[yodablocks]

What

A privacy follow-up to #28 and #30. Jev is a cloud API, so quarq's local-handling promise does not hold for that arm. The production re-ranker bge-reranker-v2-m3 is already local. This experiment asks whether the union candidate pool from #30 lifts bge the way it lifts Jev, which decides whether a fully local path can get close.

Everything is under experiments/local_rerank/. Nothing in quarq/ changes. The bands were committed before the run.

Result (one run, 3 s pause between questions)

Set	Re-ranker + pool	Hit@1	Hit@5	MRR	Query p50
38 gold	bge, embedding pool (#28)	33	36	0.901	2,079 ms
38 gold	bge, union pool	33	38	0.926	2,099 ms
38 gold	Jev, union pool (#30)	38	38	1.000	862 ms
15 hard	bge, union pool	10	12	0.733	2,049 ms
15 hard	Jev, union pool 10 (#30 runs)	12	n/a	n/a	about 850 ms

• Band: no lift at Hit@1 (33, the lift threshold was 35).
• The union puts the gold page in bge's pool (Hit@5 36 to 38), but bge does not rank it first. Its five Hit@1 misses on the 38 are the same as before, including b1-020 and b2-011, the two the union added.
• On the hard set bge misses h-001 and h-002 at rank 1; Jev gets both. h-009, h-011 and h-015 are in no pool at 10, so no re-ranker could fix those.

Caveats

• One run on small sets that have been iterated on. The hard questions are LLM-drafted (synthetic-draft) and not human-reviewed.
• Latency compares a local GPU with a network call, so it is only a rough figure. It excludes the 3 s pauses.
• The Jev hard-set figure combines runs from feat(experiments): bring the recall experiment (#29) onto master #30 (h-001..h-011 and the French rewrites h-012..h-014). h-015 was not run at pool 10 because its gold page is not in the pool, so Jev's 12 of 15 is not a single run.
• Hypothesis, not tested: the Jev question asks whether the passage states the answer, while bge scores topical relevance, so the wording may matter as much as the model.
• Not covered: other local re-rankers (needs a model download), and a local LLM as a pointwise judge (needs LM Studio running).

Reproduce

python3 experiments/local_rerank/run.py   # loads bge once, about 4 minutes, paced

Query embeddings and the BM25 index come from the on-disk caches built by experiments/recall/probe.py.