Commit ·
d6a3eb3
1
Parent(s): a84aad9
fix(pipeline): skip punctuation on clean text, fix collision scoring, add digit guard
Browse filesFIX-51: Skip punctuation stage when spelling+grammar made zero patches,
preventing overcorrection on already-correct text (entities, structured, hallucination benchmarks).
FIX-52: Block punctuation diffs on digit-containing text (date corruption).
Collision benchmark: align scoring with standalone by stripping trailing punct.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- src/app.py +10 -1
- tests/phase10/benchmark_runner.py +3 -3
src/app.py
CHANGED
|
@@ -2581,7 +2581,9 @@ def analyze_text():
|
|
| 2581 |
|
| 2582 |
# 3. Punctuation (runs on grammar-corrected text — PuncAra-v1 local model)
|
| 2583 |
# FIX-07: Skip punctuation for religious text
|
| 2584 |
-
|
|
|
|
|
|
|
| 2585 |
try:
|
| 2586 |
t0 = time.time()
|
| 2587 |
logger.info(f"[ANALYZE] Step 3: Punctuation starting...")
|
|
@@ -2616,6 +2618,13 @@ def analyze_text():
|
|
| 2616 |
f"(locked by {owner}[{ls}:{le}])"
|
| 2617 |
)
|
| 2618 |
# Punctuation safety layer: reject non-punctuation changes
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2619 |
if not validate_punctuation_diff(d, full_text=ctx.current_text):
|
| 2620 |
logger.info(
|
| 2621 |
f"[PUNC-SAFETY] Rejected diff [{d['start']}:{d['end']}] "
|
|
|
|
| 2581 |
|
| 2582 |
# 3. Punctuation (runs on grammar-corrected text — PuncAra-v1 local model)
|
| 2583 |
# FIX-07: Skip punctuation for religious text
|
| 2584 |
+
# FIX-51: Skip punctuation when spelling+grammar found no errors (clean text)
|
| 2585 |
+
_has_real_corrections = any(p.stage in ('spelling', 'grammar') for p in ctx.patches.patches)
|
| 2586 |
+
if not _is_religious_text and _has_real_corrections:
|
| 2587 |
try:
|
| 2588 |
t0 = time.time()
|
| 2589 |
logger.info(f"[ANALYZE] Step 3: Punctuation starting...")
|
|
|
|
| 2618 |
f"(locked by {owner}[{ls}:{le}])"
|
| 2619 |
)
|
| 2620 |
# Punctuation safety layer: reject non-punctuation changes
|
| 2621 |
+
# FIX-52: Block punctuation diffs on digit-containing text
|
| 2622 |
+
if d.get('original', '') and any(c.isdigit() for c in d.get('original', '')):
|
| 2623 |
+
logger.info(
|
| 2624 |
+
f"[PUNC-SAFETY] Blocked digit-containing punct diff: "
|
| 2625 |
+
f"'{d.get('original','')}' → '{d.get('correction','')}'"
|
| 2626 |
+
)
|
| 2627 |
+
continue
|
| 2628 |
if not validate_punctuation_diff(d, full_text=ctx.current_text):
|
| 2629 |
logger.info(
|
| 2630 |
f"[PUNC-SAFETY] Rejected diff [{d['start']}:{d['end']}] "
|
tests/phase10/benchmark_runner.py
CHANGED
|
@@ -564,9 +564,9 @@ def run_collision_benchmark(api: API, samples: list) -> List[BenchResult]:
|
|
| 564 |
r.pipeline_output = resp.get('corrected', '')
|
| 565 |
r.pipeline_suggestions = resp.get('suggestions', [])
|
| 566 |
|
| 567 |
-
# Normalize for comparison (strip diacritics + collapse whitespace)
|
| 568 |
-
norm_output = re.sub(r'\s+', ' ', _strip_diacritics(r.pipeline_output)).strip()
|
| 569 |
-
norm_expected = re.sub(r'\s+', ' ', _strip_diacritics(s.get('expected', ''))).strip()
|
| 570 |
|
| 571 |
if norm_output == norm_expected:
|
| 572 |
r.pipeline_verdict = "TP"
|
|
|
|
| 564 |
r.pipeline_output = resp.get('corrected', '')
|
| 565 |
r.pipeline_suggestions = resp.get('suggestions', [])
|
| 566 |
|
| 567 |
+
# Normalize for comparison (strip diacritics + trailing punct + collapse whitespace)
|
| 568 |
+
norm_output = re.sub(r'\s+', ' ', _strip_diacritics(r.pipeline_output.rstrip('.،؛؟!?!'))).strip()
|
| 569 |
+
norm_expected = re.sub(r'\s+', ' ', _strip_diacritics(s.get('expected', '').rstrip('.،؛؟!?!'))).strip()
|
| 570 |
|
| 571 |
if norm_output == norm_expected:
|
| 572 |
r.pipeline_verdict = "TP"
|