Drop the verifier/revision pass

The verifier was running a full extra LLM call against draft[:2500]
after every chunk, asking gpt-4o to either say APPROVED or rewrite the
whole excerpt. In practice it almost never caught a real terminology
mistake on flagship note models (gpt-5.1, deepseek-v4-pro, claude-4.x);
its main observable effect was the silent-truncation bug fixed in
v1.0.4 — partial revisions overwriting good drafts.

After the v1.0.4 guard, the verifier was already a no-op for any draft
>2500 chars (most of them at detail=7). Net effect: paying a 5–15s
gpt-4o call per chunk to produce something that gets discarded. Just
remove it.

Removed:
  • VERIFY_MODEL, VERIFY_NOTES constants
  • the "verify" prompt template
  • the post-draft verify block in generate_section
  • VERIFY_MODEL row from gui.py + Electron Settings dropdowns
  • verify-prompt artifact stripping in _clean_artifacts
  • run.py / gui.py / app.js fillGenInfo "Verifier" mentions

The translator-chunking + auto-pick fixes from v1.0.4 stay — those
addressed a real, recurring failure (Chinese translation hitting
gpt-4o's 16K cap). The corresponding regression tests are renamed
to TestVerifierRemoved and now assert the constants/prompt are gone.

Author: nodeeeeee

electron/package.json

@@ -1,6 +1,6 @@
 {
   "name": "auto-note",
-  "version": "1.0.4",
+  "version": "1.0.5",
   "description": "AutoNote — lecture notes generator from Canvas recordings",
   "homepage": "https://github.com/nodeeeeee/Auto-Note",
   "author": {

electron/renderer/app.js

@@ -1056,15 +1056,14 @@ async function loadLectureDropdown(cid) {
 }
 
 async function fillGenInfo() {
-  const [noteModel, verifyModel, target] = await Promise.all([
+  const [noteModel, target] = await Promise.all([
     window.api.getConstant('generate', 'NOTE_MODEL'),
-    window.api.getConstant('generate', 'VERIFY_MODEL'),
     window.api.getConstant('generate', 'QUALITY_TARGET'),
   ]);
   const el = document.getElementById('gen-info-card');
   if (el) {
     el.innerHTML = mkCard(`<div class="info-row">${I.info}
-      <span>Generator: <strong>${esc(noteModel)}</strong>   Verifier: <strong>${esc(verifyModel)}</strong>
+      <span>Generator: <strong>${esc(noteModel)}</strong>
         Quality target: <strong>${esc(target)}</strong></span></div>`);
   }
 }
@@ -1223,21 +1222,6 @@ const CONSTANTS_DEF = [
     // ── Claude CLI (uses `claude -p`, no API key needed) ────────────────
     ['Claude CLI (local)','claude-cli'],
   ]],
-  ['generate', 'VERIFY_MODEL', 'Verification LLM', 'gpt-4.1-mini', [
-    // ── OpenAI ──────────────────────────────────────────────────────────
-    ['GPT-4.1 mini ★','gpt-4.1-mini'],['GPT-4.1 nano','gpt-4.1-nano'],
-    ['GPT-4.1','gpt-4.1'],['GPT-5.1','gpt-5.1'],['o4-mini (reasoning)','o4-mini'],
-    // ── Anthropic ───────────────────────────────────────────────────────
-    ['Claude Haiku 4.5','claude-haiku-4-5-20251001'],['Claude Haiku 3.5','claude-3-5-haiku-20241022'],
-    ['Claude Sonnet 4.6','claude-sonnet-4-6'],['Claude Sonnet 3.5','claude-3-5-sonnet-20241022'],
-    // ── Google Gemini ───────────────────────────────────────────────────
-    ['Gemini 2.5 Flash','gemini-2.5-flash'],['Gemini 2.5 Flash Lite','gemini-2.5-flash-lite'],
-    ['Gemini 2.0 Flash','gemini-2.0-flash'],
-    // ── DeepSeek ────────────────────────────────────────────────────────
-    ['DeepSeek V4 Flash','deepseek-v4-flash'],['DeepSeek V4 Pro','deepseek-v4-pro'],
-    // ── xAI / Mistral ───────────────────────────────────────────────────
-    ['Grok 3 mini','grok-3-mini'],['Mistral Small','mistral-small-latest'],
-  ]],
   ['generate', 'DETAIL_LEVEL',   'Default detail level',  '8',   null],
   ['generate', 'CHAPTER_SIZE',   'Slides per GPT call',   '15',  null],
   ['generate', 'QUALITY_TARGET', 'Self-score target',     '8.0', null],

gui.py

@@ -1712,7 +1712,6 @@ def _on_course_select(e) -> None:
                          value=False, active_color=C_PRIMARY)
 
     note_model   = _read_constant("generate", "NOTE_MODEL")
-    verify_model = _read_constant("generate", "VERIFY_MODEL")
     quality      = _read_constant("generate", "QUALITY_TARGET")
 
     def _run(_) -> None:
@@ -1750,8 +1749,7 @@ def _run(_) -> None:
         _card(ft.Row(controls=[
             ft.Icon(ft.Icons.INFO_OUTLINE, color=C_PRIMARY, size=15),
             ft.Text(
-                f"Generator: {note_model}   Verifier: {verify_model}   "
-                f"Quality target: {quality}",
+                f"Generator: {note_model}   Quality target: {quality}",
                 size=12,
                 color=ft.Colors.with_opacity(0.7, ft.Colors.WHITE),
             ),
@@ -2212,22 +2210,6 @@ def _worker():
             ("Claude Sonnet 4.5",    "claude-sonnet-4-5"),
             ("Claude Haiku 4.5",     "claude-haiku-4-5-20251001"),
         ]),
-        ("generate", "VERIFY_MODEL", "Verification LLM", "gpt-4.1-mini", [
-            # OpenAI
-            ("gpt-4.1-mini",         "gpt-4.1-mini"),
-            ("gpt-4.1-nano",         "gpt-4.1-nano"),
-            ("gpt-4.1",              "gpt-4.1"),
-            ("gpt-5.1",              "gpt-5.1"),
-            # Gemini
-            ("Gemini 3.5 Flash",     "gemini-3.5-flash"),
-            ("Gemini 3.0 Flash",     "gemini-3.0-flash"),
-            ("Gemini 2.5 Flash",     "gemini-2.5-flash"),
-            ("Gemini 2.0 Flash",     "gemini-2.0-flash"),
-            # Anthropic
-            ("Claude Haiku 4.5",     "claude-haiku-4-5-20251001"),
-            ("Claude Sonnet 4.5",    "claude-sonnet-4-5"),
-            ("Claude Sonnet 4.6",    "claude-sonnet-4-6"),
-        ]),
         ("generate",   "DETAIL_LEVEL",         "Default detail level",    "8",    None),
         ("generate",   "CHAPTER_SIZE",          "Slides per GPT call",     "15",   None),
         ("generate",   "QUALITY_TARGET",        "Self-score target",       "8.0",  None),

note_generation.py

@@ -64,9 +64,7 @@
 DETAIL_LEVEL      = 7
 OUTPUT_FORMAT     = "md"
 NOTE_MODEL        = "gpt-5.1"
-VERIFY_MODEL      = "gpt-4o"
 TRANSLATE_MODEL   = "gpt-4o"   # Chinese/other translation post-pass — gpt-4o is cheap enough
-VERIFY_NOTES      = True
 QUALITY_TARGET    = 8.0
 IMAGE_RENDER_SCALE = 1.5
 NOTE_LANGUAGE     = "en"    # "en" = English | "zh" = Chinese
@@ -156,18 +154,6 @@
   Format: `![Slide N](path) *(caption)*` or `![Frame N](path) *(caption)*`
 - Code examples must be complete and compilable, using the correct language tag (```c, ```cpp, ```python).
 """,
-verify="""\
-Check the following note excerpt for technical terminology consistency with the slides, and for any obvious factual errors.
-
-**Reference glossary (from slides):**
-{term_list}
-
-**Note excerpt:**
-{draft}
-
-If there are no issues, reply APPROVED (this word only).
-If there are terminology or factual errors, return the corrected full note excerpt with no explanation.
-""",
 exam="""\
 Below are the complete lecture notes for {course_name}. Please append a concise exam cheat-sheet section at the end.
 
@@ -845,14 +831,6 @@ def _clean_artifacts(text: str) -> str:
     cleaned = []
     for line in lines:
         stripped = line.strip()
-        # Remove bare "APPROVED" lines (verifier leak)
-        if stripped == "APPROVED":
-            continue
-        # Remove verify prompt leaks
-        if "reply APPROVED (this word only)" in stripped:
-            continue
-        if "return the corrected full note excerpt" in stripped:
-            continue
         if "terminology or factual errors" in stripped:
             continue
         # Clean section-header artifacts
@@ -1116,44 +1094,6 @@ def generate_section(
         heading = f"### {lec_num}.{ci} {_chunk_title(chunk)}"
         return f"{heading}\n\n*(Section could not be generated — re-run with force to retry.)*", True
 
-    if VERIFY_NOTES and draft:
-        terms = set()
-        for s in chunk:
-            for t in re.findall(r"\b[A-Z][a-zA-Z]{3,}\b|\b[A-Z]{3,}\b", s.text):
-                terms.add(t)
-        term_list = ", ".join(sorted(terms)[:30])
-        # The verifier only sees the head of the draft. If the draft is
-        # longer than that window, accepting v_result as the new draft
-        # would silently truncate everything past the window — exactly
-        # the bug that produced the ~2KB section files. Cap the window
-        # and refuse to overwrite drafts the verifier never fully saw.
-        VERIFY_INPUT_CAP = 2500
-        verifier_saw_all = len(draft) <= VERIFY_INPUT_CAP
-        v_user = _P("verify").format(term_list=term_list,
-                                     draft=draft[:VERIFY_INPUT_CAP])
-        # Always route the verify/revision pass through VERIFY_MODEL
-        # (gpt-4o) — cheap and fast, avoids burning codex quota on a
-        # short review call. If OpenAI is unavailable (no key, quota
-        # exceeded, network issue), skip verification rather than losing
-        # the generated draft.
-        _vmodel = VERIFY_MODEL
-        try:
-            v_result = _call(_vmodel, "", v_user, 1500)
-            if not v_result.strip().upper().startswith("APPROVED"):
-                if not verifier_saw_all:
-                    # Long draft — refuse to overwrite with a verifier
-                    # revision that only saw the first 2500 chars.
-                    tqdm.write(f"  [warn] Verifier flagged issues but draft "
-                               f"({len(draft)} chars) exceeds verify window — "
-                               f"keeping full draft as-is.")
-                elif len(v_result) > len(draft) * 0.5:
-                    draft = v_result
-                else:
-                    tqdm.write(f"  [warn] Verifier suspicious response, keeping draft")
-        except Exception as _ve:
-            tqdm.write(f"  [warn] Verify pass failed ({type(_ve).__name__}: "
-                       f"{str(_ve)[:120]}) — keeping draft")
-
     # Strip pipeline artifacts that may have leaked into the draft
     draft = _clean_artifacts(draft)
 

run.py

@@ -694,7 +694,6 @@ def _const(script: str, name: str, desc: str) -> None:
     _const("align",      "OFF_SLIDE_THRESHOLD",  "Min cosine to stay on-slide")
     _const("align",      "PRIOR_SIGMA",          "Temporal prior width (slides)")
     _const("generate",   "NOTE_MODEL",           "LLM for note generation")
-    _const("generate",   "VERIFY_MODEL",         "LLM for verification")
     _const("generate",   "DETAIL_LEVEL",         "Default detail level (0-10)")
     _const("generate",   "CHAPTER_SIZE",         "Slides per GPT call")
     _const("generate",   "QUALITY_TARGET",       "Self-score target for --iterate")

test/test_truncation_fixes.py

@@ -85,64 +85,22 @@ def test_passes_through_cli_models(self, monkeypatch):
         assert ng._pick_translate_model() == "codex-cli"
 
 
-class TestVerifyOverwriteGuard:
-    """The verifier sees only ``draft[:2500]``. When the draft is longer
-    than that, accepting v_result as the new draft truncates everything
-    past the verify window. The new behavior preserves the full draft
-    and only allows the verifier to overwrite when it saw all of it.
+class TestVerifierRemoved:
+    """The verifier/revision pass was removed in v1.0.5 — modern flagship
+    note models rarely make terminology mistakes worth a separate review
+    call, and the limited verify window had been silently truncating
+    long drafts. Make sure no constant or prompt template is left over.
     """
 
-    def test_long_draft_kept_intact_when_verifier_disagrees(self, monkeypatch):
-        # Construct a draft longer than VERIFY_INPUT_CAP=2500. The mock
-        # verifier returns a "revised" version that is shorter — under
-        # the old logic it would replace the draft and silently truncate
-        # the tail. Under the fix, the original draft is kept.
+    def test_constants_gone(self):
         import note_generation as ng
+        assert not hasattr(ng, "VERIFY_NOTES")
+        assert not hasattr(ng, "VERIFY_MODEL")
+
+    def test_verify_prompt_not_in_template(self):
+        import note_generation as ng
+        # Both language tables should have lost the "verify" key
+        assert "verify" not in ng._PROMPTS["en"]
+        if "zh" in ng._PROMPTS:
+            assert "verify" not in ng._PROMPTS["zh"]
 
-        long_draft = "Sentence about TCP. " * 200    # ~3800 chars
-        revised = "TCP is a protocol. " * 50          # ~950 chars
-
-        captured = {"draft": long_draft}
-
-        def fake_call(model, system, user, max_tokens, _truncated=None):
-            # Verifier path
-            if "Reference glossary" in user or "Reference Glossary" in user:
-                return revised
-            # Translator path (skip — language is en in this test)
-            return captured["draft"]
-
-        # Build the minimal context generate_section needs, then assert
-        # the draft did not collapse to ~revised. Easier: just exercise
-        # the guard logic directly — patch _call and call generate_section
-        # would require a full LectureData mock. Instead, replicate the
-        # post-verify fragment here:
-        VERIFY_INPUT_CAP = 2500
-        draft = long_draft
-        verifier_saw_all = len(draft) <= VERIFY_INPUT_CAP
-        v_result = revised
-
-        if not v_result.strip().upper().startswith("APPROVED"):
-            if not verifier_saw_all:
-                # The fix: refuse to overwrite
-                pass
-            elif len(v_result) > len(draft) * 0.5:
-                draft = v_result
-
-        assert draft == long_draft, (
-            "Long draft must NOT be replaced by a partial verifier revision"
-        )
-
-    def test_short_draft_can_be_replaced_by_verifier(self):
-        # When the verifier saw the entire draft, replacement is safe.
-        VERIFY_INPUT_CAP = 2500
-        draft = "Short draft about UDP." * 5           # ~110 chars
-        v_result = "UDP is a connectionless protocol." * 5  # ~165 chars
-
-        verifier_saw_all = len(draft) <= VERIFY_INPUT_CAP
-        if not v_result.strip().upper().startswith("APPROVED"):
-            if not verifier_saw_all:
-                pass
-            elif len(v_result) > len(draft) * 0.5:
-                draft = v_result
-
-        assert draft == v_result