Inspect the scores.
Scores are supplied, not calibrated probabilities.
A boundary turns a score into a decision.
INVESTIGATIONS / MEASUREMENT & FRAMING
Constructed exampleWho pays when we move the line?
The same messages. A different cutoff. A different set of mistakes.
YOUR TASK / CONSTRUCTED EXAMPLE
Review the inbox rule
You’re reviewing a fictional inbox filter before it is used. Choose which mistakes deserve attention, then inspect the messages on each side of the cutoff.
An evidence trail.
Your decision.
Begin with a provisional judgment
Loading the supplied example…
Keep a private first thought (optional)
Private writing stays out of links and default downloads. A reopened link does not restore it.
This records a provisional judgment and the starting configuration. It does not change a mathematical setting or save automatically.
01 / THE QUESTION
What would you rather get wrong?
Twenty fictional messages have supplied labels and scores. A cutoff turns each score into a decision: flag it as spam, or let it through.
Keep a first thought
Optional. Included only when you explicitly save; excluded from shared links.
Constructed data · every relationship and outcome is supplied.
02 / OPEN THE MODEL
Assign costs to the mistakes
Whole fictional points 0–100. Changing a cost does not change any classification or move the cutoff.
—False positivesmessages
—False negativesmessages
—Assigned error costfictional points
A cutoff is a choice with consequences.
A false positive flags a legitimate message. A false negative lets spam through. Moving the cutoff changes which mistakes this fixed set contains.
Save this experiment
Explore the model and compare working versions
Follow the idea.
A concept, in three framesMove the cutoff.
The decision changes at an explicit boundary.
Count both errors.
False positives and false negatives differ.
HOLD ONE VERSION STILL
What changes when you change the model?
Keep the current inputs beside your next adjustment.
MODEL & LIMITS
What this model leaves open
These scores are not calibrated probabilities, and twenty supplied messages do not estimate real-world performance. Assigned costs are fictional points, not objective harm.
- InputsWhat is supplied?
- DefinitionWhat is counted?
- InferenceWhat follows?
Inspect the exact workings
| ID | Score | Supplied label | Decision | Outcome |
|---|---|---|---|---|
| S01 | 90 | Spam | Flag | True positive |
| S02 | 85 | Spam | Flag | True positive |
| S03 | 80 | Spam | Flag | True positive |
| S04 | 75 | Spam | Flag | True positive |
| S05 | 70 | Spam | Flag | True positive |
| S06 | 65 | Spam | Flag | True positive |
| S07 | 60 | Spam | Flag | True positive |
| S08 | 55 | Spam | Pass | False negative |
| L01 | 95 | Legitimate | Flag | False positive |
| L02 | 68 | Legitimate | Flag | False positive |
| L03 | 58 | Legitimate | Pass | True negative |
| L04 | 48 | Legitimate | Pass | True negative |
| L05 | 45 | Legitimate | Pass | True negative |
| L06 | 40 | Legitimate | Pass | True negative |
| L07 | 35 | Legitimate | Pass | True negative |
| L08 | 30 | Legitimate | Pass | True negative |
| L09 | 25 | Legitimate | Pass | True negative |
| L10 | 20 | Legitimate | Pass | True negative |
| L11 | 15 | Legitimate | Pass | True negative |
| L12 | 10 | Legitimate | Pass | True negative |
Precision: 7/9 = 77.78%. Recall: 7/8 = 87.50%. Assigned cost: 2 × 4 + 1 × 1 = 9 fictional points.
Assumptions, sources and evidence status
Original constructed fixture: constructed-messages-v1. thresholds-v1. No empirical observations are represented by these example counts.
- All labels and scores are supplied; these are not empirical classifier performance estimates.
- Scores are not calibrated probabilities. Error costs are fictional points chosen by the reader.
- Changing costs does not move the threshold or retrain a model. Precision is undefined if no messages are flagged.