Cormack, Clarke & Buettcher (2009) — reciprocal rank fusion, and the unexplained 60

[akash-coded]

Reading Club. One paper per thread. The house position: replicate, do not cite.
A paper we can re-run against this corpus is worth far more argued about with numbers.

Paper: SIGIR 2009

The claim, in one paragraph

Two and a half pages that beat every learned fusion method they tested. RRF(d) = Σ 1/(k + rank), with k = 60, and no derivation for the 60.

The mechanism worth arguing about: RRF consumes ranks, not scores, because any monotone
transform of a retriever's scores leaves its ranking unchanged — so scores carry retriever-specific
information that is not comparable across retrievers.

How we would test it here

notebooks/04 §4.9 sweeps RRF against weighted fusion on this corpus, and finds RRF losing
— finding 1 in the README. L05 makes you implement
both and watch k decide between confidence and agreement.

Two or three things to argue about

1. Where does 60 come from? Fitted on TREC. On this corpus the whole k sweep is worth +0.023 against α's +0.061. Is k a parameter at all, or a constant with a plausible story?
2. RRF is a positional voting rule, closer to Borda count than to a scoring function. Does framing it as social choice — interpersonal comparison of utility — buy you anything predictive, or is it a nice analogy?
3. When does RRF lose? Equal votes assume comparable competence. On this corpus the dense leg is weak and RRF drags the merge toward it. How would you detect that condition before fitting an α?

---

If you have run any part of this, post the numbers — a replication with an interval outranks
every opinion in the thread, including mine.