Prevent truncated evidence card responses

app.py

@@ -444,18 +444,29 @@ def validate_stage_with_repairs(
     raise RuntimeError(f"{stage_name} validation failed")
 
 
+def prepare_evidence_inputs(
+    papers: list[dict[str, Any]], deep_texts: dict[str, str], limit: int
+) -> list[dict[str, Any]]:
+    """Prioritize deeply read papers and bound evidence-card output size."""
+    selected_ids = set(deep_texts)
+    ordered = sorted(papers, key=lambda paper: paper.get("id") not in selected_ids)
+    evidence_inputs = []
+    for paper in ordered[:limit]:
+        item = {k: paper.get(k) for k in ("id", "title", "summary", "authors", "published", "updated", "url")}
+        if paper.get("id") in deep_texts:
+            item["selected_text_extract"] = deep_texts[paper["id"]][:9000]
+        evidence_inputs.append(item)
+    return evidence_inputs
+
+
 def build_report(
     request: ResearchRequest,
     papers: list[dict[str, Any]],
     deep_texts: dict[str, str],
     english_understanding: dict[str, Any],
 ) -> dict[str, Any]:
-    raw_evidence = []
-    for paper in papers:
-        item = {k: paper.get(k) for k in ("id", "title", "summary", "authors", "published", "updated", "url")}
-        if paper["id"] in deep_texts:
-            item["selected_text_extract"] = deep_texts[paper["id"]]
-        raw_evidence.append(item)
+    evidence_limit = min(8, DEPTHS[request.depth]["deep"])
+    raw_evidence = prepare_evidence_inputs(papers, deep_texts, evidence_limit)
 
     if raw_evidence:
         card_payload = {
@@ -468,12 +479,26 @@ def build_report(
                 "schema": EvidenceCardsEnvelope.model_json_schema(),
             },
         }
-        cards_result = chat_json(
-            "Create compact evidence cards for an algorithm literature review. Keep experimental conditions and limitations. "
-            "Do not invent results, citations, or paper IDs. Return only the required JSON object and no empty fields.",
-            json.dumps(card_payload, ensure_ascii=False),
-            3600,
-        )
+        card_prompt = json.dumps(card_payload, ensure_ascii=False)
+        try:
+            cards_result = chat_json(
+                "Create one compact evidence card per supplied paper. Keep only decision-relevant experimental conditions "
+                "and limitations. Do not invent results, citations, or paper IDs. Return a complete JSON object; stay concise.",
+                card_prompt,
+                3600,
+            )
+        except RuntimeError as first_error:
+            cards_result = chat_json(
+                "The previous evidence-card response was incomplete or invalid JSON. Regenerate it more concisely. "
+                "Return exactly one short card per supplied paper as a complete JSON object with no markdown or commentary.",
+                json.dumps({
+                    "previous_error": str(first_error),
+                    "request": request.model_dump(),
+                    "papers": [{k: item.get(k) for k in ("id", "title", "summary", "published")} for item in raw_evidence],
+                    "schema": EvidenceCardsEnvelope.model_json_schema(),
+                }, ensure_ascii=False),
+                3000,
+            )
         cards = validate_stage_with_repairs(
             cards_result,
             EvidenceCardsEnvelope,

tests/test_app.py

@@ -2,7 +2,7 @@ import unittest
 
 from pydantic import ValidationError
 
-from app import ResearchReport, deduplicate, render_presentation, render_section_content, render_value, safe_slug
+from app import ResearchReport, deduplicate, prepare_evidence_inputs, render_presentation, render_section_content, render_value, safe_slug
 
 
 def base_report(section):
@@ -23,6 +23,13 @@ class AppTests(unittest.TestCase):
     def test_deduplicate_versions(self):
         self.assertEqual(len(deduplicate([{"id": "1234.5678v1"}, {"id": "1234.5678v2"}])), 1)
 
+    def test_evidence_inputs_are_bounded_and_prioritize_deep_reads(self):
+        papers = [{"id": str(i), "title": f"Paper {i}", "summary": "Summary"} for i in range(10)]
+        result = prepare_evidence_inputs(papers, {"8": "deep text", "9": "deep text"}, 3)
+        self.assertEqual(len(result), 3)
+        self.assertEqual([item["id"] for item in result[:2]], ["8", "9"])
+        self.assertIn("selected_text_extract", result[0])
+
     def test_render_is_downloadable_document(self):
         section = {"title": "Scope", "purpose": "Define the technical decision boundary.", "takeaway": "The boundary determines the useful evidence.", "layout": "text", "content": {"points": ["A meaningful point", "Another meaningful point"]}, "source_ids": []}
         output = render_presentation(base_report(section), [], "2026-01-01")