R14-KK-G2-R1 Per-Claim Repair Entitlement + Late-Detected QA Stabilization

core/generation_runner.py

@@ -43,7 +43,7 @@ CITATION_ROLE_RULES = {
     "methods": (2, 280),
     "discussion": (2, 220),
 }
-TRANSLATION_CONTRACT_VERSION = 'hierarchical-multilingual-v6'
+TRANSLATION_CONTRACT_VERSION = 'hierarchical-multilingual-v7'
 
 
 def _clean_trailing_references(text, keywords=("reference", "references", "reference list", "参考文献")):

core/pipeline_agent.py

@@ -62,7 +62,8 @@ class CitationAuditError(DeepSeekRequestError):
 
 
 class SSciPipelineAgent:
-    TRANSLATION_CONTRACT_VERSION = 'hierarchical-multilingual-v6'
+    TRANSLATION_CONTRACT_VERSION = 'hierarchical-multilingual-v7'
+    MAX_TRANSLATION_CLAIM_REPAIR_ATTEMPTS = 2
 
     def __init__(self, ds_key, s2_key=""):
         self.ds_key = str(ds_key or "").strip()
@@ -6330,9 +6331,11 @@ Decision must be CONFIRMED_BLOCKING or NOT_CONFIRMED. evidence_code must be one
                     source_blocks, translations, quality_issues, self._translation_claim_qa_evidence,
                     self._translation_claim_repair_meta,
                 )
-                local_round = 0
-                while quality_issues and local_round < 2:
-                    local_round += 1
+                repair_cycle = 0
+                total_claims = sum(len(self._translation_claim_units(source)) for source in source_blocks.values())
+                max_repair_cycles = max(1, total_claims * self.MAX_TRANSLATION_CLAIM_REPAIR_ATTEMPTS)
+                while quality_issues and repair_cycle < max_repair_cycles:
+                    repair_cycle += 1
                     failed_pairs = [
                         (block_id, claim_id, codes)
                         for block_id, claim_map in sorted(quality_issues.items())
@@ -6340,7 +6343,7 @@ Decision must be CONFIRMED_BLOCKING or NOT_CONFIRMED. evidence_code must be one
                     ]
                     logging.warning(
                         'Translation claim-local QA recovery: section=%s round=%s failed=%s',
-                        section_name or 'section', local_round,
+                        section_name or 'section', repair_cycle,
                         [
                             {'id': f'SP-P{block_id}-{claim_id}', 'codes': list(codes)}
                             for block_id, claim_id, codes in failed_pairs
@@ -6363,6 +6366,36 @@ Decision must be CONFIRMED_BLOCKING or NOT_CONFIRMED. evidence_code must be one
                         evidence_key = (block_id, claim_id)
                         repair_evidence = dict(self._translation_claim_qa_evidence.get(evidence_key) or {})
                         code_set = {str(code or '').strip().upper() for code in (codes or [])}
+                        meta = self._translation_claim_repair_meta.setdefault(evidence_key, {
+                            'repair_attempt_count': 0, 'last_issue_codes': [], 'last_evidence_signature': '',
+                            'last_strategy': '', 'nonconvergent_state': '',
+                        })
+                        prior_attempts = int(meta.get('repair_attempt_count') or 0)
+                        prior_codes = set(meta.get('last_issue_codes') or [])
+                        already_source_regenerated = str(meta.get('last_strategy') or '').upper() == 'SOURCE_REGENERATE'
+                        already_exhausted = str(meta.get('nonconvergent_state') or '').startswith('exhausted')
+                        if (
+                            prior_attempts >= self.MAX_TRANSLATION_CLAIM_REPAIR_ATTEMPTS
+                            or already_exhausted
+                            or (already_source_regenerated and bool(prior_codes & code_set))
+                        ):
+                            meta['nonconvergent_state'] = 'exhausted_after_source_regenerate' if already_source_regenerated else 'exhausted_after_claim_budget'
+                            carried_exhausted.setdefault(block_id, {})[claim_id] = list(codes)
+                            logging.info(
+                                'Translation claim repair entitlement exhausted: section=%s id=SP-P%s-%s attempts=%s strategy=%s',
+                                section_name or 'section', block_id, claim_id, prior_attempts, meta.get('last_strategy') or '',
+                            )
+                            self._record_translation_assurance(
+                                'CLAIM_REPAIR_EXHAUSTED', target_lang=target_lang,
+                                block_id=f'SP-P{block_id}-{claim_id}', severity='critical', decision='ADJUDICATE',
+                                detail={'attempts': prior_attempts, 'strategy': str(meta.get('last_strategy') or '')},
+                            )
+                            continue
+                        if prior_attempts == 0 and repair_cycle > 1:
+                            logging.info(
+                                'Translation late-detected claim receives own repair entitlement: section=%s id=SP-P%s-%s cycle=%s',
+                                section_name or 'section', block_id, claim_id, repair_cycle,
+                            )
 
                         semantic_deltas = [
                             str(code or '').strip().upper() for code in (repair_evidence.get('semantic_delta_codes') or [])
@@ -6441,27 +6474,10 @@ Decision must be CONFIRMED_BLOCKING or NOT_CONFIRMED. evidence_code must be one
                                     )
                         self._translation_claim_qa_evidence[evidence_key] = dict(repair_evidence)
                         signature = self._translation_repair_evidence_signature(codes, repair_evidence)
-                        meta = self._translation_claim_repair_meta.setdefault(evidence_key, {
-                            'repair_attempt_count': 0, 'last_issue_codes': [], 'last_evidence_signature': '',
-                            'last_strategy': '', 'nonconvergent_state': '',
-                        })
-                        prior_codes = set(meta.get('last_issue_codes') or [])
                         same_defect = (
                             str(meta.get('last_evidence_signature') or '') == signature
                             and bool(prior_codes & code_set)
                         )
-                        if (
-                            str(meta.get('nonconvergent_state') or '').startswith('exhausted')
-                            and bool(prior_codes & code_set)
-                        ) or (same_defect and str(meta.get('last_strategy') or '').upper() == 'SOURCE_REGENERATE'):
-                            meta['nonconvergent_state'] = 'exhausted_after_source_regenerate'
-                            carried_exhausted.setdefault(block_id, {})[claim_id] = list(codes)
-                            self._record_translation_assurance(
-                                'CLAIM_REPAIR_EXHAUSTED', target_lang=target_lang,
-                                block_id=f'SP-P{block_id}-{claim_id}', severity='critical', decision='ADJUDICATE',
-                                detail={'attempts': int(meta.get('repair_attempt_count') or 0), 'strategy': 'SOURCE_REGENERATE'},
-                            )
-                            continue
 
                         force_source_regenerate = (
                             'UNRESOLVED_SEMANTIC_DELTA' in set(repair_evidence.get('semantic_delta_codes') or [])
@@ -6469,6 +6485,7 @@ Decision must be CONFIRMED_BLOCKING or NOT_CONFIRMED. evidence_code must be one
                         )
                         strategy = 'SOURCE_REGENERATE' if (
                             force_source_regenerate
+                            or prior_attempts >= 1
                             or (same_defect and str(meta.get('last_strategy') or '').upper() == 'LOCAL_EDIT')
                         ) else 'LOCAL_EDIT'
                         if strategy == 'SOURCE_REGENERATE':