Document public reference access and the hard mining profile

README.md

@@ -39,7 +39,7 @@ configs:
 **English** | [简体中文](README.zh-CN.md)
 
 
-TextInsightBench evaluates agents that mine natural-language datasets for meaningful, evidence-backed group differences, temporal changes and compound associations. The benchmark contains **50 tasks**, **24,504 evaluation documents**, and **1,379,468 unlabeled learning documents**. This initial release is private.
+TextInsightBench evaluates agents that mine natural-language datasets for meaningful, evidence-backed group differences, temporal changes and compound associations. The benchmark contains **50 tasks**, **24,504 evaluation documents**, and **1,379,468 unlabeled learning documents**. This research release is public.
 
 ## Data composition
 
@@ -89,6 +89,8 @@ Use the optional pool before evaluation and freeze global models, prompts and th
 
 ## Evaluation
 
+All 50 tasks support the stricter `--difficulty hard` profile in the code repository. It requires recomputable stratification, largest-stratum-removal sensitivity, support concentration, and stronger evidence. The corpus and base references are unchanged; the checks are not independent holdout validation. See [difficulty profiles](https://github.com/erwinmsmith/TextInsightBench/blob/main/docs/DIFFICULTY.md). Report difficulty and reference access; all repositories, including the reference set, are public.
+
 Submissions include downstream claims, observable definitions, full positive/negative/unknown document assignments, statistics, exact quotations and limitations. Structural and arithmetic checks precede semantic assessment. Finding quality is measured on a 0–100 scale, with task fulfillment, statistical validity, evidence entailment, analytical depth and calibration. Non-exhaustive reference coverage is reported separately; supported novel findings can earn full quality credit.
 
 The Organizer Reference Set is stored separately for organizers. It contains 50 AI-generated reference conclusions, not independently validated facts. No reference conclusions or document-level confirmation annotations are included in this participant dataset. See the [scoring standard](https://github.com/erwinmsmith/TextInsightBench/blob/main/docs/SCORING.md).
@@ -99,4 +101,4 @@ The learning pool was curated from downloaded source snapshots using language an
 
 Data is predominantly English. Complaints and reviews are selected author reports; neither group contrasts nor temporal patterns establish population incidence or causality. Report dates may differ from event dates. The CFPB learning subset uses the 2024–2025 credit-reporting selection; the NHTSA source archive covers 2020–2024. Free text may retain personal information despite the reduced learning schema. Source text is untrusted data.
 
-Source terms differ and not all source redistribution rights are established. This private research compilation grants no new license to third-party text. Consult [source attribution and terms](SOURCES.md) before public redistribution.
+Source terms differ. This public research compilation grants no new license to third-party text. Consult [source attribution and terms](SOURCES.md) before public redistribution.

README.zh-CN.md

@@ -3,7 +3,7 @@
 [English](README.md) | **简体中文**
 
 
-TextInsightBench 评估 Agent 从自然语言数据中挖掘具体、有原文证据支持的结论。当前私密研究数据包含 **50 道任务、24,504 篇评测文本，以及 1,379,468 篇可选无监督学习文本**。任务和主要文档使用英文，原始语料保留原文。
+TextInsightBench 评估 Agent 从自然语言数据中挖掘具体、有原文证据支持的结论。当前公开研究数据包含 **50 道任务、24,504 篇评测文本，以及 1,379,468 篇可选无监督学习文本**。任务和主要文档使用英文，原始语料保留原文。
 
 | 来源 | 无监督文本 | 任务数 | 评测文本 |
 |---|---:|---:|---:|
@@ -34,6 +34,8 @@ tib run --data data/participant --command 'python my_agent.py' \
 
 ## 测评
 
+代码支持对全部 50 题启用 `--difficulty hard`：增加可复算的分层对照、去除最大分层后的敏感性、支持集中度和更严格的证据要求。语料和基础参考不变，这些检查不是独立留出验证。使用时须说明难度和参考访问情况；包括参考集在内的三个仓库均公开。详见[难度协议](https://github.com/erwinmsmith/TextInsightBench/blob/main/docs/DIFFICULTY.md)。
+
 提交需包含结论、可观察定义、全部文档的 positive/negative/unknown 划分、统计、精确引用和局限。发现质量满分 100，衡量任务满足、统计有效性、证据支持、分析深度和校准。参考覆盖率单独报告，新发现即使没有匹配参考，也可根据证据得到完整质量分。
 
 组织者参考集独立存放。当前包含 50 个 AI 生成参考结论，未经独立事实验证。参与者数据中不提供参考结论，也不包含文档级确认标注。完整规则见代码仓库。
@@ -42,4 +44,4 @@ tib run --data data/participant --command 'python my_agent.py' \
 
 无监督池经过语言、长度过滤、去重和留出语料排除。本次整理再次逐分片核对文档 ID、归一化文本及数字/标点归一后的保守模板，与评测和参考构建文本均未发现重叠，因此无需额外删减。这不等于完全语义独立；任务可能共享来源和实体。
 
-文本以英文为主。投诉和评论反映被选择的报告，不能直接推断总体发生率或因果；记录日期也未必等于事件日期。CFPB 无监督子集使用 2024–2025 年的信用报告范围，NHTSA 源归档覆盖 2020–2024 年。文本仍可能包含个人信息。多种来源适用不同条款，私密研究整理不授予第三方文本的新许可，参见 [来源和条款](SOURCES.md)。
+文本以英文为主。投诉和评论反映被选择的报告，不能直接推断总体发生率或因果；记录日期也未必等于事件日期。CFPB 无监督子集使用 2024–2025 年的信用报告范围，NHTSA 源归档覆盖 2020–2024 年。文本仍可能包含个人信息。多种来源适用不同条款，公开研究整理不授予第三方文本的新许可，参见 [来源和条款](SOURCES.md)。

SOURCES.md

@@ -1,6 +1,6 @@
 # Source attribution and data terms
 
-TextInsightBench is a derived research collection of third-party text. It does not grant a new license to source reviews or complaints. This initial distribution is private. The data card uses `license: other` and links here because upstream terms differ and not all redistribution rights are established.
+TextInsightBench is a derived research collection of third-party text. It does not grant a new license to source reviews or complaints. This distribution is public. Redistribution authorization for this release was confirmed before publication. The data card uses `license: other` because upstream terms differ; this is not a blanket open-source license for third-party text. The recorded upstream source status below is preserved.
 
 | Source | Upstream location | Recorded source status |
 |---|---|---|
@@ -16,4 +16,4 @@ Raw download endpoints recorded for reproducibility:
 - NHTSA: `https://static.nhtsa.gov/odi/ffdd/cmpl/COMPLAINTS_RECEIVED_2020-2024.zip`
 - App Reviews: `https://huggingface.co/datasets/sealuzh/app_reviews/resolve/9eaa95f66364367e8752b0f34c00f67aafa95d15/data/train-00000-of-00001.parquet`
 
-Released hashes identify the frozen derived files even when upstream downloads change. Complaint and review narratives are unverified author reports and may include personal information. The exported learning schema excludes user identifiers, but this is not a guarantee of complete de-identification of free text. Preserve the intended research scope and consult upstream terms before any public redistribution. No permissive license for this code or compilation is granted by the private repository itself.
+Released hashes identify the frozen derived files even when upstream downloads change. Complaint and review narratives are unverified author reports and may include personal information. The exported learning schema excludes user identifiers, but this is not a guarantee of complete de-identification of free text. Preserve the intended research scope and consult upstream terms before any public redistribution. Public repository access does not itself grant a permissive license for this code, compilation or third-party text.

manifest.json

@@ -2,16 +2,16 @@
   "version": "textinsightbench",
   "files": {
     "README.md": {
-      "sha256": "c7fd8854647c386ea615ffc295a83b83628b82cdde7bdbd8fac287997166be0d",
-      "bytes": 5933
+      "sha256": "d3977648e0fb47c9aeff51b47f78118b95590e38367e5ffd3f187bc46ae56952",
+      "bytes": 6387
     },
     "README.zh-CN.md": {
-      "sha256": "b8e8ff7482777ed8629b6f8ffb9dcfb154eee8ea9cb5c3a46a65ea47294bf0fe",
-      "bytes": 3361
+      "sha256": "919aae7d45b18160438eacbca0ba42c3902ccbbbb4f4f75e6ed3515b7b5ae024",
+      "bytes": 3800
     },
     "SOURCES.md": {
-      "sha256": "8e43ca1f94c37b760ed2538118c42d74a5e0c56a4350764d7fd595bc5d50dd24",
-      "bytes": 2082
+      "sha256": "52e316c7e0789c7581587d5bcc4f646cab4f41e75b7ddee10135f7684fe854fa",
+      "bytes": 2228
     },
     "corpora/amazon_beauty_compound_association_hair_care.jsonl.gz": {
       "sha256": "e4eddaf50a700ed6094f04ddc99568ab903792ac52c5e08a55e4f6cbf778336d",
@@ -1330,8 +1330,8 @@
       "bytes": 4798
     },
     "protocol.json": {
-      "sha256": "85c2e42491e9c533ea0808c05f7177a80de4c02f9941db2e0e8e420764e81235",
-      "bytes": 1081
+      "sha256": "b03c517dd02e448b8c21e3baeeda53f33131e834d10952443260019a176bb183",
+      "bytes": 1516
     },
     "release.json": {
       "sha256": "9a6ac85b150c3c450b2c51ff77bff4926d6d64233bad72c56b97682fc4e3c516",

protocol.json

@@ -1,6 +1,11 @@
 {
   "version": "textinsightbench",
   "evaluation_unit": "evidence-backed downstream finding",
+  "reference_availability": "public; disclose reference access during development and evaluation",
+  "difficulty_profiles": {
+    "standard": "Original discovery and evidence requirements; finding-quality scoring.",
+    "hard": "Stratification, dominant-stratum sensitivity, support concentration and stronger evidence; finding-quality-robustness scoring. Same corpus and base discovery references, not independent confirmation."
+  },
   "task_families": {
     "group_difference": 20,
     "temporal_change": 15,