Several queries take turns rather than pooling their scores

Context¶

--also shipped fusing every query's backend lists into one RRF ranking. adr-b23dae55666f then measured what a wrong phrasing costs — recall@5 0.88 → 0.25 — and established that weighting the task cannot fix it, because weighting suppresses an extra query everywhere, including the documents only it can find. That decision closed with one untested idea: bound a query's share of the result rather than its score.

Decision¶

With more than one query, the per-query rankings are interleaved rather than pooled. Each query's best goes first, then each query's second, and so on, deduped. The caller's task holds every Nth slot no matter what the others rank.

Measured against the same eight tasks and the same deliberately wrong hypothetical:

configuration	pooled RRF	task weight ×2	interleaved
task only	0.88	0.88	0.88
task + correct	1.00	0.88	1.00
task + wrong	0.25	0.75	0.75
task + correct + wrong	0.69	0.81	1.00

It dominates both on every row. It keeps everything pooling had on a correct phrasing, matches what weighting bought on a wrong one, and is the only one of the three that survives a caller sending both — where pooling drowns and weighting merely dilutes.

Why the shape works where the weight did not¶

They fail differently on purpose. A weight is applied to a query's scores, so lowering it lowers every document that query found, including the ones nothing else found — which is exactly the contribution --also exists for. Interleaving leaves every query's own ordering intact and rations only how many slots it may fill, so a bad phrasing loses the argument about how much of the result it gets without losing its best pick.

Put another way: pooling asks "which document has the most support", which a confident wrong query wins by being confident. Taking turns asks "what does each query most want", which it cannot win more than once.

Consequences¶

One query is unaffected — fuse is what it always was, and docir bench scores identically before and after (0.88 / 0.20 / 0.63).

The guidance adr-b23dae55666f forced can soften but not disappear. 0.75 against a baseline of 0.88 is still a real cost for a wrong guess, so --also remains for a caller that could defend the answer it is guessing — it is no longer the 5:1 downside that made it close to unusable.

FusedScore.score stops being comparable across a multi-query result: the merged list is ordered by turn, not by score, so a document in slot 2 may carry a lower score than one in slot 3. similarity is unaffected and remains the number with absolute meaning.