Week 24 · Term Three · One inspectable audit
Turn one faulty sorter into counts a neighbor can check: 9 + 3 examples, 8 − 3 actually correct, 6 − 2 still waiting, 4 × 2 human checks, and a statement that does not say the machine is fair or that bias is gone.
Monday adds 9 + 3 examples. Tuesday subtracts 8 − 3 actually correct. Wednesday subtracts 6 − 2 still waiting. Thursday names 4 × 2 human checks. Friday writes a statement a skeptical family can inspect.
20 minutes
Statistics · Two rows, one training count
Count training examples as both rows, and refuse a whole-forest stamp the table never held.
Make
An example card
Audit handoff: The example card is filed. The sure dial is still open — twelve examples are not yet a correctness count a neighbor can subtract.
25 minutes
AI · Sure minus wrong
Count actually correct answers as sure minus wrong, and refuse a sure-means-right stamp that hides the five still on the card.
Make
A confidence card
Audit handoff: The confidence card is filed. The mistakes are still open — five correct answers are not yet unreviewed mistakes a visitor can count.
25 minutes
Ethics · Posted minus reviewed
Count waiting mistakes as posted minus reviewed, and refuse a fixed-itself stamp that hides the four still on the card.
Make
A mistake card
Audit handoff: The mistake card is filed. The review rule is still open — four waiting mistakes are not yet a rule a family can inspect.
20 minutes
Writing · Who must still look
Keep the audit as a counted human-check total, and refuse a headline that the machine can ship alone.
Make
A review card
Audit handoff: The review card is filed. The audit is still open — eight checks are not yet a statement a family can inspect.
25 minutes
Public writing · An audit families can check
Write a model audit that uses this week’s counts and states what the table still cannot prove.
Make
A model audit
Audit handoff: The model audit is filed. The year now has one inspectable machine-learning table — not a claim that the machine is fair.
The audit uses only work entered here. It does not infer a fair machine, a removed bias, a next batch, mastery, attendance, or accreditation.
