Audit TinyML score thresholds

Steps

Screens captured against Python 3 in your browser (JupyterLite) Python 3.12.7 / Pyodide 0.27.6 in JupyterLite 0.6.4; Chromium 148.0.7778.96; synthetic classroom fixtures on 2026-10-08; the tool may have moved on — the text steps are the contract.

1. 1 Step 1

   Do

   Run the first notebook cell in the editor and inspect the synthetic scored-input table.

   You will see

   Synthetic scored examples: seed=73, positives=10, negatives=10 row=00 label=1 score=0.953 row=06 label=1 score=0.611 row=09 label=1 score=0.303 row=10 label=0 score=0.828 row=12 label=0 score=0.516 row=19 label=0 score=0.061 STEP 1 synthetic inputs fixed

   Why it matters

   The labelled seeded set makes each later count reproducible; it is not a sensor trace.

2. 2 Step 2

   Do

   Run the second cell in the editor and inspect the confusion-matrix readout at threshold 0.50.

   You will see

   Decision rule: alert when score >= 0.50 actual/predicted alert quiet positive 8 2 negative 3 7 total=20; false alarms=3; missed positives=2 STEP 2 confusion matrix computed

   Why it matters

   A confusion matrix exposes false alarms and missed positives that a single accuracy number hides.

3. 3 Step 3

   Do

   Run the third cell in the editor and inspect the threshold-sweep table for precision, recall and false alarms.

   You will see

   threshold TP FP FN TN precision recall 0.30 10 5 0 5 0.667 1.000 0.40 9 4 1 6 0.692 0.900 0.50 8 3 2 7 0.727 0.800 0.60 7 2 3 8 0.778 0.700 0.70 5 1 5 9 0.833 0.500 0.80 3 1 7 9 0.750 0.300 0.90 1 0 9 10 1.000 0.100 STEP 3 threshold sweep computed

   Why it matters

   Moving the threshold changes both alert yield and false-alarm cost on the same examples.

4. 4 Step 4

   Do

   Run the fourth cell in the editor and inspect the precision and recall plot in the output panel.

   You will see

   Precision and recall plot (each # is 0.1; round to nearest block) 0.30 P ####### 0.667 R ########## 1.000 0.40 P ####### 0.692 R ######### 0.900 0.50 P ####### 0.727 R ######## 0.800 0.60 P ######## 0.778 R ####### 0.700 0.70 P ######## 0.833 R ##### 0.500 0.90 P ########## 1.000 R # 0.100 STEP 4 precision recall plotted

   Why it matters

   The paired plot makes the precision and recall tradeoff visible without smoothing the small sample.

5. 5 Step 5

   Do

   Run the final cell in the editor and inspect the budget decision and invariant checks in the output panel.

   You will see

   Budget: <= 2 false alarms in 10 synthetic negatives Chosen threshold=0.60; TP=7 FP=2 FN=3 TN=8 Precision=0.778; recall=0.700 Selection: maximize recall; break ties by precision, then lower threshold Matrix conservation and false-alarm budget: PASS This synthetic result does not establish deployed sensor or model performance. STEP 5 budget decision validated

   Why it matters

   The stated budget constrains the choice; the conservation check catches counting errors.