Failure Modes
Where Jev breaks, and how to build around it
The failure modes that bite in production: numeric blindness, date comparison, context rot, adversarial state, and the structural invariants that simply do not hold.
What you’ll be able to do
- ✓Recognize which half of a problem belongs in code, not the model
- ✓Extract dates as enumerated Choices and compare them yourself
- ✓Stop assuming Noul and Choice probabilities are interchangeable
- ✓Test adversarial state before it reaches real users
Lessons
- 0118m
It Is Not a Calculator
Counting, arithmetic, and numeric representations. Fix three broken questions by moving math into code.
- 0220m
Dates, Times, and Ordered Quantities
Jev reads dates as text. Rebuild a date comparison as extraction plus code.
- 0318m
Context Rot and Large State
Accuracy falls as irrelevant detail grows. Filter first — including with Jev itself.
- 0425m
The Invariants That Don't Hold
Noul 0.72 plus its negation at 0.47 sums to 1.19. Why that is fine, and what it means for your thresholds.
- 0528m
Adversarial State
State is data, and Jev does not treat it as hostile. Build a test suite that tries to move the answer.
- 0640m
Project: Invoice Date Extractor
Every failure mode at once: dates, numbers, and large state. Extract with Jev, compute in code.
- 0740m
Project: Calibration Harness
Measure whether Jev's confidence means what you think on your own data, then set thresholds from evidence.
- 0835m
Project: Injection Canary
A CI job that replays adversarial states against a pinned version and fails when an answer moves.
Get every lesson in all three courses
Interactive exercises, mini-projects, the playground, and 1,000 live Jev credits.
50% off launch special$39.99 $19.99Ends September 30th
Already have access? Sign in.