Reputation

Reputation

A calibrated, recency-weighted Bayesian skill estimate per council member — bounded, and penalised for overconfidence.

Why not a simple accuracy percentage

A percentage rewards the member who only speaks when certain, and punishes the one who engages with hard questions. It also treats a judgement from two years ago as equal to one from last week.

Property	What it fixes
Bayesian	an estimate from three data points is not treated like one from three hundred
Recency-weighted	recent performance counts more; a member who improved is not held to old form
Calibrated	rewards proportionate confidence, not just being right
Bounded	no member can accumulate unchallengeable authority
Overconfidence-penalised	being certain and wrong costs more than being uncertain and wrong

Calibration over accuracy

A member who says "70% confident" and is right 70% of the time is well calibrated — and more useful than one who says "95% confident" and is right 70% of the time, even though their raw accuracy is identical.

The second is not merely wrong sometimes; they are wrong about how much to trust them, which corrupts every decision that weighs their input.

Bounded on purpose

No member's reputation can grow without limit. A council where one voice becomes unchallengeable is no longer a council — and the member most likely to reach that point is the one whose past success makes disagreeing feel unreasonable.

Fed by graded outcomes only

Reputation moves when Outcome Tracking grades a prediction against reality. It does not move because a member argued persuasively, wrote at length, or was agreed with at the time.

curl http://127.0.0.1:8766/api/council/reputation

→ Next: Outcome Tracking