Lesson 04 · Jev 101

Confidence: the difference between a demo and automation

A typed decision is still a guess. Production systems need a control signal: when to act, when to ask a human, when to call a bigger model.

Why prompted “I’m 90% sure” is a weak control

TypeSafe argues RLHF-trained chat models tend to beoverconfident and inconsistent even when you ask them for a confidence field. If a model is right 95% of the time but won’t tell you when it’s in the 5%, you can’t safely automate. TheWhat’s the difference? spells out that contrast.

“Yet RLHF creates inherent issues such as mode dropping, overconfidence, and lack of reliability. These flaws mean that LLMs require humans-in-the-loop.”

TypeSafe AI homepage · fetched Sep 20, 2026 · Source

What Jev gives you (company claim)

Every Choice / Score decision ships with probability and confidence estimates trained under RLCD for calibration: higher confidence should mean higher accuracy. Official nuance: calibration is a property across many predictions, not a guarantee on one answer.

Noul is different: official API concepts say Noul has no separate confidence field. The probability is the uncertainty signal. Design thresholds on P(yes) directly.

Your job: validate calibration on your labeled data, then set thresholds.

“if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it. But if it's 95%, sure, then I can do something with it.”

Secondary Armin Ronacher via TechCrunch · Sep 18, 2026 · Source

Product design habit

  1. Pick the decision (Choice / Score / Noul)
  2. Choose thresholds (e.g. auto if confidence ≥ X and P(yes) ≥ Y)
  3. Middle band → human or LLM escalate
  4. Log outcomes; recalibrate
  5. Pin the model version you tuned against (jev-1.13.0, not a moving alias)
BandWhat your code should do
High confidence, allowed actionAuto-act. Log it.
Middle / ambiguousEscalate to a human, or to a larger generative model if you need language.
Low confidenceDo not act. Official guidance: escalate. Don’t paper over it with another chat prompt.
High stakes (money, safety, legal)Human confirm even when the model is sure. Habit pair #34 exists for a reason.

ChatGPT bridge line: stop asking the chat model tobe the policy. Use a decision model for the policy signal; keep chat for the human-facing words.

What confidence does not fix

  • Wrong labels inside a valid schema
  • Math, dates, and counts (do those in code)
  • Adversarial or injected state
  • CJK accuracy gaps — watch the numbers extra carefully