brightlab01 / Optimisation
Learning to take a smaller step
The weight follows the gradient of a quadratic loss. Watch an actual update trace rather than a prerecorded animation.
Open notebookYears 9–12 · Live Python notebooks
Marimo labs
Change a variable. Follow the evidence. Open a lab below to start—no Python installation needed.
New bonus collection: AI Understanding — separate Year 6 and Year 12 research notebooks →
Year 9 Understand learning
02 / Unsupervised learning
The reef that regrouped itself
Group 96 fixed reef observations using weighted distances and repeated centroid updates. Each dot is a synthetic observation.
Open notebook03 / Uncertainty
Confidence meets its evidence
Compare predicted probabilities with observed outcomes across reliability bins. The diagonal represents perfect bin-level calibration.
Open notebookYear 10 Build with evidence
01 / Retrieval
Where should a document break?
Cut a short source into overlapping windows, score lexical matches, and inspect exactly what the top two chunks retain.
Open notebook02 / Retrieval
An assistant that can show its working
Test a structured answerability policy on 120 seeded cases. The chart sweeps retrieval thresholds while you change the reference policy.
Open notebook03 / Security
When the document gives orders
Trace four synthetic source texts through a phrase filter, an action proposal and an independent permission check.
Open notebookYear 11 Test the whole system
01 / Deployment
The system that fits the school
Simulate a FIFO queue with 24 requests. Change load and service time to see why fast hardware alone does not define a good deployment.
Open notebook02 / Security
Valid JSON, invalid action
Sweep numeric durations through a bounded schema and a separate owner check. No real action is executed.
Open notebook03 / Monitoring
The day the stream changed
Monitor one fixed noisy stream with a rolling average. A known change point makes false alerts and detection delay inspectable.
Open notebookYear 12 Design and defend
01 / Governance
Who gets a say before launch?
Enumerate review-time allocations for learner participation, accessibility and appeals. Inspect the limits of treating governance as a set of numeric constraints.
Open notebook02 / Evaluation
Which change earned the improvement?
Run all eight retrieval/filter/review combinations on the same synthetic suite. Inspect component interactions, not just the best total score.
Open notebook03 / Governance
Make the case. Show the evidence.
Follow dependencies between claims and evidence. Failed or expired tests propagate to every claim that needs them.
Open notebook