QIW← Question-Intelligence-WorkbenchLog in

The mathematics, hands on

Two interactive demos of the rating models behind QIW ranking

QIW does more than answer questions — it puts candidates into a defensible order. Two Bayesian rating models do that work: instead of a score, each candidate carries a distribution — an estimate and how unsure we are about it. That is hard to explain and easy to show. These two demos show it: verdicts arrive year by year, the distributions move and narrow, and the ordering settles — or does not.

Watch the 20-minute story“From Elo to Evidence” — the film behind these demos (20 min).
Demos, not the running system

Both pages run on invented data. Publication counts, dossiers, the “true” quality and the judgements are all simulated; the drug targets are there as a recognisable example, and what the pages say about them is not a literature review. Only the mathematics is the same: those equations run in QIW over real dossiers assembled from sourced evidence and over verdicts from real LLM judges.

Pairwise: Bayesian Bradley–Terry

The judge sees two dossiers and names the stronger one. Every verdict moves both distributions — an uncertain candidate a long way, a well-evidenced one barely at all. It also shows why a duel between two apparent equals is the most informative call you can buy.

Where this sits in QIW
In QIW: criteria in pairwise form, ranking through the “shortlist” tier.
Open the demo
Listwise: Plackett–Luce

The judge gets K dossiers at once and orders them. One call therefore buys K(K−1)/2 pairwise constraints — far cheaper, but each dossier is touched less often. The K slider makes that trade immediately visible.

Where this sits in QIW
In QIW: criteria in listwise form, ranking through the “funnel” tier (OpenSkill).
Open the demo

Both demos come in English, German, French and Spanish — switcher at the top right.

Further reading

The equations follow Weng & Lin (2011), extended with a precision that grows with the evidence available. The base model is checked numerically against openskill.py. The formulas are on the demo pages themselves, with the citation.