Commits on May 23, 2026

1. v0.0.20 — self-review pass (2nd LLM call) + L3.5 routine …

   User feedback after v0.0.19's prompt rewrite: "继续很深入改". Deeper
   this time: not just better prompts, but a structural change to the
   pipeline itself.
   
   ## The biggest single quality multiplier we can ship
   
   A second LLM call on the stitched output catches 20-30% more L3 errors
   that the first pass missed. This is the single biggest quality lift
   short of changing the model. Default ON.
   
   ## Self-review (Stage 6) — wired into iter_events
   
   After all chunks complete:
   
   1. Build a payload of {edited_markdown, change_log, library_context,
      briefing}
   2. Call the model with the self_review system prompt
   3. Parse JSON: {additional_corrections, rollbacks,
      promotions_to_user_review, data_conflicts, format_issues}
   4. Apply additional_corrections to the markdown via string replace
      (only if `old` actually appears in the doc — model hallucination
      guard)
   5. Append review changes to change_log with stage="self_review"
   6. Surface diagnostics on the `complete` event's `self_review` field
   
   Cost: 1 extra LLM call per Run regardless of chunk count. Auto-skipped
   when stitched output > 100k chars to bound cost on monster transcripts.
   
   SSE events: self_review_start / self_review_done / self_review_error.
   UI shows "↻ Self-review — re-reading for missed corrections…" then
   "+N fixes · M data conflicts to check · K ambiguous items flagged".
   
   Pipeline param: `enable_self_review: bool = True`. Tests that exercise
   chunking/main-edit alone opt out explicitly.
   
   ## Self-review prompt (06_self_review.md)
   
   Rewritten from a thin checklist into a 5-check routine:
   
   1. Proper noun audit (the headline reason; first pass misses 20-30%)
   2. Speaker consistency across full document
   3. Cross-section data consistency (ARR / headcount / funding agree)
   4. Format hygiene (leftover [Speaker N], mixed punctuation, etc.)
   5. Over-correction rollbacks (confidence < 0.7 sanity check)
   
   Hard JSON output schema. "Don't second-guess high-confidence work."
   
   ## L3.5 rewritten
   
   Same treatment as L3 got in v0.0.19. The conservative L3.5 table
   now wraps a mandatory 7-check routine: sentence boundary correctness,
   stutter dedup (exact X X only), word-order garbling, missing function
   words, speaker-switch swallowed mid-paragraph, number/letter
   confusion in spoken digits, same-sound substitution destroying
   meaning (cohort vs co-host).
   
   Hard confidence threshold table per change type. Explicit "what
   L3.5 does NOT do" list keeps the model from over-correcting.
   
   ## Tests
   
   284 → 292 (+8). All passing. Ruff clean.
   
   test_pipeline_self_review.py covers: event ordering, additional
   corrections actually applied, token counts include both passes,
   opt-out flag, auto-skip for huge output, garbage response doesn't
   crash, diagnostics surfaced in complete, ignore corrections where
   old not in doc (hallucination guard).
   
   Existing tests that asserted exact token counts updated to reflect
   the new 2-pass total.

   

   Chen17-sq committed May 23, 2026
   Configuration menu

   • View commit details
   • Copy full SHA for d6af598
   • Browse repository at this point

   Copy the full SHA

   d6af598 View commit details

   Browse the repository at this point in the history