English | 中文

# Real case: completing a machine-learning experiment This directory preserves one public VibeSkills case that passed final delivery acceptance. The task used scikit-learn's bundled Wisconsin Breast Cancer dataset and delivered a data audit, baseline model, statistical review, result figures, scientific report, and group-meeting slides. The run ID is `20260718T041559Z-51996499`. This is a software-reproducibility case. It is not clinical validation and must not be used for diagnosis or patient decisions. ## Case execution After requirement approval, this task used the `L` workflow and proceeded in order. VibeSkills shortlisted relevant capabilities from the configured local Skill folders, arranged the work into 9 modules and 10 work units, and the current Agent completed and recorded the plan. | Item | Result | |:---|:---| | Local Skill scope | More than 100 available Skills were counted in the configured folders during publication preparation | | Selected for this run | 7 Skills | | Work plan | 9 modules completed as 10 serial work units | | Execution | 10 completed, 0 failed, 0 blocked | | Module acceptance | 18 criteria passed | | Cross-artifact checks | 17 / 17 passed | | Final acceptance | `PASS`, with readiness `fully_ready` | The root README groups the 9 modules into 5 work groups for a quicker overview. The original plan and execution records retain every module and work unit. ### How the Skills took part VibeSkills searched the configured local Skill folders, read the shortlisted candidates' `SKILL.md` files, and checked their purposes and limits. The case then used these 7 Skills. | Skill | Assigned work | |:---|:---| | `exploratory-data-analysis` | Materialize the public data and audit structure, quality, class balance, duplicates, and leakage risk | | `scikit-learn` | Build the frozen comparator and logistic-regression baseline, then verify exact replay | | `statistical-analysis` | Calculate variability and uncertainty intervals and state their assumptions | | `scientific-critical-thinking` | Review bias, leakage, generalization limits, and unsupported claims | | `scientific-visualization` | Produce 4 source-mapped PNG figures and matching SVG files | | `sciwrite` | Review the scientific report in 5 passes without changing the approved scientific content | | `presentations` | Build the 7-slide deck, render every slide, and inspect layout and data | The complete selection record is in [`selected-skills.json`](./evidence/selected-skills.json). The local Skill count comes from [`skill-inventory-snapshot.json`](./evidence/skill-inventory-snapshot.json), captured on the same host while preparing this public case. It is not a count emitted by the accepted runtime. ### Execution and checks All 10 work units completed with no failure or blocked state. The case then ran 17 consistency checks across module outputs, plan binding, data and model contracts, statistics, figures, the report, slides, and publication boundaries. All 17 passed. ## Final delivery The final delivery contains 4 result figures, a scientific report, and a 7-slide group-meeting deck. Every item comes from the same accepted execution record. | Data and model results | Delivery checks | |:---:|:---:| | ![Class balance](./assets/class-balance.png) | ![Confusion matrices](./assets/confusion-matrix.png) | | ![ROC and precision-recall curves](./assets/roc-and-pr-curves.png) | ![Cross-validation distributions](./assets/cv-performance-distribution.png) | All 4 figures also have [SVG versions](./assets/vector/). The final written report is [`scientific-report.md`](./outputs/scientific-report.md). The 7-slide group-meeting deck is available as a [`PPTX`](./outputs/group-meeting-slides.pptx), with the rendered montage below. ![Rendered overview of the seven-slide deck](./assets/slides-montage.png) Final acceptance is `PASS` with readiness `fully_ready`. See the detailed result in [`delivery-acceptance-report.md`](./evidence/delivery-acceptance-report.md). ## Execution records and reproduction materials These files preserve the path from the original request to final acceptance. Every number used on the root README can be traced to one of them. | Stage | Material | |:---|:---| | Original task | [`original-task.md`](./evidence/original-task.md) | | Confirmed requirement | [`requirement.md`](./evidence/requirement.md) | | Execution plan | [`execution-plan.md`](./evidence/execution-plan.md) · [`module-work-plan.json`](./evidence/module-work-plan.json) | | Skill selection | [`selected-skills.json`](./evidence/selected-skills.json) · [`skill-inventory-snapshot.json`](./evidence/skill-inventory-snapshot.json) | | Actual execution | [`module-execution.json`](./evidence/module-execution.json) | | Artifact consistency | [`consistency-check.json`](./evidence/consistency-check.json) | | Final acceptance | [`delivery-acceptance-report.md`](./evidence/delivery-acceptance-report.md) · [`JSON`](./evidence/delivery-acceptance-report.json) | | Public-copy notes | [`PUBLICATION-NOTES.md`](./PUBLICATION-NOTES.md) | [`case-manifest.json`](./evidence/case-manifest.json) and [`consistency-check.json`](./evidence/consistency-check.json) were generated before final acceptance, so they preserve the pending state at that point in the run. The later `module-execution.json` and `delivery-acceptance-report.*` files record the final status. ## Reproduce the core experiment The public package retains the data-audit, baseline, uncertainty, and figure scripts. Confirm that `python` points to Python 3.12.12, then run these commands from the repository root: ```powershell python --version python -m venv .\docs\cases\ml-experiment\reproduce\.venv .\docs\cases\ml-experiment\reproduce\.venv\Scripts\python.exe -m pip install -r .\docs\cases\ml-experiment\reproduce\requirements.lock pwsh .\docs\cases\ml-experiment\reproduce\reproduce.ps1 ``` The script writes to `reproduce/generated/`, runs the frozen baseline twice, and checks that the generated `metrics.json` exactly matches the accepted result. The generated directory and virtual environment are excluded from Git.