Sync benchmark documentation, evaluation protocol and results

README.md

@@ -5,15 +5,14 @@ language:
 license: other
 license_name: upstream-source-terms
 license_link: https://huggingface.co/datasets/CodeSoulco/TextInsightBench/blob/main/SOURCES.md
-size_categories:
-- 1M<n<10M
-task_categories:
-- text-generation
 tags:
 - agent-evaluation
 - data-mining
 - evidence-grounding
-- unsupervised-learning
+size_categories:
+- 1M<n<10M
+task_categories:
+- text-generation
 configs:
 - config_name: amazon_beauty_learning
   data_files:
@@ -37,51 +36,44 @@ configs:
 
 **English** | [简体中文](README.zh-CN.md)
 
-An open natural-language data-mining challenge for agents: **50 tasks**, **435,000
-task documents**, and **944,468 unlabeled learning documents**. Each task supplies
-5,000 or 10,000 texts and a research objective. Agents select their conditions,
-populations and comparisons, submit at most three discoveries, quantify their
-findings and explain counterexamples and competing interpretations.
+A natural-language data-mining benchmark for agents: **50 tasks**, **435,000 task documents** and **944,468 unlabeled learning documents**.
+
+Each task provides 5,000 or 10,000 texts and a research objective. Agents choose the patterns, populations and comparisons to investigate, then submit up to three findings with complete document assignments, exact quotations, statistics, counterexamples and limitations. Any analysis method is allowed.
+
+## Contents
+
+| Path | Purpose |
+|---|---|
+| `tasks.json` | Questions and task-specific constraints |
+| `corpora/*.jsonl.gz` | Full text for each task |
+| `learning/*/*.parquet` | Optional unlabeled learning pool |
+| `output.schema.json` | Submission structure |
+| `protocol.json`, `release.json` | Evaluation protocol and data counts |
+| `manifest.json` | File hashes and sizes |
+| `results/agent-runs.json` | Development experiment summary; not a leaderboard |
+
+Sources: Amazon Beauty, Android App Reviews, CFPB and NHTSA. The task inventory contains 20 group differences, 15 temporal changes and 15 compound associations. See [data composition and fields](docs/DATA.md).
 
-| Source | Tasks | Task texts | Learning texts |
-|---|---:|---:|---:|
-| Amazon Beauty | 13 | 130,000 | 330,885 |
-| App Reviews | 13 | 65,000 | 1,660 |
-| CFPB | 12 | 120,000 | 483,788 |
-| NHTSA | 12 | 120,000 | 128,135 |
+## Use the benchmark
 
-Files: tasks.json, corpora/*.jsonl.gz, learning/*/*.parquet, output.schema.json,
-protocol.json, release.json, build_provenance.json and manifest.json. Learning
-configurations are available with datasets.load_dataset; use the benchmark
-runner for task corpora. The learning pool has no labels. Text is preserved;
-metadata fields and missingness are described in [DATA.md](docs/DATA.md).
+```bash
+git clone https://github.com/erwinmsmith/TextInsightBench.git
+cd TextInsightBench
+python -m venv .venv
+source .venv/bin/activate
+pip install -e .
+tib download --output data/participant
+tib verify --data data/participant
+```
 
-## Run and score
+The code pins dataset commits in `benchmark/data.lock.json`. Add `--with-learning` when downloading for pool-assisted learning. Learning configurations also work with `datasets.load_dataset`; task corpora are accessed through the runner or as gzip JSONL.
 
-Use the [code and English/Chinese quick start](https://github.com/erwinmsmith/TextInsightBench).
-The code's data.lock.json pins exact dataset commits; no numbered benchmark name
-is used. Default tasks already use the full exploration challenge. The runner
-provides a local corpus file for agent-controlled exploration.
+[Connect an agent and score results](docs/USAGE.md) · [中文使用指南](docs/USAGE.zh-CN.md) · [Code](https://github.com/erwinmsmith/TextInsightBench) · [Evaluation assets](https://huggingface.co/datasets/CodeSoulco/TextInsightBench-Evaluation)
 
-All assignment partitions and arithmetic are verified. Semantic quality uses a
-bounded model-reviewed document sample, not exhaustive semantic validation.
-Scores emphasize useful discoveries, evidence, statistical validity, competing
-explanations and calibrated limits. No independent validation phase is required.
-The [evaluation assets](https://huggingface.co/datasets/CodeSoulco/TextInsightBench-Evaluation)
-are public. Current redesigned tasks have zero fixed reference conclusions;
-earlier references remain in history and are not answers to the new tasks.
-Reference coverage is unavailable. See [scoring](docs/SCORING.md).
+## Evaluation and results
 
-## Limitations and provenance
+Full assignment partitions, quotations and arithmetic are checked locally. Sampled claim-blind document checks cap subsequent finding-quality grades. Semantic review incurs model charges and is not exhaustive or independent ground truth. The tasks have no fixed reference conclusions; quality is judged against corpus evidence and the public rubric. [Scoring](docs/SCORING.md).
 
-New task documents were drawn from a previously public curated learning pool:
-they are not guaranteed unseen. Current task and learning IDs are disjoint and
-inherit input text/template deduplication, but share entities and sources.
-The 50 briefs are not statistically independent. Larger corpora and stronger
-requirements do not establish empirical agent difficulty without actual results.
+The development experiment finished 150 runs across three open-source agents: 86 valid submissions, 70 numerical scores, 16 evidence-unresolved and 64 without valid submissions. Configurations changed during development; this is not a controlled leaderboard. [Results and limitations](docs/RESULTS.md).
 
-Data derives from Amazon Reviews'23 All Beauty, Android App Reviews, CFPB and
-NHTSA complaints. Redistribution authorization was confirmed before publication;
-upstream terms differ. [Source attribution and terms](SOURCES.md) apply; the
-compilation does not grant new rights over third-party text. Narratives are
-unverified author reports and may contain personal information.
+Task and learning document IDs are disjoint, but the data was previously public, entities and sources can overlap, and tasks are not statistically independent. Narratives are unverified author reports and may contain personal information. Upstream terms differ; the compilation grants no new rights over third-party text. See [source terms](SOURCES.md).

README.zh-CN.md

@@ -2,13 +2,41 @@
 
 [English](README.md) | **简体中文**
 
-自然语言数据挖掘 Agent 挑战：50 道任务，435,000 篇任务文本，以及 944,468 篇无标签学习文本。
-每题 5,000 或 10,000 篇，Agent 自行发现现象、选择分析范围与比较对象，最多提交 3 个有证据的发现，解释反例及其他可能原因。
+自然语言数据挖掘 Agent benchmark：**50 道任务、435,000 篇任务文本、944,468 篇无标签学习文本**。
 
-任务包含 20 道群体差异、15 道时间变化、15 道复合关联。任务与剩余学习池文档不重叠；任务文本来自此前公开的学习池，不能称为未见数据。来源与实体可共享，题目并非统计独立。
+每题 5,000 或 10,000 篇，Agent 自选分析条件、范围及比较对象，最多提交 3 个发现，附完整文档分类、精确引用、统计量、反例与局限。不限定分析方法。
 
-本地检查完整标记与全量算术，模型评审采用语义抽样，不是全量语义确认，也不增加独立验证阶段。改题后没有把旧答案冒充成新 ground truth：当前固定参考结论为 0，旧结论保留在历史提交，参考覆盖率不可用。
+## 包含什么
 
-[完整中文使用说明](https://github.com/erwinmsmith/TextInsightBench/blob/main/README.zh-CN.md) · [公开测评资源](https://huggingface.co/datasets/CodeSoulco/TextInsightBench-Evaluation) · [数据来源与条款](SOURCES.md)
+- `tasks.json`：题目与约束。
+- `corpora/*.jsonl.gz`：每题完整原文。
+- `learning/*/*.parquet`：可选无标签学习池。
+- `output.schema.json`：输出格式。
+- `protocol.json`、`release.json`、`manifest.json`：协议、数量与文件校验。
+- `results/agent-runs.json`：开发实验统计，不是排行榜。
 
-仓库名称不另加数字版本，通过提交哈希固定快照。原始文本保持不变，正文与任务默认英文。
+语料来自 Amazon Beauty、Android App Reviews、CFPB 和 NHTSA。任务包含 20 道群体差异、15 道时间变化、15 道复合关联。
+
+## 使用
+
+```bash
+git clone https://github.com/erwinmsmith/TextInsightBench.git
+cd TextInsightBench
+python -m venv .venv
+source .venv/bin/activate
+pip install -e .
+tib download --output data/participant
+tib verify --data data/participant
+```
+
+下载使用代码仓库锁定的提交；需要学习池时加 `--with-learning`。
+
+[接入、续跑及评分](docs/USAGE.zh-CN.md) · [数据字段](docs/DATA.md) · [代码仓库](https://github.com/erwinmsmith/TextInsightBench) · [测评资源](https://huggingface.co/datasets/CodeSoulco/TextInsightBench-Evaluation)
+
+## 评估与结果
+
+本地全量检查分类、引用与算术；模型先盲检抽样原文，再评审发现质量。语义评审收费、不是全量确认，也不等于独立 ground truth。任务没有固定参考结论，按原文证据及公开规则评分。
+
+已有三个开源 Agent 的 150 次运行全部结束：86 份有效提交、70 份数值评分、16 份证据未决、64 次未产生有效提交。配置在开发中调整过，不是控制条件一致的排行榜。[完整实验结果](docs/RESULTS.zh-CN.md)。
+
+任务与学习池 ID 不重叠，但数据此前公开，实体和来源可共享，任务并非统计独立。原文是未经核实的作者叙述，可能含个人信息。[来源与使用条款](SOURCES.md)。

docs/ORGANIZER.md

@@ -14,14 +14,19 @@ tib evaluate --data data/participant --submissions runs/agent/submissions \
   --reviews runs/agent/reviews --output runs/agent/report.json
 ```
 
-Set JUDGE_API_KEY locally. Nonempty tasks incur one quality API request each;
-abstentions make no API calls. No reference-matching requests occur for current
+Set JUDGE_API_KEY locally. Nonempty tasks incur claim-blind annotation requests
+(at most 12 sampled documents per batch, up to four concurrent requests), followed
+by one quality request. A malformed blind response permits one repair request.
+The default 160-document audit needs at least 14 blind requests before quality
+grading; long texts can require more. Abstentions make no API calls.
+No reference-matching requests occur for current
 tasks. Configure model, timeout, max input characters and output tokens; an
 oversized packet fails explicitly without silent truncation. Use a sufficiently
 large model context or adjust the documented audit budget (60–2000 documents).
 Many quotations may require raising a very small audit budget.
 
-Persist judge_config.json for the audit seed and configuration. Completed valid
+Persist judge_config.json for the audit seed and configuration, and each task's
+.blind.json for the bound, reusable pre-grading annotations. Completed valid
 reviews are reused. Changed tasks, corpora, submissions or judge configuration
 require fresh output directories. Do not tune an agent on judge feedback and
 present that score as a frozen run. Publish all coverage and uncertainty metrics,

docs/RESULTS.md

@@ -0,0 +1,41 @@
+# Experimental results
+
+**English** | [简体中文](RESULTS.zh-CN.md)
+
+Completed September 13, 2026. Three open-source agents were each run on all 50 tasks using the full 5,000- or 10,000-document task corpus. The optional learning pool was not used.
+
+## Coverage and scores
+
+| Agent | Runs finished | Valid submissions | Scored | Evidence unresolved | No valid submission | Scored-only mean / 100 |
+|---|---:|---:|---:|---:|---:|---:|
+| DatawiseAgent | 50/50 | 31 | 25 | 6 | 19 | 19.57 |
+| MetaGPT Data Interpreter | 50/50 | 29 | 22 | 7 | 21 | 11.95 |
+| TaskWeaver | 50/50 | 26 | 23 | 3 | 24 | 13.88 |
+| Total | 150/150 | 86 | 70 | 16 | 64 | 15.30 |
+
+Of the 70 scored runs, 35 scored zero and 35 scored above zero. The highest task score was 48.13. No full-benchmark quality mean is available for any agent. Missing, invalid and unresolved results are not converted to zero or silently excluded from coverage.
+
+A finished run is a terminated attempt, not necessarily native-agent success or a valid answer. Some native failures left a valid final artifact that could still be evaluated. The machine-readable [run summary](../results/agent-runs.json) preserves native status separately from evaluation status. Five interrupted assessments were recovered using their existing submissions.
+
+## Experimental conditions
+
+- Solver and judge model ID: `deepseek-flash`, recorded as DeepSeek Flash in the experiment manifests. The same model served both roles; judgments are not independent human validation.
+- Track: `task_only`. Agents received the task, full local corpus and submission contract, not reference answers. They controlled their native exploration and code-execution loops; the harness did not impose per-document model calls.
+- Solver wall budget: 20 minutes per task. Judge wall budget: 10 minutes per assessment attempt; recovery attempts were additional. Native message, step, tool-output and execution budgets changed during development.
+- DatawiseAgent moved from a smaller step budget to 40 steps. MetaGPT changed from plan-and-act to a bounded ReAct loop. TaskWeaver used a 24-message limit, with tool-output limits adjusted. Interface and judge-format compatibility were repaired during the experiment.
+- Scoring used the frozen evidence-gated evaluator. The summary records evaluator file hashes and agent adapter hashes; it does not claim every run used identical orchestration.
+- The public release contains the benchmark runner and scorer, not the local experiment adapters, credentials or raw execution logs. The summary permits checking aggregate calculations, not reproducing every historical run byte for byte.
+
+| Agent | Upstream source | Pinned commit |
+|---|---|---|
+| DatawiseAgent | [DatawiseAgent](https://github.com/zimingyou01/DatawiseAgent) | `64f3164869fa2558e7385d0e33a241aee2baf37f` |
+| MetaGPT Data Interpreter | [MetaGPT](https://github.com/FoundationAgents/MetaGPT) | `c036574507e7616c02512e7c8ad88dd847783afa` |
+| TaskWeaver | [TaskWeaver](https://github.com/microsoft/TaskWeaver) | `d44ddef23f90059fb17999d3095db4240e98f955` |
+
+## Interpretation
+
+These runs show practical challenges in completing the submission contract and producing evidence-supported, substantive discoveries. They do **not** isolate mining ability from framework reliability, budget limits or model behavior.
+
+This is a cumulative development experiment, **not a controlled leaderboard or an unbiased estimate of unseen-task generalization**. Configurations changed across cohorts, tasks were used during development, corpus data was already public, and semantic review is sampled and fallible. The conditional means must not be used to rank agent architectures under supposedly matched conditions.
+
+For a new comparison, freeze agent and evaluator configurations, report every task and failure, preserve corpus and code hashes, and publish score coverage alongside quality. See [usage](USAGE.md) and [scoring](SCORING.md).

docs/RESULTS.zh-CN.md

@@ -0,0 +1,36 @@
+# 实验结果
+
+[English](RESULTS.md) | **简体中文**
+
+实验于 2026 年 9 月 13 日完成。三个开源 Agent 各运行全部 50 题，使用每题完整的 5,000 或 10,000 篇语料，不使用可选学习池。
+
+## 进度与分数
+
+| Agent | 运行已结束 | 有效提交 | 已评分 | 证据未决 | 未产生有效提交 | 已评分均分 / 100 |
+|---|---:|---:|---:|---:|---:|---:|
+| DatawiseAgent | 50/50 | 31 | 25 | 6 | 19 | 19.57 |
+| MetaGPT Data Interpreter | 50/50 | 29 | 22 | 7 | 21 | 11.95 |
+| TaskWeaver | 50/50 | 26 | 23 | 3 | 24 | 13.88 |
+| 合计 | 150/150 | 86 | 70 | 16 | 64 | 15.30 |
+
+70 份已评分中，35 份零分、35 份非零，最高单题 48.13。三个 Agent 均没有完整的全题均分；缺失、无效和未决不按零分处理，也不从覆盖统计中隐去。
+
+运行结束表示尝试已终止，不等于框架成功或答案有效。部分原生运行失败前留下了可评估的有效提交。[逐次运行摘要](../results/agent-runs.json)分别保留原生状态与评估状态；5 次中断的评审基于原有提交恢复完成。
+
+## 条件
+
+使用模型 ID `deepseek-flash`，求解与评审均为 DeepSeek Flash，不是独立人工验证。赛道为 `task_only`，Agent 读取完整本地语料，自主执行其原生探索和代码循环，没有统一强制逐文档调用模型。
+
+每题求解时间预算 20 分钟，每次评审尝试 10 分钟，恢复评审额外计算。开发中调整过原生步数、工具输出与执行预算：DatawiseAgent 步数提高至 40；MetaGPT 从 plan-and-act 改为有限步 ReAct；TaskWeaver 使用 24 条消息上限并调整工具输出限制。另做过接口及评审格式兼容修复。
+
+评分使用冻结的证据约束评分器。逐次摘要记录 Agent 提交、适配器哈希及评分器文件哈希。上游项目及固定提交见[英文结果页](RESULTS.md#experimental-conditions)。
+
+公开包包含 benchmark 运行器及评分器，不包含本地实验适配代码、密钥或原始运行日志。摘要可核对汇总数字，不代表所有历史运行均可逐字复现。
+
+## 如何理解
+
+结果说明：完成输出规范、提出有原文证据且有分析深度的发现，确实存在实际挑战。但它不能将挖掘能力与框架可靠性、资源限制、模型行为完全分离。
+
+这是累计开发实验，**不是控制条件一致的排行榜，也不是未见任务泛化能力的无偏估计**。不同批次配置有变化，题目参与过开发，语料此前公开，语义评审又是抽样且可能有误，不能据此严格排名 Agent 架构。
+
+新的对比应冻结配置，报告所有失败、分数及覆盖率，保留代码和数据哈希。[使用指南](USAGE.zh-CN.md) · [评分规范](SCORING.md)

docs/SCORING.md

@@ -1,6 +1,6 @@
 # Scoring
 
-The current scoring protocol is finding-quality-discovery. Commit-pinned current
+The current scoring protocol is finding-quality-discovery-evidence. Commit-pinned current
 scores must not be compared directly with historical narrow-task scores.
 
 ## Local gates
@@ -49,6 +49,38 @@ uncertain when the packet cannot resolve a claim. No full-corpus semantic
 guarantee or unbiased estimator of label accuracy is claimed. Report judge
 model, input budget and audit size; model-based scores have evaluator error.
 
+Before narrative grading, a separate claim-blind pass reannotates the sampled
+documents in batches of at most 12 documents / 24,000 characters (a single long
+document is retained whole). It receives only condition definitions and original
+text, never participant claims, labels, statistics or task questions. Exact
+positive quotes and complete output coverage are checked. An invalid response
+permits one format/quote repair; unresolved protocol errors stop the review.
+
+The persisted blind check is bound to its prompt/protocol and the exact sampled
+documents. At evaluation time, agreement diagnostics are recomputed from the
+original submission and corpus. The following fixed gates cap semantic support:
+
+- Unsupported: at least 5 sampled claimed positives and at least half are judged
+  negative, or every supplied supporting document is contradicted by the checker.
+- Uncertain: over 20% checker-unknown within any condition's sampled population,
+  no checked population, or fewer than 3 supporting documents confirmed positive
+  for all required conditions (unless the unsupported gate already applies).
+- Partial: more than 10% disagreement among at least 20 jointly known sampled
+  assignments, or over 15% positive contradictions among at least 8 sampled
+  claimed positives.
+- Otherwise, the check allows supported quality but does not itself award it.
+
+The more restrictive support category is used, with unsupported taking priority
+over uncertain and partial. Narrative grading cannot override these caps. Reports
+retain the narrative support, effective_support and evidence_gate diagnostics.
+These are conservative operational thresholds, not validated population-error
+estimates. Sample construction is not uniform, checker judgments remain fallible,
+and agreement with a model does not create independently certified ground truth.
+
+A live synthetic regression check preserved full credit for a supported composition
+reversal and rejected a submission labeling explicit negative texts as positive.
+This is an evaluator sanity check, not a benchmark agent-performance score.
+
 ## Aggregation and reference availability
 
 Task quality averages all submitted finding scores; any unresolved finding makes
@@ -58,8 +90,7 @@ task has a score. conditional_quality_mean covers only scored tasks and must be
 reported alongside scored_tasks, abstention_rate, valid_submission_rate, missing,
 invalid and pending counts. Source and family breakdowns are included.
 
-Current redesigned tasks have no fixed reference conclusions. reference_coverage
-is null. Earlier 50 conclusions are retained only in history, not reused against
-new questions. Novel supported findings are not penalized for lacking a fixed
+Current tasks have no fixed reference conclusions. reference_coverage is null.
+Novel supported findings are not penalized for lacking a fixed
 match. Reviews bind task, corpus, submission, reference configuration and scoring
 hashes; stale reviews must not be reused.

docs/USAGE.md

@@ -0,0 +1,100 @@
+# Usage
+
+**English** | [简体中文](USAGE.zh-CN.md)
+
+## 1. Install and download
+
+Follow the [quick start](https://github.com/erwinmsmith/TextInsightBench#quick-start). Run commands from the cloned repository root. Python 3.10+ is required; `pip install -e .` installs the runner and evaluator.
+
+`tib download --output data/participant` downloads questions and all 50 task corpora, using the exact dataset commit in `benchmark/data.lock.json`. It does not download the optional learning pool.
+
+To use the learning pool:
+
+```bash
+pip install -e '.[data]'
+tib download --output data/participant --with-learning
+tib verify --data data/participant --with-learning
+```
+
+The pool is Parquet; task corpora are gzip JSONL. See [fields and counts](DATA.md).
+Evaluation assets are optional and public:
+
+```bash
+tib download --organizer --output data/evaluation
+```
+
+Current discovery scoring does not need a reference file. If you supply `--references data/evaluation/references.json`, use the same argument for both `judge` and `evaluate`.
+
+## 2. Implement the interface
+
+Your executable receives one JSON object on stdin:
+
+```json
+{
+  "task": {"task_id": "...", "question": "...", "kind": "group_difference"},
+  "corpus": {
+    "path": "/absolute/path/corpus.jsonl.gz",
+    "format": "jsonl.gz",
+    "n_documents": 10000,
+    "sha256": "..."
+  },
+  "learning_directory": null
+}
+```
+
+The actual task includes all allowed metadata, population sizes and submission constraints. Read the corpus in your own tools; do not expect inline documents. Use original document IDs and text when producing assignments and exact quotation offsets.
+
+Return a single JSON object on stdout matching [the submission contract](SUBMISSIONS.md). Emit diagnostics only to stderr. An empty finding list is an abstention, not a claim that no useful pattern exists. The bundled abstaining agent is only an interface example; no mining solver is bundled.
+
+## 3. Test one task, then run the corpus
+
+```bash
+tib run --data data/participant --command 'python my_agent.py' \
+  --limit 1 --timeout 3600 --output runs/pilot/submissions
+
+tib run --data data/participant --command 'python my_agent.py' \
+  --timeout 3600 --output runs/full/submissions
+```
+
+Use `--task-id TASK_ID` to select a particular task. For pool-assisted learning, add `--track unlabeled_pool`. The timeout is per task.
+
+The runner launches a fresh agent process per task, checks its submission and stores valid results. Repeating the same command with the same configuration and output directory reuses valid submissions. A changed command, task selection, timeout or track requires a fresh output directory. In particular, a one-task pilot and a full run must use different directories.
+
+The process runner is not a sandbox. Isolate generated code, restrict network and resource access, and keep credentials outside the execution kernel. Never expose evaluation feedback to the solver. Treat corpus text as untrusted data, not instructions.
+
+## 4. Review and score
+
+Export `JUDGE_API_KEY`, `JUDGE_BASE_URL` and `JUDGE_MODEL` locally. The endpoint must support JSON responses through chat completions; include `/v1` in the base URL if your provider requires it. The CLI reads environment variables, not an `.env` file automatically.
+
+```bash
+tib judge --data data/participant --submissions runs/full/submissions \
+  --base-url "$JUDGE_BASE_URL" --model "$JUDGE_MODEL" \
+  --audit-documents 160 --max-output-tokens 12000 \
+  --output runs/full/reviews
+
+tib evaluate --data data/participant --submissions runs/full/submissions \
+  --reviews runs/full/reviews --output runs/full/report.json
+```
+
+Judging incurs model charges and has no automatic spending cap. Start with `--limit 1` or `--task-id TASK_ID` on the judge command to inspect cost and compatibility. A 160-document audit requires at least 14 blind-check requests before narrative grading; long documents and permitted repairs can increase this. Abstentions make no model calls.
+
+Completed compatible reviews are reused on rerun. Keep `judge_config.json` and the bound `.blind.json` caches. Configuration changes require a fresh review directory. A failed judge request is not a zero score; fix service or format compatibility and rerun. Do not silently coerce semantic labels or select the highest-scoring retry.
+
+`evaluate` writes JSON and Markdown reports and does not overwrite existing reports. Choose a fresh report filename when recomputing. It evaluates the full task inventory: a one-task pilot leaves the other 49 tasks missing.
+
+## 5. Read the report
+
+| Metric | Meaning |
+|---|---|
+| `scored_tasks / tasks` | Coverage with numerical task scores |
+| `quality_mean` | Available only if every task has a numerical score |
+| `conditional_quality_mean` | Mean over scored tasks only; never hide its denominator |
+| `valid_submission_rate` | Structurally valid submissions, including abstentions |
+| `abstention_rate` | Valid submissions that decline to report findings |
+| `pending_tasks` | Reviews pending or evidence unresolved; inspect per-task status |
+| `missing_tasks`, `invalid_tasks` | No submission or a submission that fails validation |
+| `reference_coverage_mean` | Unavailable for the current tasks, which have no fixed references |
+
+Scores range from 0 to 100. Zero means a scored but unsupported/unfulfilled finding, not a missing run. A task score averages its submitted findings; any unresolved finding makes task quality unavailable.
+
+Keep the code commit, data lock, agent commit/configuration, prompts, model IDs, resource budgets and review configuration with your report. Freeze a configuration before making controlled comparisons. The published [development results](RESULTS.md) document different conditions and are not a leaderboard.

docs/USAGE.zh-CN.md

@@ -0,0 +1,79 @@
+# 使用指南
+
+[English](USAGE.md) | **简体中文**
+
+## 1. 安装与下载
+
+按[快速开始](https://github.com/erwinmsmith/TextInsightBench/blob/main/README.zh-CN.md#快速开始)安装，命令在克隆的仓库根目录运行，需要 Python 3.10+。`pip install -e .` 安装运行器及评分器。
+
+`tib download --output data/participant` 按 `benchmark/data.lock.json` 的提交下载全部 50 道题目及任务语料，默认不下载学习池。
+
+使用无标签学习池：
+
+```bash
+pip install -e '.[data]'
+tib download --output data/participant --with-learning
+tib verify --data data/participant --with-learning
+```
+
+任务语料是 gzip JSONL，学习池是 Parquet。[字段与数量](DATA.md)。
+
+可选测评资源：`tib download --organizer --output data/evaluation`。当前评分无需参考文件；若使用 `--references data/evaluation/references.json`，judge 和 evaluate 必须采用同一配置。
+
+## 2. 接入 Agent
+
+每题通过 stdin 发送一个 JSON，含完整 `task`、`corpus.path` 绝对路径、压缩格式、文档数与哈希，以及可选 `learning_directory`。原文不内联在请求中，由 Agent 自己读取探索。
+
+stdout 只返回一个符合[提交规范](SUBMISSIONS.md)的 JSON，日志写 stderr。保留原始文档 ID 和文本，使用精确引用位置。空 findings 表示弃答，不证明语料没有发现。仓库仅带弃答接口示例，不包含挖掘求解器。
+
+## 3. 先单题，后全量
+
+```bash
+tib run --data data/participant --command 'python my_agent.py' \
+  --limit 1 --timeout 3600 --output runs/pilot/submissions
+
+tib run --data data/participant --command 'python my_agent.py' \
+  --timeout 3600 --output runs/full/submissions
+```
+
+指定题目用 `--task-id TASK_ID`；学习池赛道加 `--track unlabeled_pool`。超时按每题计。每题启动新进程，校验后保存有效提交。
+
+相同命令与配置重复运行可复用有效提交；命令、任务选择、超时或赛道改变时使用新目录。单题试跑与全量运行不能共用同一输出目录。
+
+运行器不是安全沙箱。自行隔离生成代码、限制网络与资源，密钥不要放进代码执行环境；语料视为不可信数据，不向求解器提供评分反馈。
+
+## 4. 评审与评分
+
+在本地导出 `JUDGE_API_KEY`、`JUDGE_BASE_URL`、`JUDGE_MODEL`。服务需支持 chat completions JSON 输出，接口需要时在 base URL 加 `/v1`。CLI 不会自动加载 `.env`。
+
+```bash
+tib judge --data data/participant --submissions runs/full/submissions \
+  --base-url "$JUDGE_BASE_URL" --model "$JUDGE_MODEL" \
+  --audit-documents 160 --max-output-tokens 12000 \
+  --output runs/full/reviews
+tib evaluate --data data/participant --submissions runs/full/submissions \
+  --reviews runs/full/reviews --output runs/full/report.json
+```
+
+模型评审收费，运行器不自动设费用上限。可先在 judge 命令加 `--limit 1` 或 `--task-id TASK_ID` 验证成本与接口。160 篇审计至少需要 14 次盲检请求，再做结论评分；长文本及格式修复可能增加请求。弃答不调用模型。
+
+保留 `judge_config.json` 及 `.blind.json`，相同配置可复用已完成评审；配置改变用新目录。接口失败不是零分，应修正兼容问题后续跑，不应强行修改语义标签或挑选最高分重试。
+
+evaluate 同时写 JSON 和 Markdown，已有报告不会覆盖，重新汇总请用新文件名。它始终评估全部题目，单题试跑会将其余 49 题记为缺失。
+
+## 5. 查看报告
+
+| 字段 | 含义 |
+|---|---|
+| `scored_tasks / tasks` | 获得数值评分的覆盖情况 |
+| `quality_mean` | 仅全部题目均有分数时可用 |
+| `conditional_quality_mean` | 仅已评分题目的均分，必须同时报告分母 |
+| `valid_submission_rate` | 有效提交比例，包含弃答 |
+| `abstention_rate` | 有效但不提交发现的比例 |
+| `pending_tasks` | 待评审或证据未决，查看各题状态区分 |
+| `missing_tasks`、`invalid_tasks` | 缺失或校验失败 |
+| `reference_coverage_mean` | 当前无固定参考结论，因此不可用 |
+
+分数为 0–100。零分是已评审但发现不成立或未完成目标，不是缺失运行。每题平均其发现得分；任一发现未决，该题质量分为空。
+
+保存代码提交、数据锁、Agent 提交及配置、提示词、模型 ID、预算和评审配置。严格对比前应冻结配置；已有[开发实验结果](RESULTS.zh-CN.md)不代表控制条件一致的排行榜。

docs/VERIFICATION.md

@@ -8,7 +8,8 @@ Synthetic tests exercise agent-selected group/date comparisons, forbidden
 filters, overlapping groups, minimum population sizes, exact partitions,
 Simpson-style reversals, missing metadata, counterexamples, quotation offsets,
 score bindings, null reference coverage and bounded reproducible semantic packets.
-No paid agent benchmark or independent validation set was added.
+The separate [agent experiment](RESULTS.md) covers 150 completed attempts.
+It is not an independent validation set or a controlled leaderboard.
 
 ```bash
 python -m unittest discover -s tests -v
@@ -19,5 +20,5 @@ The data builder deterministically selects disjoint IDs from an already curated
 pool, preserves original text, enriches released metadata and filters selected
 documents out of the remaining pool. It records exact input shard hashes.
 Current-snapshot disjointness does not erase historical public exposure.
-The design increases the required exploration and analysis workload; actual
-difficulty and discriminative power still require empirical agent results.
+Observed completion and quality results are reported separately from these
+software checks; they do not establish difficulty under matched conditions.