Update · Stronger notebook checks in four lessons, and all 103 lessons now independently reviewed

An update to Synapsa Commons: free, hands-on AI courses that run anywhere. No new lessons this time. The four lessons of Drop 14 have been through confirmation reviews, and this release ships what those reviews fixed. With them, all 103 lessons are now counted as independently reviewed.

What changed

• The notebook checks now test what the grader tests, in all four Drop 14 lessons. Before, some wrong answers turned a row green on the progress board and then lost marks in the full grading. Each rule the grader applies now has a matching check in the notebook, so a wrong answer fails its own row, with a message that names the likely mistake. Two rare edge cases in Threat-model an agent, each costing 2 of its 70 points, still have no notebook check and are recorded for a later update. Where a rule was only implied, the exercise text now states it (for example, that a value must be a plain Python float and not a numpy float, or that a text goes into a prompt exactly as given).
• Correct answers now score full marks. In the full grading, some stand-ins for the lessons' own objects behaved differently from the real ones, so some correct answers lost points. They now match the lessons' objects.
• An environment for a tool-using agent (Post-training). The docstring example of the training function printed a table shape that did not match its table. It is now correct. The OpenEnv source links point to the project's new home at github.com/openenv, and the lesson now names OpenEnv as the environment hub it follows.
• Designing tools an agent can use (Agents and protocols). The text now says exactly which calls each toolset checks and what a failed call says. A new printed line counts where the silent failures ended. An example in a docstring that was never run is now a runnable example.
• Five workflow patterns in plain Python, each with an eval (AI engineering). The hard-item demo no longer relies on a fixed item name: the code picks the first hard item, prints that it is hard, and counts how often its decoy comes up over 400 samples. The three prompt-chaining rows now use the same draws, so they can be compared directly. The voting section now explains that a majority amplifies whichever label is most likely on one draw, and the conclusion about the orchestrator now follows from the rates it prints.
• Threat-model an agent (Evaluation and governance). The full grading has one more row (68 points became 70), and the notebook checks the same case. The exercise texts now state how flows with several tags, unknown components and taint are handled. The OWASP source quote is now the full sentence, re-read on 8 November 2026.
• The front page is regenerated: 103 lessons, every one passing all 14 gates in QUALITY.md and independently reviewed.

Checked before release

• Today all 103 lessons passed every quality gate on macOS. Every worked solution scores 100%. Every blank notebook scores 0%, except five C and C++ lessons, which score 5 to 9% because their "it compiles" rows pass.
• The four updated lessons run to the end unanswered, and print the same results on Python 3.11 and 3.12 when completed. All 17 links in their "Before you start" boxes are live.
• These four lessons were last checked on Linux on 6 November, before these fixes. That check was not repeated, because nothing in the fixes depends on the operating system.
• Limits still apply. On native Windows there is no time or memory limit, so use WSL, Colab or Codespaces there. One long native call, such as a very large numpy operation, is stopped only when it returns.
• Worked solutions and full grading come with enrolment on Synapsa, opening in the coming weeks.