| # Full strict-prompt validation census |
|
|
| Ran on 2026-08-01 in response to a direct question: is the current validation split |
| (444 facts, post-rebalance) manually reviewed at all? It was not — the earlier |
| validation review (300/600-fact sample) was drawn from the pre-rebalance 1,464-fact |
| pool, and `rebalance_splits.py` moved 1,020 of those facts into test. Diffing |
| fact_ids confirmed **zero overlap** between that sample and the current 444-fact |
| validation set. This census closes that gap: all 444 validation facts, all 11 |
| non-English languages plus English-vs-gold, same strict prompt as the full test |
| census (`data_analysis/make_full_test_census.py --split validation`). |
| |
| 12 worksheets, 12 subagents, one per worksheet. All 12 verdict files passed |
| structural validation (row count, fact_id/order match against the worksheet, |
| no duplicates) before aggregation. |
|
|
| ## Translation defects (11 languages, exhaustive — not a sample) |
|
|
| | lang | n | ok | subject | type | relation | unsure | defect rate | |
| |---|---|---|---|---|---|---|---| |
| | ar | 444 | 416 | 10 | 16 | 1 | 1 | 6.1% | |
| | sw | 444 | 417 | 4 | 20 | 2 | 1 | 5.9% | |
| | ru | 444 | 419 | 13 | 6 | 3 | 3 | 5.0% | |
| | ja | 444 | 426 | 5 | 11 | 0 | 2 | 3.6% | |
| | zh | 444 | 423 | 8 | 7 | 1 | 5 | 3.6% | |
| | de | 444 | 430 | 12 | 2 | 0 | 0 | 3.2% | |
| | bn | 444 | 430 | 5 | 3 | 4 | 2 | 2.7% | |
| | es | 444 | 425 | 9 | 2 | 1 | 7 | 2.7% | |
| | pt | 444 | 434 | 3 | 5 | 2 | 0 | 2.3% | |
| | fr | 444 | 427 | 5 | 4 | 1 | 7 | 2.3% | |
| | id | 444 | 435 | 6 | 1 | 1 | 1 | 1.8% | |
| | **all** | **4,884** | **4,513** | **80** | **77** | **16** | **29** | **3.54%** (95% Wilson CI 3.06–4.10%) | |
|
|
| ## English-vs-gold |
|
|
| 431 ok / 9 conflict / 3 vague / 1 unsure — **conflict rate 2.03%** (95% Wilson CI |
| 1.07–3.81%). |
|
|
| ## Comparison to the test census |
|
|
| This is the first *controlled* split comparison this project has run — same strict |
| prompt on both splits, unlike the earlier validation-sample-vs-original-test-review |
| comparison, which confounded prompt strictness with split. |
|
|
| | | test (2,523 facts) | validation (444 facts) | |
| |---|---|---| |
| | translation defect rate | 3.85% [3.63, 4.08] | 3.54% [3.06, 4.10] | |
| | English-conflict rate | 4.08% [3.38, 4.93] | 2.03% [1.07, 3.81] | |
|
|
| Both CIs overlap. Validation is not measurably dirtier or cleaner than test under |
| the same rubric — the two splits appear to carry the same underlying defect rate. |
|
|
| ## Known bug recurrence |
|
|
| - **Arabic "municipality of citizenship" (بلدية) — confirmed NOT recurring.** The |
| reviewing agent explicitly checked all 43 `country of citizenship` rows in this |
| batch; all correctly use بلد. The prior fix holds under fresh, independent review. |
| - **"Place/work read as a language" — recurs, and is broader than previously |
| documented.** Confirmed again in sw/ja/zh (the known pattern: a name with no |
| target-language Wikidata label gets rendered as "the X language," then asked |
| its own official language). New in this batch: Swahili shows the same reversed |
| construction on **P407 (language of work/name)**, not just P37 (official |
| language) — 14 of Swahili's 20 `type` flags are this pattern across both |
| relations. Previously only documented as a P37 bug. |
| - **Truncated Spanish questions — recurs.** 2 more instances in this batch |
| (rows 50, 245 of `census_validation_es_00.md`), consistent with the class found |
| in the test census. |
|
|
| ## New finding: Korean-script (Hangul) contamination |
|
|
| Not caught by any prior review pass because it's orthogonal to the subject/type/ |
| relation rubric — the subject and property are usually still correct, only a |
| Korean-origin proper name or work title is left in raw Hangul instead of being |
| transliterated into the target script (e.g. `Q495870`'s `ja`/`zh` questions ask |
| about "오윤교" instead of a transliterated form). Two reviewing agents independently |
| flagged the same fact_id as `unsure` for exactly this reason, which prompted a |
| full-dataset mechanical scan (Hangul is not a legitimate character range in any of |
| this dataset's 12 languages, so any hit is unambiguous — zero false-positive risk). |
| |
| **49 (fact, language) hits across the whole 59,291-fact release**: 47 in train |
| (~40 distinct facts, several hit in 2+ languages), 2 in validation (1 fact, |
| `Q495870`, both `ja` and `zh`), 2 in test (`Q7640060` in `fr`+`zh`, `Q2743574` in |
| `bn`). All traced to Korean-origin subjects/titles (e.g. `오윤교`, `권정혁`, |
| `일성록`/Ilseongnok, `트롤 슬레이어`/Troll Slayer) whose name/title failed to |
| transliterate. |
| |
| Scope is small (49/711,492 language-rows ≈ 0.007%) and the fix pattern already |
| exists (`make_retranslation_files.py` → per-language subagent → gated |
| `apply_full_census.py`-style acceptance). Given the 3-day submission deadline, |
| recommend: fix the 2 test-split and 1 validation-split facts now (cheap, and |
| these are the reviewer-visible splits), document train's ~40 facts as a known |
| residual defect class in the dataset card rather than fixing before the deadline. |
| |
| ## Recommendation |
| |
| - Validation census is now complete and at parity with test's rigor. |
| - Fix-vs-flag decision for the 173 translation defects + 9 English conflicts |
| follows the same policy already established for test (retranslate confirmed |
| question defects via the gated pipeline; flag English conflicts for editorial |
| review; no automatic drops). Not yet applied — pending a go-ahead, given |
| competing deadline priorities. |
| - The Hangul-contamination fix (2 test + 1 validation fact) is small enough to |
| bundle into that same fix pass. |
| |
