Skip to practice lab
QentraCOMPANION / 1.0.0

SHIPPING RELIABLE AI / BY QENTRA.TECH

See a failure.
Understand the fix.

A guided practice space for the book. Generate fictional support cases, run checks and see exactly what changed. Start in your browser, even if you have never used a terminal.

Start the browser lab

Free and public · No account · No API key · No model charges

Practice, with a clear boundary.All cases are synthetic. Responses are scripted, not generated by a live AI model. Checks run in this browser; your inputs are not submitted to a demo server. A green result is not proof that a real AI application is safe.

01 / NO-INSTALL PATH

Your first evaluation.

Use the steps in order. Every red result has a reason and a next action.

Start with Generate practice data. No installation needed.

INSPECT THE EVIDENCE

Your cases appear here

No setup before your first result.

Select Generate practice data. Then run the baseline and open a row to see the response and the check that caught it.

Five families: refund policy, human escalation, privacy, document retrieval and safe tool sequence.

02 / OPTIONAL DEVELOPER PATH

Take the same lesson
to your own machine.

Reference runtime: Python 3.12.14. Standard library only. No pip install, account or API key.

  1. Download and unzip the kit.

    The ZIP contains source scripts, a dataset, tests, an example CI workflow, a license and a troubleshooting README. The book manuscript is not included.

    Download companion ZIPInspect file hashes
  2. Open a terminal in the extracted folder.

    Windows: type powershell in File Explorer’s address bar. macOS: open Terminal, type cd , drag the folder into the window and press Return. Linux: use Open in Terminal from your file manager.

  3. Check Python, then run one line at a time.

    Windows: py -3.12 --version. macOS/Linux: python3 --version. If missing, stay in the browser lab or use the official Python installer.

These commands match the generated browser dataset: 40 cases, seed 42. If you have not generated data, they use the book walkthrough defaults.

python3 scripts/generate_data.py --output data/dataset.jsonl --count 40 --seed 42
python3 scripts/run_evals.py --dataset data/dataset.jsonl --variant baseline --report reports/baseline.json
python3 scripts/run_evals.py --dataset data/dataset.jsonl --variant improved --report reports/improved.json
python3 scripts/compare_reports.py reports/baseline.json reports/improved.json --output reports/comparison.json
python3 -m unittest discover -s tests -v

Expected: baseline fails 40/40 with exit 1. Improved passes 40/40 with exit 0. Comparison has no new failures. Do not join these lines with &&: the expected baseline failure stops that chain.

Read the full setup and recovery guide →

A website cannot automatically open your local terminal. The browser terminal above runs here; these commands run locally after you open your terminal.

03 / READ → PRACTICE → REFLECT

Keep the book
and the lab together.

You do not need to understand every line of Python to explain why a case passes or fails.

CHAPTER 1–3

Define the risk

Write the reader’s goal and one unacceptable outcome. The demo is fictional, not a customer deployment.

CHAPTER 4

Inspect the data

Generate cases; open a row. Explain its input and expectation before running a check.

CHAPTER 5

Run deterministic checks

Run baseline, read each failure, then compare improved on exactly the same cases.

CHAPTER 6 + 13

Understand human judgment

Use metrics_demo.py for synthetic label calculations. This lab does not call or calibrate a live model judge.

CHAPTER 7–8

Inspect retrieval and actions

Find the obsolete document and wrong-order lookup failures. Understand the narrow scope of each check.

CHAPTER 9–14

Build a responsible release process

Introduce a regression, download a report and inspect the CI example. Add real permitted evidence and controls before deployment.

04 / GET UNSTUCK

A failure is useful
when the next step is clear.

The baseline is all red. Did I break something?

No. It deliberately uses wrong scripted responses. Open a case and read the reason. Run improved next. Do not weaken the rule to hide the failure.

The improved version is green. Is my AI ready to ship?

No. These are synthetic scripted teaching responses. Your real model, customer policies, privacy, authorization, monitoring and release decisions need their own evidence.

Python or a file is not found.

Use the browser route with no installation, or install official Python and reopen a terminal in the extracted kit folder. The folder must contain README.md and scripts/. Windows uses py -3.12; macOS/Linux use python3.

Can I use real customer data or an API key here?

No. This lab generates only fictional cases and has no key or customer-data upload field. Redact and obtain permission before building a separate private evaluation workflow.

Can I resume later?

Download the dataset and reports before leaving. This demo does not save to an account or local storage. Reloading starts fresh. Keep the same count and seed to regenerate the same cases.

What is public, and what is licensed?

The companion code and synthetic fixtures are public under MIT. The book manuscript, private research, admin records and production systems are separate and are not bundled.

Found an issue? Email support@qentra.tech with the kit version, case ID and fictional reproduction. Never include secrets or customer data.