# Verification scope and limitations

Edition: 2026-10-14.1. The teaching package is intended for review and future integration; no site or service was published.

## Actual execution evidence

- CPython 3.12.14, Linux x86_64; exact dependency versions are in requirements.txt and the generated reference report
- A new isolated virtual environment installed every pinned requirement successfully; all 26 tests passed there
- Deterministic CSV regeneration checked byte-for-byte against all four supplied partitions
- All ten example scripts executed successfully
- Reference development and final evaluation executed with recorded versions and frozen source/development-data hashes
- Independent instrumentation verified that development does not open test.csv and the final path performs no threshold or hyperparameter selection
- Twenty-six reference unit tests passed, covering data integrity, split/label availability, preprocessing, frozen calibration, metrics and policy edge cases
- Separate learner function checks and classifier/regressor capstone smoke checks passed against the reference implementations; incomplete starters are expected to fail
- Independent review manually recomputed final regression errors, confusion metrics, log loss, Brier score, pairwise ROC AUC and average precision from saved predictions
- All eight original SVGs were rendered and visually inspected; labels and source-derived numbers were checked
- All 36 exercise answers and 24 exit-check answers were independently reviewed
- Independent final review passed 10 core review groups, 5 learner-runner groups and 8 static/build groups; both HTML pages, the coursebook and all SVGs regenerated byte-identically
- Learner final evaluation rejected changed source/runtime and deliberately corrupted frozen policy predictions before reading the test; the learner threshold selector was not called during final evaluation
- Both HTML pages have valid local links/fragments and unique IDs; all eight guide images have alternative text and no scripts are embedded

Full reference unit tests inspect final-test data for invariant checking and belong after the learner's final reveal. They are not a pre-evaluation training step. Public learner checks use training fixtures only. The final-test files and answers are physically accessible in this transparent course; hashes discourage accidental drift but do not enforce exam security.

## Interpretation of the reference result

The reference case was generated and executed, not assigned invented model scores. Its final results are in outputs/final_test_results.json and the separate answer key. The final operational conclusion must account for both predictive performance and capacity; a threshold feasible on validation can fail on the future window. Do not tune on that revealed test and call the changed result independent.

Synthetic inputs include missingness, correlated predictors, repeat depots, after-outcome traps and a deliberately changing future process. These are pedagogical choices, not calibrated simulations of any real organization's data. There is no evidence here about production accuracy, causal intervention effects, fairness for real populations or actual financial savings.

## Unverified scope

Browser layout remains unverified: an isolated headless Chromium launch was attempted on the authoring cloud computer, but environment process/socket restrictions prevented it from starting. HTML structure, local navigation targets and asset paths were checked statically. macOS and Windows installation has not been executed. The browser-based Aabha Python lab is not a verified execution route for this package. Screen-reader behavior, all assistive technology combinations, printed page layouts and a beginner learner pilot remain unverified. Semantic HTML, image descriptions, visible keyboard focus and responsive styling are provided, but these do not constitute an accessibility certification. Suggested duration and assessment thresholds are editorial proposals.

No model is deployed, no monitoring service runs, no users are scored, and no credentials or paid services are required. Reference source and generated CSV/JSON/SVG are supplied; no pickle, joblib model binary or embedded executable is distributed.