Confidence Calibration and Uncertainty Expression
Confidence Calibration and Uncertainty Expression
Priority: MEDIUM
Problem:
Claude Code expresses overconfidence, claiming success before end-to-end verification. Examples: "Excellent! Credentials generated successfully!" when builds still failed. This creates false expectations.
Current Behavior:
Premature success claims: "Fixed!" "Excellent!" "This will work!"
No expression of uncertainty
User believes task is done, but it isn't
Expected Behavior:
Express calibrated confidence: "I'm 70% confident this will work because..."
Distinguish between "completed a step" and "verified end-to-end"
Show uncertainty when guessing: "I think this might work, but I'm not certain because..."
Only claim success after verification
Impact:
User wastes time thinking problem is solved
False confidence damages trust
User can't assess risk of proposed approach
Suggested Implementation:
Require end-to-end verification before claiming success
Express confidence levels with reasoning
Distinguish "I completed X" from "X solved the problem"
Show what still needs verification
Example:
❌ Bad: "Excellent! Credentials generated successfully!"
✅ Good: "I've generated the certificate and profile (saved to ios/certs/). However, I haven't verified that EAS will accept them yet. Let me try a build to confirm these credentials work end-to-end."
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗