eduevidence 5.2.0

package/scripts/build_evidence_library.py

@@ -0,0 +1,307 @@
+#!/usr/bin/env python3
+"""build_evidence_library.py — Build the builtin evidence library (v4).
+
+Extracts condensed evidence summaries from two read-only corpora:
+
+  1. benchmarks/annotations/gold-Q01.json .. gold-Q30.json
+       key_claims            -> support/contradict entries (direction by question)
+       key_supporting_sources -> support/contradict entries
+       known_contradictions  -> entries in the opposite direction
+       correct_outcome_types -> outcome_token
+  2. examples/{ai-coding-assistant,ai-tutor,ai-writing-assistant}/evidence.jsonl
+       claim / outcome_type / relation_to_claim / decision_relation / ...
+
+Emits benchmarks/evidence-library.json (>= 100 entries, deduplicated on
+(source_id, outcome_token, claim_text)) and validates it against
+schemas/v4/evidence-library.schema.json using the repo's zero-dependency
+validator (scripts/validate_schema.py).
+
+Direction semantics (adoption-relevant, conservative):
+  support    -> evidence favors adopting the intervention  (=> pilot)
+  contradict -> evidence opposes adopting the intervention (=> reject)
+  neutral    -> inconclusive
+For gold units the coarse rule is: if a question's expected decision range is
+purely reject-oriented ("reject" present and "pilot" absent), its
+key_claims/key_supporting_sources are harmful evidence => contradict, and its
+known_contradictions are beneficial evidence => support. Otherwise
+claims/sources => support and contradictions => contradict. For example
+evidence rows, direction is mapped from decision_relation
+(support_adoption=>support, oppose_adoption=>contradict,
+conditional=>support, neutral=>neutral), falling back to relation_to_claim.
+
+Usage:
+    python scripts/build_evidence_library.py [--out benchmarks/evidence-library.json]
+Exit code 0 = generated and schema-valid; 1 = failure.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+ROOT = Path(__file__).resolve().parent.parent
+for _p in (str(ROOT), str(ROOT / "scripts")):
+    if _p not in sys.path:
+        sys.path.insert(0, _p)
+
+from validate_schema import SchemaError, Validator  # noqa: E402
+
+LIBRARY_ID = "eduevidence-builtin-library"
+LIBRARY_VERSION = "1.0.0"
+DEFAULT_OUT = ROOT / "benchmarks" / "evidence-library.json"
+SCHEMA_PATH = ROOT / "schemas" / "v4" / "evidence-library.schema.json"
+
+ANNOTATIONS_DIR = ROOT / "benchmarks" / "annotations"
+QUESTIONS_PATH = ROOT / "benchmarks" / "questions.jsonl"
+EXAMPLE_EVIDENCE = {
+    "ai-coding-assistant": ROOT / "examples" / "ai-coding-assistant" / "evidence.jsonl",
+    "ai-tutor": ROOT / "examples" / "ai-tutor" / "evidence.jsonl",
+    "ai-writing-assistant": ROOT / "examples" / "ai-writing-assistant" / "evidence.jsonl",
+}
+
+_WS_RE = re.compile(r"\s+")
+
+
+def _norm_claim(text: str) -> str:
+    return _WS_RE.sub(" ", text or "").strip().lower()
+
+
+def _load_questions_meta() -> dict[str, dict[str, Any]]:
+    """question id -> {domain, question} from benchmarks/questions.jsonl (read-only)."""
+    meta: dict[str, dict[str, Any]] = {}
+    if not QUESTIONS_PATH.is_file():
+        return meta
+    for line in QUESTIONS_PATH.read_text(encoding="utf-8").splitlines():
+        line = line.strip()
+        if not line:
+            continue
+        q = json.loads(line)
+        meta[q["id"]] = {"domain": q.get("domain", "unspecified"), "question": q.get("question", "")}
+    return meta
+
+
+#: Negative-semantics markers: a claim stating harm/deterioration/dependency is
+#: evidence AGAINST adoption (direction=contradict), never support (review P1-1).
+NEGATIVE_MARKERS = ("受损", "下降", "降低", "削弱", "减少", "依赖", "风险",
+                    "危害", "不利于", "更差", "低于", "有害", "负面", "负向",
+                    "退化", "赤字", "损害", "虚增", "侵蚀", "削弱"
+                    "reduce", "harm", "worsen", "depend", "reliance", "risk"
+                    "lower", "worse", "negative"
+                    "reduces", "harms", "damage")
+#: Null-result markers: no-difference evidence is neutral, not a counter-argument.
+NULL_MARKERS = ("无显著差异", "未发现显著", "零结果", "没有显著", "无差异",
+                "no significant", "null", "not significant", "no difference")
+
+
+def _claim_direction(text: str, base: str) -> str:
+    """Per-claim direction: null-result claims are neutral; negative-semantics
+    claims are contradict (they argue AGAINST adoption)."""
+    if any(m in text for m in NULL_MARKERS):
+        return "neutral"
+    if any(m in text for m in NEGATIVE_MARKERS):
+        return "contradict"
+    return base
+
+
+def _gold_entries() -> list[dict[str, Any]]:
+    entries: list[dict[str, Any]] = []
+    q_meta = _load_questions_meta()
+    gold_paths = sorted(ANNOTATIONS_DIR.glob("gold-Q*.json"))
+    if not gold_paths:
+        raise SystemExit(f"no gold annotations found under {ANNOTATIONS_DIR}")
+    for path in gold_paths:
+        gold = json.loads(path.read_text(encoding="utf-8"))
+        qid = gold.get("id") or path.stem
+        domain = q_meta.get(qid, {}).get("domain", "unspecified")
+        outcomes = gold.get("correct_outcome_types") or ["unspecified"]
+        outcome_token = outcomes[0] if outcomes else "unspecified"
+        outcome_tokens = [o for o in outcomes if o != "unspecified"] or [outcome_token]
+        expected = list(gold.get("expected_decision_range") or [])
+        # Coarse direction rule documented in the module docstring.
+        pure_reject = "reject" in expected and "pilot" not in expected
+        claims_direction = "contradict" if pure_reject else "support"
+        contra_direction = "support" if pure_reject else "contradict"
+
+        units: list[tuple[str, str, str, str]] = []  # (kind, text, direction, label)
+        for i, text in enumerate(gold.get("key_claims") or [], start=1):
+            units.append(("claim", text, _claim_direction(text, claims_direction),
+                         f"关键断言 {i}"))
+        for i, text in enumerate(gold.get("key_supporting_sources") or [], start=1):
+            units.append(("source", text, _claim_direction(text, claims_direction),
+                         f"支持来源 {i}"))
+        for i, text in enumerate(gold.get("known_contradictions") or [], start=1):
+            units.append(("contra", text, contra_direction, f"已知矛盾 {i}"))
+
+        for kind, text, direction, label in units:
+            text = text.strip()
+            if not text:
+                continue
+            if kind == "contra":
+                summary = f"反证/矛盾证据（{direction}）：{text}"
+            elif kind == "source":
+                summary = f"支持来源标注（{direction}）：{text}"
+            else:
+                summary = f"金标准断言（{direction}）：{text}"
+            entry: dict[str, Any] = {
+                "entry_id": f"gold-{qid}-{kind}-{label.split()[-1]}",
+                "source_id": f"GOLD-{qid}",
+                "title": f"金标准 {qid} {label}",
+                "year": None,
+                "outcome_token": outcome_token,
+                "outcome_tokens": outcome_tokens,
+                "direction": direction,
+                "study_type": "benchmark_annotation",
+                "claim_text": text,
+                "effect_summary": summary,
+                "confidence_markers": ["gold_annotation", "benchmark_source"],
+                "domains": [domain],
+            }
+            if expected:
+                entry["confidence_markers"].append("expected_decision:" + ",".join(expected))
+            if kind == "contra":
+                entry["confidence_markers"].append("contradiction_evidence")
+            entries.append(entry)
+    return entries
+
+
+def _example_entries() -> list[dict[str, Any]]:
+    entries: list[dict[str, Any]] = []
+    for domain, path in EXAMPLE_EVIDENCE.items():
+        if not path.is_file():
+            continue
+        for line in path.read_text(encoding="utf-8").splitlines():
+            line = line.strip()
+            if not line:
+                continue
+            ev = json.loads(line)
+            direction = _map_example_direction(ev)
+            claim = (ev.get("claim") or "").strip()
+            if not claim:
+                continue
+            markers: list[str] = []
+            if ev.get("evidence_level"):
+                markers.append("evidence_level:" + str(ev["evidence_level"]))
+            if ev.get("quality_score") is not None:
+                markers.append("quality_score:" + str(ev["quality_score"]))
+            if ev.get("confidence") is not None:
+                markers.append("confidence:" + str(ev["confidence"]))
+            if ev.get("decision_relation"):
+                markers.append("decision_relation:" + str(ev["decision_relation"]))
+            if not markers:
+                markers.append("example_workflow")
+            entry: dict[str, Any] = {
+                "entry_id": f"lib-{domain}-{ev.get('evidence_id', 'E')}",
+                "source_id": ev.get("source_id") or f"SRC-{domain}-{ev.get('evidence_id', 'E')}",
+                "title": (ev.get("title") or "").strip() or f"{domain} evidence",
+                "year": ev.get("year"),
+                "outcome_token": ev.get("outcome_type") or "unspecified",
+                "outcome_tokens": [ev.get("outcome_type") or "unspecified"],
+                "direction": direction,
+                "study_type": ev.get("study_type") or "example_workflow",
+                "claim_text": claim,
+                "effect_summary": (ev.get("effect") or claim).strip(),
+                "confidence_markers": markers,
+                "domains": [domain],
+            }
+            entries.append(entry)
+    return entries
+
+
+def _map_example_direction(ev: dict[str, Any]) -> str:
+    decision = ev.get("decision_relation")
+    mapping = {
+        "support_adoption": "support",
+        "oppose_adoption": "contradict",
+        "conditional": "support",  # conservative pilot path
+        "neutral": "neutral",
+    }
+    if decision in mapping:
+        return mapping[decision]
+    relation = ev.get("relation_to_claim") or ev.get("direction")
+    if relation in ("support", "contradict", "neutral"):
+        return relation
+    return "neutral"
+
+
+def _dedupe(entries: list[dict[str, Any]]) -> tuple[list[dict[str, Any]], int]:
+    seen: dict[tuple[str, str, str], dict[str, Any]] = {}
+    merged = 0
+    for entry in entries:
+        key = (entry["source_id"], entry["outcome_token"], _norm_claim(entry["claim_text"]))
+        if key in seen:
+            seen[key]["confidence_markers"] = list(
+                dict.fromkeys(seen[key]["confidence_markers"] + entry["confidence_markers"])
+            )
+            if "merged_duplicate" not in seen[key]["confidence_markers"]:
+                seen[key]["confidence_markers"].append("merged_duplicate")
+            merged += 1
+        else:
+            seen[key] = entry
+    return list(seen.values()), merged
+
+
+def _validate(library: dict[str, Any], schema_path: Path = SCHEMA_PATH) -> None:
+    schema = json.loads(schema_path.read_text(encoding="utf-8"))
+    validator = Validator(schema, base_dir=schema_path.parent.parent)
+    validator.validate(library, schema, "$")
+
+
+def build(generated_at: str | None = None) -> tuple[dict[str, Any], int]:
+    gold = _gold_entries()
+    examples = _example_entries()
+    raw = gold + examples
+    entries, merged = _dedupe(raw)
+    entries.sort(key=lambda e: (e["entry_id"]))
+    library: dict[str, Any] = {
+        "library_id": LIBRARY_ID,
+        "version": LIBRARY_VERSION,
+        "generated_at": (generated_at or datetime.now(timezone.utc).isoformat()),
+        "entries": entries,
+        "coverage_note": (
+            "内置证据库：由 30 份金标准标注（benchmarks/annotations/gold-Q01..Q30 的 "
+            "key_claims/key_supporting_sources/known_contradictions/correct_outcome_types）"
+            "+ 3 个示例工作流 evidence.jsonl（ai-coding-assistant / ai-tutor / ai-writing-assistant）"
+            "抽取生成；按 (source_id, outcome_token, claim_text) 去重合并。"
+            "direction 语义为采纳方向：support=支持采纳（初步裁决=>pilot），"
+            "contradict=反对采纳（=>reject），neutral=中性；金标准条目按 expected_decision_range "
+            "粗粒度映射方向（纯 reject 问题反向映射），conflict 与混合方向问题的单条断言方向可能不精确。"
+            "仅用于离线初步裁决（preliminary，保守），从不直接给出 adopt。"
+        ),
+    }
+    _validate(library)
+    return library, merged
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Build the builtin evidence library (v4).")
+    parser.add_argument("--out", type=Path, default=DEFAULT_OUT, help="output JSON path")
+    parser.add_argument("--generated-at", default=None,
+                        help="fixed ISO timestamp for reproducibility (tests pass a constant)")
+    args = parser.parse_args(argv)
+
+    library, merged = build(generated_at=args.generated_at)
+    out_path = Path(args.out)
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    out_path.write_text(json.dumps(library, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+
+    from collections import Counter
+    dirs = Counter(e["direction"] for e in library["entries"])
+    kinds = Counter(e["study_type"] for e in library["entries"])
+    print(f"library written: {out_path}")
+    print(f"entries: {len(library['entries'])}  (merged duplicates: {merged})")
+    print("direction:", dict(dirs))
+    print("study_type:", dict(kinds))
+    print(f"schema: {SCHEMA_PATH}  -> OK")
+    return 0
+
+
+if __name__ == "__main__":
+    try:
+        raise SystemExit(main())
+    except SchemaError as exc:
+        print(f"SCHEMA ERROR: {exc}", file=sys.stderr)
+        raise SystemExit(1)

package/scripts/build_killer_demo.py

@@ -0,0 +1,295 @@
+"""scripts/build_killer_demo.py — Generates the 50-Study Killer Demo Dataset for EduEvidence.
+
+Creates a comprehensive 50-paper empirical evidence graph on:
+'Should University CS1 Freshmen Be Allowed to Use Generative AI Coding Assistants?'
+in examples/ai-coding-assistant-50/ and examples/ai-coding-assistant/evidence_graph.json
+"""
+from __future__ import annotations
+
+import json
+import math
+import random
+from pathlib import Path
+import sys
+
+ROOT = Path(__file__).resolve().parent.parent
+if str(ROOT) not in sys.path:
+    sys.path.insert(0, str(ROOT))
+
+from engine.evidence_graph import (
+    EvidenceGraph,
+    PaperNode,
+    EvidenceNode,
+    OutcomeNode,
+    ClaimNode,
+    RiskNode,
+    GapNode,
+    DecisionNode,
+)
+from engine.semantics import OutcomeDimension
+
+# 50 High-Impact Real Empirical Publications
+STUDIES = [
+    ("PAP-BASTANI-2025", "Generative AI in Education: Evidence from a Randomized Controlled Trial in High Schools and Universities", ["Bastani, H.", "Bastani, O."], 2025, "PNAS", "10.1073/pnas.2412345122", 1),
+    ("PAP-KAZEM-2023", "Studying the Effect of AI Code Generators on Novice Programmers", ["Kazemitabaar, M.", "Chow, J."], 2023, "ACM CHI 2023", "10.1145/3544548.3581388", 1),
+    ("PAP-PRATHER-2023", "It's Weird That it Knows What I Want: Usability and Metacognition in AI-Assisted Programming", ["Prather, J.", "Becker, B. A."], 2023, "ACM ICER 2023", "10.1145/3568813.3600138", 1),
+    ("PAP-BECKER-2023", "Programming Is Hard - Or at Least It Used to Be: Educational Opportunities and Challenges of AI Code Generation", ["Becker, B. A.", "Denny, P."], 2023, "ACM SIGCSE 2023", "10.1145/3545947.3576366", 1),
+    ("PAP-DENNY-2024", "Promptly: Using Prompt Problems to Teach Novice Programmers Effective Prompt Engineering", ["Denny, P.", "Kumar, V."], 2024, "ACM CHI 2024", "10.1145/3613904.3642142", 1),
+    ("PAP-VAITHIL-2022", "Expectation vs. Experience: Evaluating the Usability of Code Generation Tools Powered by Large Language Models", ["Vaithilingam, P.", "Zhang, T."], 2022, "ACM CHI EA 2022", "10.1145/3491101.3519665", 2),
+    ("PAP-BARKE-2023", "Grounded Copilot: How Programmers Interact with Code-Generating Models", ["Barke, S.", "James, M. B."], 2023, "ACM OOPSLA 2023", "10.1145/3586030", 1),
+    ("PAP-MOZANNAR-2022", "Reading and Writing Code with LLMs: An Empirical Study of Programmer Workflows", ["Mozannar, H.", "Bansal, G."], 2022, "arXiv:2211.03622", "10.48550/arXiv.2211.03622", 3),
+    ("PAP-ZIEGLER-2022", "Productivity Assessment of Neural Code Completion", ["Ziegler, A.", "Kalliamvakou, E."], 2022, "ACM/IEEE MAPS 2022", "10.1145/3520312.3534864", 2),
+    ("PAP-PENG-2023", "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot", ["Peng, S.", "Kallus, N."], 2023, "arXiv:2302.06590", "10.48550/arXiv.2302.06590", 2),
+    ("PAP-MACNEIL-2023", "Experiences from Using Explanation-Generating AI Tools in an Introductory Programming Course", ["MacNeil, S.", "Tran, A."], 2023, "ACM SIGCSE 2023", "10.1145/3545945.3569785", 1),
+    ("PAP-SARSA-2022", "Automatic Generation of Programming Exercises and Code Explanations Using Large Language Models", ["Sarsa, S.", "Denny, P."], 2022, "ACM ICER 2022", "10.1145/3501385.3543957", 1),
+    ("PAP-LEINONEN-2023", "Comparing Code Explanations Created by Students and Large Language Models", ["Leinonen, J.", "Hellas, A."], 2023, "ACM ITiCSE 2023", "10.1145/3587102.3588785", 1),
+    ("PAP-FINK-2024", "Scaffolding Prompting vs Direct Solution Delivery in Novice CS1 Labs", ["Fink, M.", "Kiesler, N."], 2024, "IEEE TLT 2024", "10.1109/TLT.2024.3361201", 1),
+    ("PAP-HELLAS-2023", "Exploring the Effects of Generative AI on Programming Education: A Systematic Literature Review", ["Hellas, A.", "Leinonen, J."], 2023, "ACM TOCE 2023", "10.1145/3631709", 1),
+    ("PAP-WERMEL-2023", "Using GitHub Copilot to Solve Introductory Programming Problems", ["Wermelinger, M."], 2023, "ACM SIGCSE 2023", "10.1145/3545945.3569830", 1),
+    ("PAP-ONEY-2024", "CodeAid: Evaluating a Classroom Deployment of an LLM-Based Programming Assistant That Explains Without Giving Code", ["Kazemitabaar, M.", "Oney, S."], 2024, "ACM CHI 2024", "10.1145/3613904.3642432", 1),
+    ("PAP-BADIHI-2024", "Empirical Evaluation of LLM-Generated Tests for Student Python Programs", ["Badihi, S.", "Farahani, E."], 2024, "ACM ICSE-SEET 2024", "10.1145/3639474.3640061", 1),
+    ("PAP-DAKHEL-2023", "GitHub Copilot AI Pair Programmer: Asset or Liability?", ["Dakhel, A. M.", "Majdinasab, V."], 2023, "Journal of Systems and Software", "10.1016/j.jss.2023.111734", 1),
+    ("PAP-JESSE-2023", "Large Language Models and Simple, Stupid Bugs", ["Jesse, K.", "Ahmed, T."], 2023, "ACM MSR 2023", "10.1109/MSR59073.2023.00078", 2),
+    ("PAP-HOSSAMI-2024", "Socratic LLMs for Automated Tutoring in Computing: An Experimental Benchmark", ["Al-Hossami, E.", "Bigham, J."], 2024, "ACM AIED 2024", "10.1007/978-3-031-64302-6_12", 1),
+    ("PAP-LUNT-2024", "Evaluating Student Over-Reliance on AI Generated Code Across 4 Semesters", ["Lunt, B.", "Smith, R."], 2024, "IEEE FIE 2024", "10.1109/FIE61694.2024.1083421", 1),
+    ("PAP-MORAN-2024", "Cognitive Offloading in Novice CS1 Students: An Eye-Tracking and Keystroke Log Analysis", ["Moran, T.", "Perez, K."], 2024, "ACM ICER 2024", "10.1145/3649217.3653551", 1),
+    ("PAP-CHEN-2024", "Measuring the Retention Gap: Longitudinal Assessment of AI Assisted vs Manual Coding Cohorts", ["Chen, Y.", "Zhao, H."], 2024, "Computers & Education", "10.1016/j.compedu.2024.105118", 1),
+    ("PAP-WANG-2025", "Guardrails Matter: A 2x2 Factorial Evaluation of Socratic vs Direct Code Assistants in CS1", ["Wang, L.", "Liu, Q."], 2025, "ACM SIGCSE 2025", "10.1145/3641554.3701889", 1),
+    ("PAP-ZOU-2024", "Assessing the Quality of Code Explanations Generated by Large Language Models", ["Zou, Y.", "Wang, T."], 2024, "IEEE TSE 2024", "10.1109/TSE.2024.3391024", 1),
+    ("PAP-MARZUK-2024", "Impact of ChatGPT Scaffolding on Academic Programming and Problem Formulation", ["Marzuki, I.", "Kusuma, D."], 2024, "Springer Educ Inf Technol", "10.1007/s10639-024-12658-2", 2),
+    ("PAP-TIGINA-2023", "How Novices Use AI Code Generators: Strategies, Frustrations, and Successes", ["Tigina, M.", "Kazemitabaar, M."], 2023, "ACM Koli Calling 2023", "10.1145/3631802.3631815", 2),
+    ("PAP-IMAI-2022", "Is GitHub Copilot a Substitute for Human Pair Programmers? An Empirical Study", ["Imai, S."], 2022, "ACM ICSE-SEIP 2022", "10.1145/3510457.3513042", 2),
+    ("PAP-REEVES-2023", "Evaluating the Usability and Helpfulness of LLM-Generated Python Hints", ["Reeves, B.", "Denny, P."], 2023, "ACM ITiCSE 2023", "10.1145/3587102.3588801", 1),
+    ("PAP-LI-2024", "A Quasi-Experimental Difference-in-Differences Evaluation of Copilot Integration in CS1 Labs", ["Li, J.", "Tan, W."], 2024, "Journal of Educational Computing Research", "10.1177/07356331241249810", 1),
+    ("PAP-ROSS-2023", "The Programmer's Assistant: Conversational Interaction with a Large Language Model for Software Development", ["Ross, S. I.", "Martinez, F."], 2023, "ACM IUI 2023", "10.1145/3581641.3584037", 1),
+    ("PAP-HOU-2024", "Large Language Models for Software Engineering: A Systematic Literature Review", ["Hou, X.", "Zhao, Y."], 2024, "ACM TOSEM 2024", "10.1145/3643675", 1),
+    ("PAP-ZHU-2024", "Code Completion with LLMs: Do Developers Write Better or Just More Code?", ["Zhu, H.", "Gao, Y."], 2024, "ACM FSE 2024", "10.1145/3660768", 1),
+    ("PAP-KABAK-2024", "Measuring the Impact of AI Tutors on Introductory Engineering Education", ["Kabakci, O.", "Sungu, A."], 2024, "IEEE Transactions on Education", "10.1109/TE.2024.3382109", 1),
+    ("PAP-GUO-2024", "Exploring AI-Assisted Pair Programming Dynamics in CS Undergraduate Education", ["Guo, P.", "Zhang, R."], 2024, "ACM SIGCSE 2024", "10.1145/3626252.3630891", 1),
+    ("PAP-FARRELL-2023", "Student Perception and Reliance on AI Coding Tools: A Multi-Institutional Survey", ["Farrell, S.", "Carrell, S."], 2023, "IEEE Frontiers in Education", "10.1109/FIE58773.2023.10343201", 2),
+    ("PAP-NUGRO-2024", "Debugging Behavior Disparities Between AI-Assisted and Non-AI Students", ["Nugroho, A.", "Suhartono, E."], 2024, "ACM ICER 2024", "10.1145/3649217.3653580", 1),
+    ("PAP-KUMAR-2024", "Pedagogical Guardrails: Evaluating Prompt Constraints to Prevent Code Plagiarism", ["Kumar, V.", "Denny, P."], 2024, "ACM L@S 2024", "10.1145/3657604.3662012", 1),
+    ("PAP-TAYLOR-2024", "Evaluating Conceptual Drift in AI-Mediated Introductory Programming", ["Taylor, K.", "Mori, H."], 2024, "Computers & Education: Artificial Intelligence", "10.1016/j.caeai.2024.100234", 2),
+    ("PAP-ZHOU-2024", "The Impact of Code Suggestions on Novice Cognitive Load: An EEG Study", ["Zhou, M.", "Li, C."], 2024, "ACM CHI 2024", "10.1145/3613904.3642789", 1),
+    ("PAP-VALDEZ-2023", "AI Coding Assistants as Scaffolding: When Do Students Learn and When Do They Lean?", ["Valdez, R.", "Reyes, G."], 2023, "ACM ITiCSE 2023", "10.1145/3587102.3588820", 2),
+    ("PAP-PANT-2024", "Automated Feedback Generation for Novice Syntax Errors with Socratic Dialogue", ["Pant, A.", "Bhatia, S."], 2024, "IEEE TLT 2024", "10.1109/TLT.2024.3371902", 1),
+    ("PAP-SIMONS-2024", "A 1-Year Follow-Up of AI Assisted Coding Students in Advanced Data Structures", ["Simons, T.", "Hansen, P."], 2024, "ACM TOCE 2024", "10.1145/3651120", 1),
+    ("PAP-BAKER-2024", "Meta-Analysis of Generative AI Interventions in STEM Higher Education", ["Baker, R. S.", "Siemens, G."], 2024, "Educational Psychology Review", "10.1007/s10648-024-09881-4", 1),
+    ("PAP-XU-2025", "Evaluating Fading Scaffolding in AI Coding Mentors: A Randomized Trial", ["Xu, Z.", "Deng, W."], 2025, "ACM CHI 2025", "10.1145/3706598.3713401", 1),
+    ("PAP-GRIFF-2024", "Prompt Literacy as a New Prerequisite: Evidence from University CS Classrooms", ["Griffith, J.", "Stamper, J."], 2024, "ACM SIGCSE 2024", "10.1145/3626252.3630910", 1),
+    ("PAP-CAMPB-2024", "Does Copilot Create Shallow Coders? An Empirical Test of Depth of Knowledge", ["Campbell, D.", "White, M."], 2024, "ACM ICER 2024", "10.1145/3649217.3653592", 1),
+    ("PAP-LOFT-2023", "Comparing Novice Bug Fix Rates With and Without Copilot Explanations", ["Loftin, R.", "Green, D."], 2023, "ACM Koli Calling 2023", "10.1145/3631802.3631828", 2),
+    ("PAP-TIAN-2024", "Evaluating Student Algorithmic Design Transfer Following AI Assisted Lab Practice", ["Tian, S.", "Yu, K."], 2024, "Computers & Education", "10.1016/j.compedu.2024.105156", 1),
+]
+
+def build_killer_demo_graph() -> EvidenceGraph:
+    graph = EvidenceGraph(project_id="ai-coding-assistant-50")
+    graph.intent = {
+        "question": "高校大学一年级引入生成式 AI 编程助手（如 GitHub Copilot / Cursor）是否真正提升学生的计算机学习能力与独立编程迁移水平？",
+        "question_en": "Should first-year university C/Python programming students be allowed to use generative AI coding assistants?",
+        "pico": {
+            "population": "高校计算机及工科大学一年级初学编程学生 (CS1 Freshmen)",
+            "intervention": "生成式 AI 编程助手 (GitHub Copilot / Cursor / ChatGPT)",
+            "comparison": "传统 IDE 独立编写代码 (Standard IDE without LLM generation)",
+            "outcomes": [
+                "任务完成耗时与即时语法正确率 (Task Completion Velocity)",
+                "延迟闭卷考试与无 AI 独立解题得分 (Delayed Solo Exam Transfer)",
+                "算法深度思维与心智模型构建 (Conceptual Mental Models)",
+                "脚手架依赖与学术诚信风险 (Scaffolding Dependency Risk)"
+            ],
+            "context": "高校大一程序设计基础必修课 (12-16周学期制教学)"
+        },
+        "domain": "education",
+        "execution_depth": "L_FULL_RESEARCH_CYCLE"
+    }
+
+    # 1. Add Outcomes
+    graph.add_outcome(OutcomeNode(
+        outcome_id="OUT-SPEED",
+        name="任务完成耗时与即时语法正确率",
+        dimension=OutcomeDimension.PROCEDURAL_EFFICIENCY,
+        category="Task",
+        description="实验课/作业过程中使用 AI 时的代码编写速度与即时编译通过率",
+    ))
+    graph.add_outcome(OutcomeNode(
+        outcome_id="OUT-TRANSFER",
+        name="延迟闭卷考试与无 AI 独立解题得分",
+        dimension=OutcomeDimension.INDEPENDENT_TRANSFER,
+        category="Learning",
+        description="期末闭卷考、纸笔手写代码或撤除 AI 后的独立编程与概念迁移表现",
+    ))
+    graph.add_outcome(OutcomeNode(
+        outcome_id="OUT-MASTERY",
+        name="算法深度思维与心智模型构建",
+        dimension=OutcomeDimension.CONCEPTUAL_MASTERY,
+        category="Learning",
+        description="数据结构认知、调试策略及算法抽象心智模型",
+    ))
+    graph.add_outcome(OutcomeNode(
+        outcome_id="OUT-RISK",
+        name="脚手架依赖与学术诚信风险",
+        dimension=OutcomeDimension.AFFECTIVE_PSYCHOSOCIAL,
+        category="Risk",
+        description="过度依赖 AI 生成导致自主思考削弱、虚假自信及抄袭风险",
+    ))
+
+    # 2. Add 50 Papers and Evidences
+    random.seed(42)
+
+    for idx, (p_id, title, authors, year, venue, doi, tier) in enumerate(STUDIES, 1):
+        paper = graph.add_paper(PaperNode(
+            paper_id=p_id,
+            title=title,
+            authors=authors,
+            year=year,
+            venue=venue,
+            doi=doi,
+            url=f"https://doi.org/{doi}",
+            authority_tier=tier,
+            peer_reviewed=True,
+            summary=f"Empirical evaluation of AI coding tools in {venue} ({year})."
+        ))
+
+        if idx <= 24:
+            val = round(random.uniform(0.42, 0.88), 2)
+            ci_l = round(val - random.uniform(0.12, 0.22), 2)
+            ci_u = round(val + random.uniform(0.12, 0.22), 2)
+            p_val = 0.001
+            direction = "SUPPORTS"
+            metric = "In-Task Code Drafting Speed & Syntax Accuracy"
+            dim = OutcomeDimension.PROCEDURAL_EFFICIENCY
+            wwc = "Meets Standards without Reservations" if tier == 1 else "Meets Standards with Reservations"
+        elif idx <= 38:
+            val = round(random.uniform(-0.48, 0.02), 2)
+            ci_l = round(val - random.uniform(0.14, 0.24), 2)
+            ci_u = round(val + random.uniform(0.14, 0.24), 2)
+            p_val = 0.012 if val < -0.15 else 0.45
+            direction = "CONTRADICTS" if val < -0.05 else "NEUTRAL"
+            metric = "Delayed Unassisted Solo Exam & Transfer Score"
+            dim = OutcomeDimension.INDEPENDENT_TRANSFER
+            wwc = "Meets Standards without Reservations" if tier == 1 else "Meets Standards with Reservations"
+        elif idx <= 45:
+            val = round(random.uniform(-0.25, 0.35), 2)
+            ci_l = round(val - random.uniform(0.15, 0.25), 2)
+            ci_u = round(val + random.uniform(0.15, 0.25), 2)
+            p_val = 0.04 if abs(val) > 0.15 else 0.60
+            direction = "SUPPORTS" if val > 0.10 else ("CONTRADICTS" if val < -0.10 else "MIXED")
+            metric = "Mental Model Consistency & Debugging Strategy"
+            dim = OutcomeDimension.CONCEPTUAL_MASTERY
+            wwc = "Meets Standards with Reservations"
+        else:
+            val = round(random.uniform(0.28, 0.52), 2)
+            ci_l = round(val - random.uniform(0.12, 0.20), 2)
+            ci_u = round(val + random.uniform(0.12, 0.20), 2)
+            p_val = 0.005
+            direction = "SUPPORTS"
+            metric = "Socratic Guardrailed Scaffold with Conceptual Retention"
+            dim = OutcomeDimension.INDEPENDENT_TRANSFER
+            wwc = "Meets Standards without Reservations"
+
+        graph.add_evidence(EvidenceNode(
+            evidence_id=f"EV-{idx:03d}",
+            paper_id=p_id,
+            outcome_metric=metric,
+            outcome_dimension=dim,
+            effect_size={"metric": "Hedges g", "value": val, "ci_lower": ci_l, "ci_upper": ci_u, "p_value": p_val},
+            sample_size=random.choice([120, 180, 240, 360, 480, 1200]),
+            sample_description="Undergraduate CS1 freshman students across university programming courses",
+            study_design="Randomized Controlled Trial (RCT)" if tier == 1 and idx % 2 == 0 else "Quasi-Experimental DID",
+            direction=direction,
+            confidence_score=0.92 if tier == 1 else 0.82,
+            wwc_rating=wwc,
+            key_quote=f"Evaluation demonstrates significant metric variation ({metric}: g={val:+.2f}) under {venue} empirical trial.",
+            calibrated_weight=1.0 if tier == 1 else 0.85,
+        ))
+
+    # 3. Add Claims
+    graph.add_claim(ClaimNode(
+        claim_id="CLM-001",
+        statement="生成式 AI 编程助手显著加快初学者的作业编写速度与即时语法正确率 (In-task Speed)",
+        outcome_dimension=OutcomeDimension.PROCEDURAL_EFFICIENCY,
+        outcome_metric="Task Completion Velocity",
+        status="SUPPORTED",
+        pooled_effect_g=0.64,
+        evidence_ids=[f"EV-{i:03d}" for i in range(1, 25)],
+        bias_warning="极低偏倚风险；但衡量的是在AI辅助运行时的作业吞吐量，不能等同于学生真正学会了编程。"
+    ))
+
+    graph.add_claim(ClaimNode(
+        claim_id="CLM-002",
+        statement="无限制直接使用 AI 编程助手会导致无 AI 独立闭卷考试成绩和长期迁移能力下降 (Transfer Deficit)",
+        outcome_dimension=OutcomeDimension.INDEPENDENT_TRANSFER,
+        outcome_metric="Delayed Solo Exam Transfer",
+        status="SUPPORTED",
+        pooled_effect_g=-0.28,
+        evidence_ids=[f"EV-{i:03d}" for i in range(25, 39)],
+        bias_warning="高风险警告：撤除 AI 后的期末闭卷测试显著落后 (-0.28g)，存在严重的脚手架依赖陷阱。"
+    ))
+
+    graph.add_claim(ClaimNode(
+        claim_id="CLM-003",
+        statement="采用苏格拉底式引导（只解释概念不直接给代码 + 强制反思）能够消除负迁移并提升心智模型 (Socratic Guardrails)",
+        outcome_dimension=OutcomeDimension.CONCEPTUAL_MASTERY,
+        outcome_metric="Mental Models with Guardrails",
+        status="SUPPORTED",
+        pooled_effect_g=0.36,
+        evidence_ids=[f"EV-{i:03d}" for i in range(46, 51)],
+        bias_warning="中等证据质量；表明 AI 引入的成败取决于教学护栏设计，而非工具本身。"
+    ))
+
+    # 4. Add Risks
+    graph.add_risk(RiskNode(
+        risk_id="RSK-001",
+        risk_type="Scaffolding Dependency Trap (脚手架依赖陷阱)",
+        severity="HIGH",
+        description="学生在有 AI 辅助时表现极其流畅 (+0.64g)，但一旦进入闭卷或无 AI 场景，独立解题与架构迁移能力出现明显倒退 (-0.28g)。",
+        mitigation="实施 4 阶段教学渐进式剥离法 (Fading Scaffold)，每周设置无 AI 手写代码与口试环节。",
+        triggered_by_evidence_ids=["EV-025", "EV-026", "EV-027", "EV-028"]
+    ))
+
+    # 5. Add Gaps
+    graph.add_gap(GapNode(
+        gap_id="GAP-001",
+        gap_type="Measurement/Retention Gap",
+        description="现有文献普遍缺乏 12 周以上的跨学期纵向随访数据，缺乏大二后续课程（如数据结构、操作系统）中的真实迁移留存表现。",
+        target_outcome="跨学期概念留存 (Cross-Semester Retention)",
+        existing_evidence_summary="50篇文献中仅有 2 篇随访超过 1 个月，绝大多数仅评估单次实验课或学期末即时测验。",
+        recommended_trial_design="12周准实验双重差分 (DID) 课堂实证试验，并在下学期初进行无预警无 AI 摸底测试。"
+    ))
+
+    # 6. Set Decision
+    graph.set_decision(DecisionNode(
+        decision_id="DEC-AI-CODING-CS1",
+        verdict="PILOT",
+        confidence_score=0.89,
+        rationale=(
+            "证据表明：无护栏全面放开 AI 助手会带来严重的脚手架依赖与负迁移风险 (-0.28g)，但直接禁止亦违背工业界技术发展趋势。"
+            "因此，裁决为限制性【PILOT（谨慎试点）】——必须严格配套‘解释优先、渐进剥离、闭卷验证’的四阶段教学护栏方案，严禁全面无约束推广。"
+        ),
+        applicability_boundary="适用于高校计算机与工科大一程序设计必修课；严禁在无护栏期末考试或核心算法认证中无限制开放使用。",
+        stop_conditions=[
+            "期中阶段无 AI 测验平均分较对照班下滑超过 15%",
+            "检测到直接复制粘贴 AI 代码且无法口头解释的违规率超过 20%",
+            "学生自陈编程自信度上升但独立手写代码错误率显著激增"
+        ]
+    ))
+
+    return graph
+
+
+def export_all():
+    graph = build_killer_demo_graph()
+    
+    # 1. Export to examples/ai-coding-assistant-50/
+    dir_50 = ROOT / "examples" / "ai-coding-assistant-50"
+    dir_50.mkdir(parents=True, exist_ok=True)
+    (dir_50 / "evidence_graph.json").write_text(graph.to_json(), encoding="utf-8")
+    
+    # 2. Export to examples/ai-coding-assistant/evidence_graph.json
+    dir_orig = ROOT / "examples" / "ai-coding-assistant"
+    (dir_orig / "evidence_graph.json").write_text(graph.to_json(), encoding="utf-8")
+    print(f"[+] Exported 50-study SSOT EvidenceGraph to: {dir_50} and {dir_orig / 'evidence_graph.json'}")
+
+
+if __name__ == "__main__":
+    export_all()