Edition and verification scope
Edition 0.7.0 contains all 14 modules, eight teaching pages per module, public scripts, clean notebooks, worked feedback, instructor guidance and the independent phishing case. Begin with readiness and learning routes. Reviewers should use the independent-review guide.
How this edition is verified
Every change to main runs the full workflow: core checks on Linux (Python 3.12) and Windows (Python 3.12 and 3.13), framework tests and notebooks on Linux CPU, and the complete site, download and browser checks. Check the workflow results for the exact commit you use; a failed, cancelled or timed-out run does not verify that revision.
Reference environments and evidence
| Area | Evidence for this edition | Scope |
|---|---|---|
| Core processing/modeling | Linux CI, Python 3.12; Windows CI, Python 3.12 and 3.13 | Scripts, behavioral tests and fresh-kernel notebooks |
| Frameworks | Linux CPU CI, Python 3.12 | TensorFlow learning/reload, local LLM generation, adapters, app forms and health |
| Notebook execution | All 14 module notebooks and the phishing notebook | Clean-kernel execution; interface usability is a separate question |
| Reading site | Source/generated links, HTML structure, HTTP routes and downloads | Selected core, optional-project and advanced pages also receive scripted focus, 320px reflow and selected axe checks |
| Distribution | Tracked-source archive integrity, attributed data caches and installed-wheel inference | Source/provenance/schema/cache behavior |
Use the workflow results for exact commit-specific results. The maintainer guide gives reproduction commands. macOS, Windows framework execution and live hosted-provider sessions are outside the demonstrated platform scope; use the Linux CPU reference for framework work.
Interpreting the examples
Early fixtures are labeled teaching data. Later modeling uses attributed public datasets and a fixed local-model revision. Tiny neural/LLM runs demonstrate mechanisms and may underperform simple baselines. Current framework results include successful adaptation as well as adverse comparisons; use the linked results and protocol-specific denominators.
Adult results are descriptive audits, batch replay is not a measured timeline, and a legal tool plan may misunderstand intent. Historical benchmark performance and successful execution do not establish operational suitability.
Verified results
Modules 8–11 and 14: use the current framework results. The archived framework baseline describes earlier weak configurations and is not the current result.
The core rows below were re-executed from the pinned sources during the October 28, 2026 review (Windows, Python 3.13). Module 12 is quoted from its worked feedback; its framework run is covered by the Linux framework CI job.
| Module | Data and split | Selected method | Final result | Comparison |
|---|---|---|---|---|
| 5 | UCI Bike Sharing; 504/110/110 days | previous_day persistence (validation MAE 733.3) |
MAE 926.6 rentals/day; largest residual 4,470 on 2012-10-31 | Both linear candidates lost to persistence on validation (888.5 and 1,102.9) |
| 6 | UCI red wine, quality ≥ 7; 243 test | Logistic regression, threshold 0.3 | TP 19, FP 30, FN 10, TN 184; recall 0.66, balanced accuracy 0.757 | Dummy predicts no positives |
| 7 | UCI red wine chemistry; 243 test | k = 2 (validation silhouette 0.206) | Cluster IDs are not quality classes | k = 3: 0.184; k = 4: 0.162 |
| 12 | Module 6 model replayed on 243 held-out records in 4 batches | Logistic, threshold 0.3 | Batch recall 0.40–0.73; batch 3 selection rate 0.41 vs 0.10–0.16 | Batches follow record order, not time |
| 13 | UCI Adult; 16,253 test | Logistic on 5 numeric fields, threshold 0.3 | Recall 0.62, balanced accuracy 0.729 | Selection rate: female 0.20, male 0.31; smallest group n = 135 |
| Phishing | UCI Phishing Websites; 2,107 test | Random forest, threshold 0.5 | TP 844, FN 60, FP 58; recall 0.934, balanced accuracy 0.943 | Dummy predicts no phishing |
Recorded development executions and current CI status are separate evidence. Check the exact commit in Actions: a failed, cancelled or timed-out job does not verify that revision. A learner pilot and human accessibility reviews have not yet been carried out; see below.
Independent-review questions
Review learner workload, unaided task completion, explanations, feedback and transfer across starting experience. Human screen-reader/keyboard/real-zoom sessions and notebook/app interface evaluation have no recorded human test results in this edition; the accessibility evidence log identifies their scope. Automated checks do not establish WCAG conformance or learner effectiveness. Use the pilot instrument to record observations rather than treating an unperformed study as a pass.
The architecture, source register and archived learning-support review provide implementation context.