Fix phrase span leaks, safer length budget, disable phrase T5 by default.

Reject object-front and duplicate-verb span candidates, keep length budget from dropping source content, and bake ENGINE_PHRASE_USE_T5=false into Space defaults.

Co-authored-by: Cursor <cursoragent@cursor.com>

.env.example

@@ -27,13 +27,29 @@ ENGINE_SAFETY_MIN=0.80
 ENGINE_USE_MINILM_SAFETY=true
 MINILM_MODEL=sentence-transformers/all-MiniLM-L6-v2
 
-# Local CPU paraphraser for uncovered sentences (preferred over fixed clefts)
+# Local CPU paraphraser for meaning-safe but newer-looking rewrites
 ENGINE_PARAPHRASE=true
+ENGINE_PARAPHRASE_PRIMARY=true
 ENGINE_PARAPHRASE_MODEL=Vamsi/T5_Paraphrase_Paws
 ENGINE_PARAPHRASE_MIN_SIM=0.72
-ENGINE_PARAPHRASE_NUM_RETURN=3
-ENGINE_PARAPHRASE_MAX_NEW_TOKENS=64
-ENGINE_PARAPHRASE_MAX_SURFACE=0.92
+ENGINE_PARAPHRASE_NUM_RETURN=5
+ENGINE_PARAPHRASE_MAX_NEW_TOKENS=72
+ENGINE_PARAPHRASE_MAX_SURFACE=0.88
+ENGINE_PARAPHRASE_MIN_DIVERGENCE=0.14
+
+# Phrase-level rewrite (verb–object spans via WordNet hyponyms + optional T5)
+ENGINE_PHRASE_REWRITE=true
+# 0 = dynamic (1 normally, 2 with polish). Set 1–4 to hard-cap.
+ENGINE_PHRASE_MAX_CHANGES=0
+ENGINE_PHRASE_MIN_SIM=0.74
+# Keep T5 span paraphrase off by default (WordNet phrase path is safer).
+ENGINE_PHRASE_USE_T5=false
+
+# Keep rewrite length near the source; trim only when clearly bloated.
+ENGINE_PRESERVE_LENGTH=true
+# Split long sentences on clause joins during structural fallback.
+ENGINE_SPLIT_LONG=true
+ENGINE_SPLIT_MIN_WORDS=16
 
 # Legacy fixed cleft fallback (off by default)
 ENGINE_FORCE_REWRITE=false
@@ -43,7 +59,9 @@ ENGINE_REQUIRE_WORDING_CHANGE=true
 
 # Low-impact vocabulary refinement after structure/paraphrase
 ENGINE_LEXICAL_REFINEMENT=true
-ENGINE_LEXICAL_MAX_CHANGES=3
+# 0 = fully dynamic synonym budget from sentence length (recommended).
+# Set >0 only to hard-cap the dynamic budget (max 15).
+ENGINE_LEXICAL_MAX_CHANGES=0
 ENGINE_LEXICAL_MIN_WSD=0.14
 ENGINE_LEXICAL_MIN_ZIPF=4.0
 ENGINE_LEXICAL_MAX_FREQUENCY_GAP=0.60

Dockerfile

@@ -20,7 +20,11 @@ ENV PYTHONDONTWRITEBYTECODE=1 \
     ENGINE_PARAPHRASE=true \
     ENGINE_FORCE_REWRITE=false \
     ENGINE_USE_MINILM_SAFETY=true \
-    ENGINE_LEXICAL_REFINEMENT=true \
+    ENGINE_LEXICAL_REFINEMENT=true \
+    ENGINE_PHRASE_REWRITE=true \
+    ENGINE_PHRASE_USE_T5=false \
+    ENGINE_PRESERVE_LENGTH=true \
+    ENGINE_SPLIT_LONG=true \\
     HF_HOME=/opt/hf_cache \
     TRANSFORMERS_CACHE=/opt/hf_cache \
     LANGUAGETOOL_HOME=/opt/languagetool \

app/config.py

@@ -209,6 +209,9 @@ ENGINE_PHRASE_MIN_SIM = max(
     0.0,
     min(float(os.environ.get("ENGINE_PHRASE_MIN_SIM", "0.74") or "0.74"), 1.0),
 )
+# T5 span paraphrase is optional; WordNet phrase swaps are the default wording path.
+_ephr_t5 = (os.environ.get("ENGINE_PHRASE_USE_T5") or "false").strip().lower()
+ENGINE_PHRASE_USE_T5 = _ephr_t5 in {"1", "true", "yes", "on"}
 
 # Keep rewritten length near the source (trims only when clearly bloated).
 _epl = (os.environ.get("ENGINE_PRESERVE_LENGTH") or "true").strip().lower()

app/engine/mechanics.py

@@ -22,7 +22,8 @@ def enforce_length_budget(
 ) -> str:
     """Trim only when the rewrite clearly ballooned past the original.
 
-    Preserves paragraph breaks (\\n\\n) — never flattens the document.
+    Preserves paragraph breaks (\\n\\n). Never trims so hard that most of the
+    source content disappears.
     """
     source = original or ""
     output = rewritten or ""
@@ -31,8 +32,11 @@ def enforce_length_budget(
     if source_words == 0:
         return output
 
-    max_ratio = 1.08 if preserve_length else 1.6
+    # Slightly looser ceiling so normal phrase/lexical growth is kept.
+    max_ratio = 1.20 if preserve_length else 1.6
     max_words = max(1, int(source_words * max_ratio))
+    # Floor: do not destroy content just to hit the ceiling.
+    min_words = max(1, int(source_words * 0.90))
     if output_words <= max_words:
         return output
 
@@ -59,4 +63,8 @@ def enforce_length_budget(
             kept_paragraphs.append(" ".join(kept_sentences))
         if count >= max_words:
             break
-    return "\n\n".join(kept_paragraphs).strip() or output
+    trimmed = "\n\n".join(kept_paragraphs).strip()
+    if not trimmed or len(trimmed.split()) < min_words:
+        # Trimming would drop too much meaning — keep the full rewrite.
+        return output
+    return trimmed

app/engine/orchestrator.py

@@ -18,6 +18,7 @@ from app.config import (
     ENGINE_PHRASE_MAX_CHANGES,
     ENGINE_PHRASE_MIN_SIM,
     ENGINE_PHRASE_REWRITE,
+    ENGINE_PHRASE_USE_T5,
     ENGINE_PRESERVE_LENGTH,
     ENGINE_REQUIRE_WORDING_CHANGE,
     ENGINE_SAFETY_MIN,
@@ -211,6 +212,7 @@ def _apply_phrase_rewrite(
         max_changes=max_changes,
         polish=polish,
         min_sim=min(ENGINE_PHRASE_MIN_SIM, safety_min),
+        use_t5=ENGINE_PHRASE_USE_T5,
     )
     if not refined.changes or refined.text == before:
         return record

app/engine/paraphrase/__init__.py

@@ -414,19 +414,35 @@ def _span_candidate_ok(source: str, candidate: str) -> bool:
         return False
     if ":" in cand or ";" in cand:
         return False
-    src_tokens = _WORD.findall(src)
-    cand_tokens = _WORD.findall(cand)
+    src_tokens = [token.lower() for token in _WORD.findall(src)]
+    cand_tokens = [token.lower() for token in _WORD.findall(cand)]
     if not src_tokens or not cand_tokens:
         return False
     # Keep span length close so splicing stays grammatical.
-    if len(cand_tokens) > len(src_tokens) + 2 or len(cand_tokens) + 2 < len(src_tokens):
+    if len(cand_tokens) > len(src_tokens) + 1 or len(cand_tokens) + 2 < len(src_tokens):
         return False
-    if len(cand) > max(12, int(len(src) * 1.6)):
+    if len(cand) > max(12, int(len(src) * 1.45)):
         return False
     # Mid-sentence VO spans usually start with a verb; reject full-clause flips
-    # like "Customer loyalty strengthens".
+    # like "Customer loyalty strengthens" / "customer loyalty strengthens".
     if src[0].islower() and cand[0].isupper():
         return False
+    src_tail = set(src_tokens[1:])
+    if cand_tokens[0] in src_tail:
+        return False
+    # Reject duplicated content words ("…experience to create").
+    from collections import Counter
+
+    src_counts = Counter(src_tokens)
+    cand_counts = Counter(cand_tokens)
+    for token, count in cand_counts.items():
+        if len(token) < 4:
+            continue
+        if count > max(1, src_counts.get(token, 0)):
+            return False
+    # Reject dangling infinitive tails invented by T5.
+    if re.search(r"\bto\s+[a-z]{3,}$", low) and " to " not in src.lower():
+        return False
     return True