--- license: cc-by-4.0 language: - en pretty_name: "MissionReport: Evaluating AI Judgment in Government Reporting (sample)" task_categories: - text-generation annotations_creators: - expert-generated language_creators: - expert-generated source_datasets: - original size_categories: - n<1K tags: - government - defense - report-writing - document-generation - agents - benchmark configs: - config_name: tasks data_files: - split: test path: data/tasks/*.parquet --- # MissionReport (micro1): Sample Release *Evaluating AI Judgment in Government Reporting* MissionReport measures how well AI agents write government status reports. Each task is an assignment written by an expert who has done the job: a role, a reader, the working files and the organization's template. The agent reads the files and produces a finished Word document. The expert's rubric then scores each decision the report makes: what leads, what the facts mean together, how to report what is not yet known, and what to leave out. The full benchmark has 20 tasks in four report families (weekly activity reports, after-action reports, program status reports, and performance and accomplishment reports) and five settings (Army, Air Force, Navy, Marine Corps, and Department of Defense offices and agencies). In the published run, nine frontier models scored between 54.3% and 64.1%, and 3 of 540 reports met every criterion in their rubric. **Published report:** https://www.micro1.ai/benchmark/mission-report ## Citation Please cite as: **"MissionReport: Evaluating AI Judgment in Government Reporting (micro1). Available at: https://www.micro1.ai/benchmark/mission-report"** ## Scope of this release This is a sample of **5 tasks (25% of MissionReport)**, released for evaluation and benchmarking. The remaining 15 tasks are held out. To request access to the full dataset, contact **maria@micro1.ai** or fill out the enquiry form at https://www.micro1.ai/. Task IDs and titles match the report's Appendix B, and Appendix A shows each task's expert report beside one model's report. Mean score is the task's average over nine frontier models with three attempts each, in the published run. | Task | Report | Setting | Reader | Length limit | Source files | Criteria (penalties) | Mean score | |---|---|---|---|---|---:|---:|---:| | `war-02` · WAR-02 | Staff trip entry for a mobility command | Air Force | the directorate head and the command's commander | 200 words | 1 | 23 (4) | 60.9% | | `war-08` · WAR-08 | Division gunnery qualification weekly activity report | Army | a deputy commanding general | 150 words | 10 | 18 (2) | 54.3% | | `aar-01` · AAR-01 | Wing emergency operations exercise after-action report | Air Force | a wing commander | one full page | 15 | 16 (3) | 60.4% | | `aar-04` · AAR-04 | Armor company training rotation after-action report | Army | the training center chain of command | 1,500 words | 9 | 27 (3) | 52.3% | | `psr-01` · PSR-01 | Career field hiring read-ahead memo | DoD offices and agencies | a senior executive director | two pages, 350 words | 7 | 13 (2) | 40.5% | 42 source files (Word, PDF and Excel), 83 required criteria worth 155 points and 14 penalties worth 20 points. PSR-01 has the lowest mean score of the 20 tasks. **Task versions.** These are the current task versions. Compared with the versions behind the published scores, the prompts, source files, criterion descriptions and point weights are unchanged. Two expert reports (PSR-01 and AAR-04), one grading note and some criterion source notes received typo and formatting fixes. A full re-run on these versions matched the published scores within run-to-run variation. ## Contents A `tasks/` tree of **5 runnable [Harbor](https://harborframework.com) tasks**, plus a Parquet index for the Dataset Viewer and a script that summarizes a finished run. ``` tasks/psr-01/ ├── task.toml # metadata, timeouts, judge settings ├── instruction.md # the assignment shown to the agent ├── environment/Dockerfile # sandbox: Python document libraries, pandoc, poppler-utils, LibreOffice, OpenCode ├── steps/answer/ │ ├── instruction.md # the same assignment, for Harbor's single step │ └── workdir/attachments/ # source files, copied to /app/attachments/ ├── tests/ # verifier, NOT visible to the agent │ ├── test.sh # entrypoint: format check, then the judge, then the reward │ ├── agent_judge.py # the agentic judge │ ├── fixtures/rubric.json # the weighted criteria │ ├── fixtures/config.json # prompt, rubric, grading guidance, judge settings │ └── golden/ # the expert's report, the judge's reference └── solution/solve.sh # oracle: submits the expert's report ``` The agent works in `/app` and must save its report as `/app/outputs/report.docx`, the only file the grader reads. The instruction wraps the expert's prompt with a short role preamble, file-reading tips and the list of attachments. ## Dataset Viewer columns The `tasks/test` split has one row per task. - `task_id`, `title`, `report_family`, `setting`, `reader`, `length_limit`: as in the report's Appendix B - `prompt`: the expert's prompt - `attachments`, `num_attachments`: paths of the source files in the Harbor task - `rubric_count`, `penalty_count`, `positive_points`, `penalty_points` - `rubric`: JSON list of criteria with `criterion_id`, `points` (negative for penalties), `polarity`, `criterion`, `source` (the prompt line or file the criterion rests on) and `justification` - `grading_guidance`: the expert's grading notes. The rubric judge does not see them. - `mean_score_pct`: the task's mean score in the published run - `golden_report_path`, `instruction_path`, `rubric_path`, `task_path`: paths into the Harbor task ## Scoring - **Criteria.** Each criterion is judged true or false. A required criterion earns its points when the report meets it. A penalty subtracts its points when the report commits the error it describes, such as including material the reader does not need. - **Task reward.** Positive points earned minus penalty points triggered, divided by the task's positive points, clamped to [0, 1]. A missing, corrupt or empty `report.docx` scores 0. - **Judge.** Claude Opus 5.5, running as an agent in OpenCode at medium reasoning effort, reads the submitted document, can render its pages, and compares it with the expert's report. It decides each criterion with a written reason. It never sees the point weights, the grading guidance or the source files. ## Running ```bash uv tool install harbor export ANTHROPIC_API_KEY=... # the judge (Claude Opus 5.5) # Check the setup end to end: the oracle submits the expert's report and should score 1.0 harbor run -p tasks/psr-01 -a oracle # Run a model on all five tasks, three attempts each (also export that model's provider key) harbor run -p tasks -a opencode -m -k 3 # Rewards per task, plus each missed criterion and triggered penalty with the judge's reason python scripts/summarize_job.py jobs/ --criteria ``` Harbor runs each task in Docker by default; `-e` selects another sandbox provider. Harbor asks before passing your environment to a task; add `-y` to confirm when running unattended. The published results used the OpenCode agent at version 1.18.31 (add `--ak version=1.18.31` to match) and default reasoning effort, with a 60-minute limit per attempt and three attempts per task. The verifier allows up to 85 minutes per report; a judging pass usually takes a few minutes. ## Important notes - **Sources.** Each task's source files were assembled, written or adapted by an expert with reporting experience in that organization. They name real organizations, ranks, places and publications. AAR-04 includes the Army's public doctrine manual, FM 6-0. - **Not for operational use.** Reports produced on these tasks are not official products. Do not use this dataset for real reporting decisions without independent validation. - **No endorsement.** MissionReport is not affiliated with or endorsed by the U.S. Department of Defense or any military service. Use of this dataset does not imply endorsement by micro1. ## License CC-BY-4.0.