018 min
Try one gate
Sort paper items with a threshold at 50. Turn over reference labels and identify errors.
Ask: “Is there a gap between all A and all B scores?”
Listen for: “No; they overlap.”
Brightlab
brightlab.Made for curious mindsbrightlab.Made for curious mindsBrightlab
Year 4 · Evaluation
When is asking a person better than choosing a bin?
Go to the investigation ↓
Investigate with Wattle the wombatA recycling gate policy with a three-outcome audit and review instructions.
Which mistakes does your recycling rule make?
Our sorter uses a synthetic material score. Scores for two material types overlap, so no single threshold separates every item correctly. A reject zone sends uncertain items to a person. This can reduce wrong-bin decisions while increasing review work. These values are invented measurements, not real recycling advice or calibrated probabilities.
Before this lesson: Count missed targets separately from false alarms. Useful earlier investigations: A rover that waits for evidence.
Our sorter uses a synthetic material score. Scores for two material types overlap, so no single threshold separates every item correctly. A reject zone sends uncertain items to a person. This can reduce wrong-bin decisions while increasing review work. These values are invented measurements, not real recycling advice or calibrated probabilities.
Count missed targets separately from false alarms. Useful earlier investigations: A rover that waits for evidence.
Print the fixed set of material scores and labels. Use two containers and a review tray. Explain that actual recycling rules depend on local facilities; the invented materials are called A and B.
ACARA V9 AI curriculum connection ↗ · Technologies ↗
A complete teaching sequence · 60 minutes
Sort paper items with a threshold at 50. Turn over reference labels and identify errors.
Ask: “Is there a gap between all A and all B scores?”
Listen for: “No; they overlap.”
Record how a wider review band might affect wrong bins and workload.
Ask: “Could fewer automatic mistakes require more human work?”
Listen for: “Yes, more items might be reviewed.”
Adjust threshold with reject width fixed, then width with threshold fixed. Record both counts for three settings.
Ask: “Which items moved from a wrong bin into review?”
Listen for: “Those near the boundary.”
Inspect a confidently wrong item far from the overlap.
Ask: “Will a narrow reject zone catch this one?”
Listen for: “No, it is far from the line.”
Choose a setting under a review capacity of four items per batch. Label unresolved errors and design a clear review message.
Ask: “Can you meet the review budget without hiding mistakes?”
Listen for: “We need to show both counts, even if some errors remain.”
Partners exchange settings and reproduce counts.
Ask: “What could make your threshold unsuitable elsewhere?”
Listen for: “Different materials or different score measurements.”
One threshold can remove all errors.
A labelled outlier far from the boundary is confidently mis-sorted and bypasses a narrow reject zone.
Meet a four-item review capacity while documenting remaining wrong-bin decisions and a usable review message.
Check that reviewed items are not quietly counted as correct automatic predictions. Ask learners to locate one error excluded from review.
| Criterion | Beginning | Secure | Extending |
|---|---|---|---|
| Error accounting | Combines review and correct outcomes | Separates mistakes, correct decisions and review | Explains which cases each measure hides |
| Design justification | Chooses a threshold by appearance | Uses counts and review capacity | Documents a residual outlier and a monitoring need |
Work with eight score cards and count outcomes physically before using the full batch.
Assign different costs to contaminating each bin and compare the preferred policy with the accuracy-only choice.
Synthetic A/B materials only. Do not handle waste or apply this threshold to real recycling decisions.
Shared device? Turn remembering off. A project file lets you continue on another device.
Your browser is the laboratory
Widen the reject zone with scores and threshold fixed. Predict the direction of wrong-bin and review counts.
A scores: 12,20,32,42,49,55,90. B scores: 43,52,62,71,83. Threshold starts at 50. A review band around it can defer decisions. There are 12 items and capacity to review four.
The experiment opens after your prediction.
There is no penalty for being surprised.
Apple Silicon · PyTorch MPS
Sweep classification thresholds over a large fixed synthetic batch. Change the review width and compare automatic-error and review-work curves.
Students extend the experiment in teams, documenting parameters, outputs and limitations.
The bundle contains lesson-specific working code, a configuration file, a reactive notebook, a deterministic CPU check and hardware setup instructions. Acceleration is reported only after a tensor operation and result read-back succeed.
Download Mac Studio investigation ↓unzip y4-threshold-mac-pathway.zip -d y4-threshold-mac cd y4-threshold-mac bash setup-mac.sh source .venv/bin/activate python experiment.py --device mps --output results marimo edit notebook.py
CPU and available-device execution status is recorded in the downloaded README and validation report. DGX Spark execution requires that hardware; static validation alone does not establish GPU compatibility or performance. The browser lesson remains fully available without this extension.
NVIDIA DGX Spark · PyTorch CUDA
Sweep classification thresholds over a large fixed synthetic batch. Change the review width and compare automatic-error and review-work curves.
Students extend the experiment in teams, documenting parameters, outputs and limitations.
The bundle contains lesson-specific working code, a configuration file, a reactive notebook, a deterministic CPU check and hardware setup instructions. Acceleration is reported only after a tensor operation and result read-back succeed.
Download DGX Spark investigation ↓unzip y4-threshold-dgx-pathway.zip -d y4-threshold-dgx cd y4-threshold-dgx bash run-dgx.sh # Open the localhost notebook URL printed by the container.
CPU and available-device execution status is recorded in the downloaded README and validation report. DGX Spark execution requires that hardware; static validation alone does not establish GPU compatibility or performance. The browser lesson remains fully available without this extension.
A recycling gate policy with a three-outcome audit and review instructions.
Download the editable handout →