Step 4A-6-C-R3.2.8.5-R2 Source-Derived Numeric Equivalence + Claim-Local Fact Recovery

.gitattributes

@@ -33,4 +33,3 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
-core/__pycache__/pipeline_agent.cpython-313.pyc filter=lfs diff=lfs merge=lfs -text

__pycache__/desktop.ini

@@ -1,2 +0,0 @@
-[.ShellClassInfo]
-IconResource=C:\Program Files\Google\Drive File Stream\129.0.1.0\GoogleDriveFS.exe,27

core/__pycache__/desktop.ini

@@ -1,2 +0,0 @@
-[.ShellClassInfo]
-IconResource=C:\Program Files\Google\Drive File Stream\129.0.1.0\GoogleDriveFS.exe,27

core/__pycache__/pipeline_agent.cpython-313.pyc

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:5ff1e1c8ae14f93f19fb2d92015df442e858db9e6cf8deccff732c4e0c8b1973
-size 273806

core/generation_runner.py

@@ -33,7 +33,7 @@ ProgressCallback = Callable[..., None]
 MIN_FINAL_REFERENCES = 50
 LITERATURE_EVIDENCE_TARGET = 75
 CITATION_PLAN_TARGET = 70
-TRANSLATION_CONTRACT_VERSION = 'claim-bound-v1'
+TRANSLATION_CONTRACT_VERSION = 'claim-bound-v2'
 
 
 def _clean_trailing_references(text, keywords=("reference", "references", "reference list", "参考文献")):

core/pipeline_agent.py

@@ -56,7 +56,7 @@ class CitationAuditError(DeepSeekRequestError):
 
 
 class SSciPipelineAgent:
-    TRANSLATION_CONTRACT_VERSION = 'claim-bound-v1'
+    TRANSLATION_CONTRACT_VERSION = 'claim-bound-v2'
 
     def __init__(self, ds_key, s2_key=""):
         self.ds_key = str(ds_key or "").strip()
@@ -3589,16 +3589,21 @@ Output only the repaired section."""
 
     @classmethod
     def _translation_lexical_quantifier_candidates(cls, value):
-        """High-confidence source lexical quantities that may render as Arabic digits.
+        """High-confidence source numeric semantics that may render as Arabic digits.
 
-        These candidates are private attribution evidence, not model-visible facts.
-        Rhetorical ordinals are intentionally excluded unless attached to a bounded
-        research/structural noun.
+        R3.2.8.5-R2 keeps attribution claim-local and source-derived.  It accepts
+        ordinary and hyphenated academic quantity forms (``eleven-item``,
+        ``three-wave``, ``second-order``), relational forms, and English month
+        names that are commonly rendered numerically in Chinese.  These values
+        are private validation evidence and are never exposed to the model.
         """
         text = str(value or '')
         cardinal = {
             'zero':'0','one':'1','two':'2','three':'3','four':'4','five':'5','six':'6',
             'seven':'7','eight':'8','nine':'9','ten':'10','eleven':'11','twelve':'12',
+            'thirteen':'13','fourteen':'14','fifteen':'15','sixteen':'16','seventeen':'17',
+            'eighteen':'18','nineteen':'19','twenty':'20','thirty':'30','forty':'40',
+            'fifty':'50','sixty':'60','seventy':'70','eighty':'80','ninety':'90',
         }
         ordinal = {
             'first':'1','second':'2','third':'3','fourth':'4','fifth':'5','sixth':'6',
@@ -3607,22 +3612,42 @@ Output only the repaired section."""
         measurable_nouns = (
             r'mechanisms?|groups?|stages?|steps?|waves?|models?|hypotheses?|dimensions?|factors?|items?|'
             r'rounds?|phases?|samples?|studies?|analyses?|conditions?|paths?|constructs?|measures?|components?|'
-            r'categories?|levels?|questions?|options?|variables?|sections?|claims?|explanations?|participants?|teams?'
+            r'categories?|levels?|questions?|options?|variables?|sections?|claims?|explanations?|participants?|teams?|'
+            r'orders?|points?|scales?|types?|classes?|periods?|years?|months?|weeks?|days?|trials?|arms?|effects?|generations?'
         )
         found = []
         cardinal_words = '|'.join(cardinal)
-        for match in re.finditer(rf'(?i)\b({cardinal_words})\s+({measurable_nouns})\b', text):
+        # Academic attributive compounds are especially important because a
+        # source such as ``eleven-item scale`` naturally becomes ``11项量表``.
+        for match in re.finditer(rf'(?i)\b({cardinal_words})(?:\s+|-)({measurable_nouns})\b', text):
             found.append(cardinal[match.group(1).casefold()])
         ordinal_words = '|'.join(ordinal)
-        for match in re.finditer(rf'(?i)\b({ordinal_words})\s+({measurable_nouns})\b', text):
+        for match in re.finditer(rf'(?i)\b({ordinal_words})(?:\s+|-)({measurable_nouns})\b', text):
             found.append(ordinal[match.group(1).casefold()])
         for match in re.finditer(r'(?i)\b(twice|twofold|threefold|fourfold)\b', text):
             found.append({'twice':'2','twofold':'2','threefold':'3','fourfold':'4'}[match.group(1).casefold()])
         # one-to-one is a relational form with two explicit quantitative roles.
         if re.search(r'(?i)\bone[- ]to[- ]one\b', text):
             found.extend(['1', '1'])
+        # Month names can legitimately surface as Arabic month numbers in
+        # Chinese even when the source contains no digit (e.g. November -> 11月).
+        months = {
+            'january':'1','february':'2','march':'3','april':'4','may':'5','june':'6',
+            'july':'7','august':'8','september':'9','october':'10','november':'11','december':'12',
+        }
+        for month, number in months.items():
+            if re.search(rf'(?i)\b{month}\b', text):
+                found.append(number)
         return Counter(found)
 
+    @staticmethod
+    def _is_translation_unowned_numeric_error(error):
+        message = str(error or '')
+        return (
+            message.startswith('Translation introduced unowned factual content')
+            and "{'numeric':" in message
+        )
+
     @classmethod
     def _protect_translation_tokens(cls, value, claim_id='C000001'):
         """Protect one claim while keeping factual values in a private server ledger."""
@@ -4322,10 +4347,37 @@ Translate headings and terminology naturally. Preserve standard variable abbrevi
             temp=0.0, model_role='translation', thinking=False, allow_fallback=False,
         )
         parsed = self._parse_marked_translation(result, [block_id])
-        restored = self._restore_translation_tokens(
-            parsed[block_id], mapping, f'SP-P{block_id}', target_lang,
-            claim_sources={claim_id: source_claim}, keep_claim_markers=True,
-        )
+        try:
+            restored = self._restore_translation_tokens(
+                parsed[block_id], mapping, f'SP-P{block_id}', target_lang,
+                claim_sources={claim_id: source_claim}, keep_claim_markers=True,
+            )
+        except DeepSeekRequestError as error:
+            if not self._is_translation_unowned_numeric_error(error):
+                raise
+            logging.warning(
+                'Translation claim-local numeric recovery: block=SP-P%s claim=%s error=%s',
+                block_id, claim_id, self._safe_source_error(error),
+            )
+            numeric_prompt = system_prompt + (
+                '\nNUMERIC FACT RECOVERY FOR ONE CLAIM. Repair only this claim. Preserve every SP-P, SP-C, and SP-T marker byte-for-byte. '
+                'Do not introduce any Arabic digit, percentage, sample size, scale point, model number, date number, or other numeric surface unless it is represented by an SP-T marker or is directly and unambiguously expressed by an English number-word, ordinal, fold expression, relational quantity, or month name in this source claim. '
+                'If a lexical quantity is not protected by SP-T and you are uncertain, express it lexically in the target language rather than inventing an Arabic digit. '
+                'Do not add, remove, merge, or reinterpret claims. Return only the marked translated claim.'
+                + '\n<SEMANTIC_ANCHOR_MANIFEST>\n'
+                + json.dumps({block_id: safe_manifest}, ensure_ascii=False)
+                + '\n</SEMANTIC_ANCHOR_MANIFEST>'
+            )
+            retry_result = self.call_deepseek(
+                numeric_prompt, marked_source,
+                max_tokens=max(1200, min(2600, self._count_tokens(marked_source) * 2 + 900)),
+                temp=0.0, model_role='translation', thinking=False, allow_fallback=False,
+            )
+            retry_parsed = self._parse_marked_translation(retry_result, [block_id])
+            restored = self._restore_translation_tokens(
+                retry_parsed[block_id], mapping, f'SP-P{block_id}', target_lang,
+                claim_sources={claim_id: source_claim}, keep_claim_markers=True,
+            )
         clean = self._strip_translation_claim_markers(restored)
         self._validate_translation_block_protected(block_id, source_claim, clean, target_lang)
         _order, segments = self._parse_translation_claim_segments(restored, [claim_id])
@@ -4473,6 +4525,7 @@ Translate headings and terminology naturally. Preserve standard variable abbrevi
                 message = str(error)
                 if (
                     message.startswith('Translation claim-local QA failed')
+                    or message.startswith('Translation introduced unowned factual content')
                     or message.startswith('Translation alignment')
                     or message.startswith('Translation changed ')
                     or message.startswith('Translation protected ')

smartpaper_restore_bridge/__pycache__/desktop.ini

@@ -1,2 +0,0 @@
-[.ShellClassInfo]
-IconResource=C:\Program Files\Google\Drive File Stream\129.0.1.0\GoogleDriveFS.exe,27

utils/__pycache__/desktop.ini

@@ -1,2 +0,0 @@
-[.ShellClassInfo]
-IconResource=C:\Program Files\Google\Drive File Stream\129.0.1.0\GoogleDriveFS.exe,27