7.3 — 字段测试修复 Field-fix train

Choose a tag to compare

View all tags
@mianaz mianaz released this 31 Jul 06:10
· 78 commits to main since this release
v0.7.3
a091889

字段测试修复版 · Field-fix release

First release driven by a real field-test cycle — six meeting transcripts, 43 dictionary cards, 12 of them wrong.

Fixed

  • API keys are now per-provider. Switching the LLM provider previously transmitted the previous provider's key to the new provider's host. Keys are isolated per provider, migrated automatically, and stay in the macOS Keychain.
  • Speaker rename. Renaming a speaker left the roster and the live picker stuck on the old name, so every later segment kept the stale label — in one session for 50 minutes.
  • Jargon card context. Cards quoted whole 60-word ASR run-ons; they now narrow to the clause around the match.
  • False positives. "in the red" no longer fires on "the red graph"; broken dictionary glosses (mean shipped as literally "1.") repaired at source with a build-time lint.
  • Meeting-domain inference. "thank you for your attention" no longer produces an attention card in a biology talk, while precision/recall still fire in an ML talk.
  • Wrong-sense cards. CAC now resolves to Cancer-Associated Cachexia rather than 获客成本, NDA to New Drug Application, PR to precision-recall.

Notes

  • No binaries attached: DMG signing in CI is still pending the Apple-credentials decision, and this tag predates the release workflow. From v0.7.4 onward artifacts are built and attached automatically.

Commits

  • Test-count badges 4745 -> 4761 (item-1 domain inference lane)
  • feat(detect): infer meeting domains for common jargon
  • v0.7.3 version stamp (8 sites + both lockfiles)
  • fix: reject grammatical false positives for prior/mean/attention/recall
  • Test-count badges 4718 -> 4745 (v0.7.3 field-fix train)
  • fix: isolate LLM API keys by provider
  • fix literal expression context collisions
  • fix detection context and expression false positives
  • fix: close dangling-colon gloss debt for four ml-stats terms
  • fix: reject broken dictionary glosses and repair mean/precision/recall
  • fix speaker rename leaving roster/latch stuck on the old name
Assets 2
Loading