[yodablocks]

Fixes from a two-axis code review (standards and spec) of PRs #28 to #36. One code fix and documentation corrections. The standards findings about code style in experiments/ (type hints, docstrings, magic numbers, model names) are not fixed here: they turn on whether experiments/ should be exempt from the repo rules, which is the owner's decision.

Code

• experiments/jev_rerank/run.py: one Jev client was shared across arms, so a later arm's fresh_calls, cost_usd and latency percentiles included the earlier arms. Each Jev arm now gets its own client. The affected results file (run-20261004T033321.json, the jev:v3 fields) is left as recorded and annotated in the experiment README; the per-wording figures quoted there were already right.

Docs

• Model-name rule stated backwards. experiments/jev_rerank/README.md said the rule "bans model names inside rag/ and llm/". CLAUDE.md says "No hardcoded model names outside config.py and llm/ backends", so jev-latest and google/gemma-4-12b-qat in experiments/ break it. The README now says so plainly.
• Jev's variation attached to the right number. The top-level table put the one-question run-to-run variation next to 36/38, an embedding-pool single run; the repeat was on the union pool. Reworded.
• "About 5 questions behind Jev" is now "4 to 5", because Jev scored 38, 38 and 37 in three passes. Also in the local re-ranker README, where the 862 ms Jev latency is now labelled as measured on the embedding pool.
• Hard-set caveats carried forward. Only h-003 to h-011 were checked against the PDFs; the French rewrites only through their originals; and h-001 and h-002 are the two gold questions that motivated the union merge, so the hard set is not independent of it.
• Open items: the source finding now says "none of the sources checked (ECB, Banque de France, AMF)" instead of a blanket claim, and a sixth item records that the planned ingest-time tagging experiment was never run (and that "experiment 2" and "3" were reused as names).
• Stale line: the local re-ranker README said LM Studio was "not on this host"; it is the same Mac.
• The repeat-run call count (1,060) is now said to come from the run's printed output, since repeat.json does not store it.

No experiment was re-run. pytest is unaffected: nothing under quarq/ or tests/ changed.