eval/README.md

Evidence for the v11b Wikipedia index

Every number in WIKI_INDEX_V11.md, in REVIEW.md and in the dataset cards is traceable to a file here. Two builds are represented, and the distinction matters:

• v11b is what ships. Chunk counts: en 15,648,632 · ja 3,009,820 · zh 3,367,718.
• v11a was the first build. Three commissioned reviewers rejected it, it was never published, and it was rebuilt. Chunk counts: en 15,696,612 · ja 3,008,848 · zh 3,360,260. Logs measuring it carry _v11a in the name and are kept because REVIEW.md cites them as the "before" side.

Documents

file	what it is
WIKI_INDEX_V11.md	the release note: what was wrong, what changed, what was measured
WIKI_INDEX_FORENSICS_PAPER_v1_0_rc.md	the working paper, with the three review rounds reported at the same level of detail as the results
REVIEW.md	all three review rounds, their findings, and what was done about each

Acceptance — the shipped artifacts

file	what it shows
verify_en_final.log, verify_ja_final.log, verify_zh_final.log	the acceptance evidence for the shipped artifacts. verify_wiki_index.py: row count = FAISS ntotal = offsets length = manifest count, chunk_id unique, token distribution, two live queries. Exits non-zero on failure. All three PASSED
verify_en.log, verify_ja.log, verify_zh.log	superseded. These ran before the indexes were re-encoded with the residual-quantile range, so they describe files that were then overwritten. Round 3 of the review caught this; see REVIEW.md
build_en.log, build_ja.log, build_zh.log	the shipped build runs, including end-to-end throughput
offsets_{en,ja,zh}.log	the offset table and gzip seek index builds, with their file sizes
reindex_sq4.log, reindex_clip.log, reindex_clip_en.log	the two re-encodings of all three indexes from the retained fp16 vectors: FAISS min/max first, then the residual-quantile range that ships

Retrieval

file	what it shows
final_eval_en.log, final_eval_ja.log, final_eval_zh.log	the numbers in the cards: whole corpus at nprobe 32 and 128, plus, for ja and zh, the head-to-head against the previously published store on both question sets
final_eval_en_h2h.log	the English head-to-head, added 2026-09-18. The published English store's cleaning script no longer exists, so its control questions are generated from that store's own chunk text (build_en_control.py, en_control_build.log, questions_en_oldstore.json): 400 articles sampled by uniform stride from the 5,328,397 titles present in both stores
questions_*.json	the question sets themselves, including the _oldextract controls
eval_{en,ja,zh}.json	the title-query A/B of §6: the benchmark that ranked the defective extractor higher, with chunk counts and footer rates per arm
evalscale_{en,ja,zh}.json	the same three extractor arms scored with generated questions instead of the title, on ~7,000–8,800-chunk pools
evalqa_{en,ja,zh}.json	the small-pool question sets those two used
old_vs_new_v11a.log	the same head-to-head run against v11a. Superseded by final_eval_{ja,zh}.log
acceptance_full_corpus_v11a.log	whole-corpus retrieval on v11a. Superseded by final_eval_*.log

Design decisions, each measured before it was made

file	decision it settles
window_footer_uniform.log	how much of the published stores ran past the 512-token window, and how much of each store carried the licence footer
sweep_budget.log	the 256-token budget, against 160 / 360 / 480
infobox_ab_gpu1.log	keeping infobox rows in a separate chunk stream, and the arm where that loses
index_types.log, index_e2e.log	IVF,SQ4 over PQ64, PQ256, PCA+SQ8 and SQ8, by set fidelity and end to end
pq_loss.log	how much of the exact top-10 the previous IVF,PQ64 returned
quantizer_lab.log	percentile clipping against Lloyd-Max levels and a decorrelating rotation
faiss_rangestat.log, faiss_range_fix.log, faiss_range_residual.log	the quantiser range. The first attempt clipped raw-vector quantiles and made the index worse; the index encodes residuals. The three runs disagree about RS_optim, which REVIEW.md records rather than resolves
sizes.log	bytes per vector, measured on the index files rather than derived

After-the-fact verification

file	what it shows
fix_verification.log	each v11a defect counted on both stores: navigation-keyword deletion, abbreviation splits, mid-token starts, title leakage in the question set, chunks per gold article
resource_and_coverage.log	embedding throughput per GPU and precision, fp16-vs-fp32 equivalence, indexed-article coverage. §C (mean pairwise cosine) is superseded by remeasure_v2.log §B: it drew 2,000 pairs from the head of one embedding file, not 4,000 random pairs of the store
misc_measure.log	deep-read cost with and without the gzip seek index, the 60-character prose/facts pattern rate, the chunk_id collision arithmetic. §A (licence-string length) is superseded by remeasure_v2.log §A: it measured a string that occurs in no artifact
remeasure_v2.log	the corrected licence-footer length (184 characters, as it appears in the stores) and the corrected mean pairwise cosine over 20,000 disjoint random pairs of the whole shipped vector set
noise_floor.log	the standard error of one MRR figure on these question sets, reconstructed from the reported rates because per-question ranks were not retained
forensics.log	the first pass over the published artifacts, which started all of this

WIKI_INDEX_V11.md (the release note) and WIKI_INDEX_FORENSICS_PAPER_v1_0_rc.md (the working paper) also ship here, so the evidence and the documents that cite it travel together.

The .py files beside the logs are the scripts that produced them. In the GitHub repository the build pipeline itself — build_wiki_index_me5_bplus.py, build_offsets.py, reindex_sq4.py and the acceptance script verify_wiki_index.py — lives in scripts/, not here; copies ship beside the logs in each dataset repository.

Where two logs measure the same quantity, the later one is named as superseding the earlier in the tables above. Three such corrections exist, all found by round 3 of the review; REVIEW.md lists what each earlier measurement got wrong.

What this evidence does not establish

The question sets are synthetic, generated by one local model, with no human relevance judgements and no check that a question is answerable from its gold article; the reconstructed standard error of one MRR figure on them is 2.2–2.9 points. Every reviewer was a model instance commissioned by the author. Every index-type and quantiser number comes from 150 k–206 k-vector pools at three different partition settings, not from the shipped 3–15.6 M-vector indexes. Nothing here is peer reviewed, certified, or independently replicated.