Make deep report validation semantically tolerant

app.py

@@ -109,14 +109,10 @@ class ReportSection(BaseModel):
             raise ValueError("comparison content requires columns and at least two rows")
         if self.layout == "comparison" and not all(isinstance(row, dict) for row in content.get("rows", [])):
             raise ValueError("comparison rows must be objects")
-        if self.layout == "paper-analysis" and not all(
-            content.get(key) for key in ("method", "evidence", "limitations")
-        ):
-            raise ValueError("paper-analysis requires method, evidence, and limitations")
-        if self.layout == "evidence-gap" and not all(
-            content.get(key) for key in ("interpretation", "hypotheses", "required_inputs")
-        ):
-            raise ValueError("evidence-gap requires interpretation, hypotheses, and required_inputs")
+        if self.layout in {"paper-analysis", "evidence-gap"}:
+            meaningful_blocks = [value for value in content.values() if value not in (None, "", [], {})]
+            if len(meaningful_blocks) < 2:
+                raise ValueError(f"{self.layout} requires at least two substantive information blocks")
         return self
 
 
@@ -666,7 +662,8 @@ def build_report(
         "{columns:[...],rows:[{name,values:[...]}]}; process uses {steps:[{name,description}]} or "
         "{common_steps:[...],branches:[{name,steps:[...]}]}; equation uses {equation,symbols,interpretation}; "
         "paper-analysis uses {method,evidence,limitations}; evidence-gap uses "
-        "{interpretation,related_topics,hypotheses,required_inputs}. Write for slides, not an essay: one conclusion per "
+        "{interpretation,related_topics,hypotheses,required_inputs}. Equivalent topic-appropriate field names are allowed; "
+        "never add empty placeholder fields merely to match an example. Write for slides, not an essay: one conclusion per "
         "section, concise points, explicit assumptions, comparable dimensions, and no repeated background. Distinguish "
         "retrieved facts from analysis or hypotheses. Never invent sources, papers, metrics, or recency claims."
     )
@@ -694,14 +691,17 @@ def build_report(
             4200,
         )
         sections_result = align_sections_to_plan(sections_result, batch)
-        validated_batch = validate_stage_with_repairs(
-            sections_result,
-            SectionsEnvelope,
-            f"report sections batch {batch_index // 4 + 1}",
-            {**shared_context, "approved_section_plan": [item.model_dump() for item in batch]},
-            4200,
-        )
-        if len(validated_batch.sections) != len(batch):
+        try:
+            validated_batch = validate_stage_with_repairs(
+                sections_result,
+                SectionsEnvelope,
+                f"report sections batch {batch_index // 4 + 1}",
+                {**shared_context, "approved_section_plan": [item.model_dump() for item in batch]},
+                4200,
+            )
+        except RuntimeError:
+            validated_batch = None
+        if validated_batch is None or len(validated_batch.sections) != len(batch):
             recovered_sections: list[ReportSection] = []
             for plan in batch:
                 one_result = chat_json(
@@ -800,6 +800,8 @@ def field_label(key: Any) -> str:
         "method": "方法", "mechanism": "机制", "problem": "问题", "innovation": "创新",
         "tradeoffs": "权衡", "interpretation": "解释", "related_topics": "相关主题",
         "hypotheses": "待验证假设", "required_inputs": "所需补充信息", "findings": "发现",
+        "known_evidence": "已有证据", "unknowns": "证据缺口", "next_steps": "下一步验证",
+        "open_questions": "待回答问题", "method_summary": "方法概述", "experimental_results": "实验结果",
     }
     normalized = str(key).strip().lower().replace("-", "_")
     return labels.get(normalized, str(key).replace("_", " ").replace("-", " ").strip().title())

tests/test_app.py

@@ -117,6 +117,15 @@ class AppTests(unittest.TestCase):
         section = {"title": "Evidence gap", "purpose": "Separate known facts from hypotheses.", "takeaway": "More evidence is required before choosing a method.", "layout": "evidence-gap", "content": {"interpretation": "The term may be ambiguous.", "related_topics": ["Adjacent field"], "hypotheses": ["Possible interpretation"], "required_inputs": ["Definition or internal data"]}, "source_ids": []}
         self.assertEqual(ResearchReport.model_validate(base_report(section)).sections[0].layout, "evidence-gap")
 
+    def test_evidence_gap_accepts_equivalent_semantic_fields(self):
+        section = {"title": "Evidence gap", "purpose": "Separate available evidence from unresolved questions.", "takeaway": "Deployment requires evidence that is not yet public.", "layout": "evidence-gap", "content": {"known_evidence": ["A related result exists"], "unknowns": ["No direct benchmark"], "next_steps": ["Run an internal evaluation"]}, "source_ids": []}
+        validated = ResearchReport.model_validate(base_report(section))
+        self.assertEqual(validated.sections[0].content["unknowns"], ["No direct benchmark"])
+
+    def test_paper_analysis_accepts_equivalent_semantic_fields(self):
+        section = {"title": "Paper analysis", "purpose": "Analyze a paper without enforcing field spelling.", "takeaway": "The result is useful under a bounded condition.", "layout": "paper-analysis", "content": {"method_summary": "A compact architecture", "experimental_results": ["A controlled comparison"], "caveats": ["A narrow dataset"]}, "source_ids": []}
+        self.assertEqual(ResearchReport.model_validate(base_report(section)).sections[0].layout, "paper-analysis")
+
     def test_invalid_empty_section(self):
         section = {"title": "Empty", "purpose": "This section should fail validation.", "takeaway": "There is no substantive supporting content.", "layout": "text", "content": {}, "source_ids": []}
         with self.assertRaises(ValidationError):