📚 2026-08-19 question-corpus audit

Superseded — 2026-08-19

The 679-question / 456-subject bench state measured here at 06:45 UTC is superseded by the 2026-08-19 corpus import and the 2026-08-19 legacy WXR import, which closed the same day at 9,069 questions and 483 subject terms. Preserved as observed.

One sentence

The live bench still holds 679 published questions with a matching 679-row index and 0 QuestionValidator violations; the repaired local-corpus pass finds 5,701 distinct local stems and 360 overlaps with the bench; one exact published duplicate and the English progressive subject pool remain the live defects.

Status vocabulary on this page: done · verified · open · blocked(<what>) · retracted.

What this page already proved before the 06:45 UTC bench pass

The existing page already proved the core corpus shape: 679 published questions, 679 index rows, 0 content-vs-index drift, 456 subject terms, 908 question-subject rows, language rows en=508 / de=57 / es=57 / fr=57, 0 validator violations over published questions, one exact normalised Mona Lisa duplicate, local union 1,425 fingerprints, and 324 bench/local overlaps. The 06:45 UTC pass below re-measured the bench half with a single mounted wp eval-file audit script and supersedes stale sufficiency text: the deployed gate now exposes only 2 specific subjects for default English progressive criteria (general-knowledge, geography) plus the synthetic any card, not 3 specific subjects. The local-corpus handoff then superseded the earlier local union/overlap numbers after fixing XLSX shared-string parsing and adding more corpus paths.

Bench command transcript

All bench measurements in the State, Taxonomy, Conformance, Duplicates, and Sufficiency sections came from this mounted read-only PHP audit command:

ssh loca@178.18.250.170 'cd ~/dev/wrdp && sudo -n docker compose --profile cli run --rm -v /tmp/qw_corpus_audit.php:/tmp/qw_corpus_audit.php wpcli wp eval-file /tmp/qw_corpus_audit.php'

Script source path on the bench: /tmp/qw_corpus_audit.php. The script reads wp_posts, wp_qwizz_question_index, wp_qwizz_question_subjects, term assignments, and question meta only; it does not update options, post meta, posts, or plugin files.

Headline cross-check command:

ssh loca@178.18.250.170 'cd ~/dev/wrdp && sudo -n docker compose --profile cli run --rm wpcli wp post list --post_type=qw_question --post_status=publish --format=count' → 679.

Code contracts cited by this page:

Contract	Source
Validator payload for indexed questions is built from qw_type, post content/title, qw_choices, qw_correct_idx, qw_explanation, qw_lang, and exactly one qw_difficulty term.	wp/wp-content/plugins/quizwizz/includes/QuestionIndex.php:223-233
QuestionValidator accepts only multiple/boolean, non-empty stem ≤2000 chars, 2-8 non-empty string choices, boolean exactly 2 choices, each choice ≤1000 chars, distinct normalised choices, in-range integer correct_idx, explanation ≤4000 chars, valid language slug, and difficulty `easy	normal
Progressive capacity is 2 * min(easy, harder) + (easy > harder ? 1 : 0), where harder = normal + hard; a progressive selection of count 10 needs easy >= 5 and normal+hard >= 5.	wp/wp-content/plugins/quizwizz/includes/QuestionSelector.php:160-177, wp/wp-content/plugins/quizwizz/includes/QuestionSelector.php:181-196
Builder cards are only advertised when enabled, pool ≥1, and SubjectTaxonomy::can_satisfy(...) returns true.	wp/wp-content/plugins/quizwizz/includes/SubjectTaxonomy.php:395-428, wp/wp-content/plugins/quizwizz/includes/SubjectTaxonomy.php:436-502

State

Re-measured 2026-08-19 at 06:45 UTC. Headline publish count equals the wp post list cross-check above: 679.

Metric	Value	Command / key
Published qw_question	679	count.qw_question.publish=679 from the mounted wp eval-file; cross-check wp post list --post_type=qw_question --post_status=publish --format=count → 679
Auto-draft qw_question	1 (ID 1351, empty title)	count.qw_question.auto-draft=1; local-corpus bench census dumped all qw_question posts and identified ID 1351
Draft / pending / private / trash questions	0 / 0 / 0 / 0	no grouped count.qw_question.* rows for those statuses in the mounted output
Published qw_quiz	13	count.qw_quiz.publish=13
Draft qw_quiz	10	count.qw_quiz.draft=10
Draft quiz IDs	598, 613, 694, 710, 938, 1556, 1657, 1758, 1859, 1960	quiz.draft_ids=598,613,694,710,938,1556,1657,1758,1859,1960
Index rows	679	index.rows=679 (SELECT COUNT(*) FROM wp_qwizz_question_index)
Valid index rows	679	index.valid_rows=679
wp_qwizz_question_subjects rows	908	subjects.rows=908
Published questions missing from index	0	drift.published_not_indexed=0
Index rows without published question	0	drift.indexed_not_published=0
Subject rows without indexed question	0	drift.subject_rows_orphan_question=0
Content-vs-index drift	0	drift.published_not_indexed=0 and drift.indexed_not_published=0

The five imported drafts named on now (1556, 1657, 1758, 1859, 1960) are still draft. Five older draft quizzes also remain: 598, 613, 694, 710, 938. The previous page note about a deleted workstream-E scratch quiz is historical evidence; the current grouped query reports no such draft ID.

Language

Rows by lang from SELECT lang, COUNT(*) FROM wp_qwizz_question_index GROUP BY lang in /tmp/qw_corpus_audit.php:

lang	rows
en	508
es	57
fr	57
de	57

Retracted

now.md previously said content languages were en only. That is retracted. The index still holds 171 non-English rows (57 × de/es/fr). Interface language is a separate setting and is not the content-language count.

Taxonomy

Difficulty

Index difficulty rows and taxonomy term assignments both agree in the 06:45 UTC pass:

difficulty	index rows	qw_difficulty term rows
easy	276	276
normal	240	240
hard	163	163
no qw_difficulty term on published questions	0	0
multiple qw_difficulty terms on one published question	0	0

Command keys: index.difficulty.easy=276, index.difficulty.normal=240, index.difficulty.hard=163, term.difficulty.easy=276, term.difficulty.normal=240, term.difficulty.hard=163, term.difficulty.none_count=0, term.difficulty.multi_count=0.

Difficulty is a qw_difficulty term for validator/index purposes (QuestionIndex.php:207-233), not qw_difficulty post meta. A validator payload that reads the empty meta key would report a false qw_invalid_difficulty on every row.

Subjects

qw_subject terms: 456 (subject.terms=456). Published-question pool size per qw_subject term, counting distinct indexed published questions via wp_qwizz_question_subjects:

pool	subjects	Command key
20+	3	subject.hist.20+=3
10-19	16	subject.hist.10-19=16
5-9	11	subject.hist.5-9=11
1-4	408	subject.hist.1-4=408
0	18	subject.hist.0=18

Top 20 by published count (subject.top20=...): geography 26, history 26, general-knowledge 22, science 16, language 13, society 13, animals 12, art 12, film-tv 12, food-drink 12, games 12, literature 12, math 12, entertainment-music 12, mythology 12, politics 12, sports 12, tech 12, science-nature 10, science-physics 9.

Zero-pool subjects: 18 (subject.zero_pool_count=18). First 15 by term order (subject.zero_pool_first15=...): animals-nature, entertainment-japanese-anime-manga, entertainment-board-games, entertainment-books, entertainment-cartoon-animations, celebrities, entertainment-comics, science-computers, entertainment, entertainment-film, games-technology, literature-arts, science-mathematics, people-society, qw-svg-probe-4lhqftjo.

Conformance

Conformance was re-run over all published and draft qw_question posts. The scan set is conformance.scanned=679, with conformance.published=679 and conformance.draft=0; the one qw_question auto-draft is intentionally outside the requested published+draft set.

Implementation notes: the audit script mirrored the indexed payload contract from QuestionIndex.php:223-233 and also called QuizWizz\QuestionValidator::validate_array(...) for each scanned question. The exact rule source is QuestionValidator.php:35-120.

rule	violations	sample ids	Command key
type not in `multiple	boolean`	0	—
stem empty or > 2000	0	—	absent rule.stem_empty_or_over_2000; conformance.rule_violations=0
choices not a 2-8 list / empty / non-string	0	—	absent rule.choices_not_list, rule.choice_not_string, rule.choices_outside_2_8_or_empty; conformance.rule_violations=0
boolean not exactly 2 choices	0	—	absent rule.boolean_not_exactly_2; conformance.rule_violations=0
any choice > 1000 chars	0	—	absent rule.choice_over_1000; conformance.rule_violations=0
choices not distinct after validator normalisation	0	—	absent rule.choices_not_distinct; conformance.rule_violations=0
correct_idx missing or out of range	0	—	absent rule.correct_idx_missing_or_out_of_range; conformance.rule_violations=0
explanation > 4000 chars	0	—	absent rule.explanation_over_4000; conformance.rule_violations=0
invalid lang slug	0	—	absent rule.lang_invalid; conformance.rule_violations=0
no / multiple difficulty terms	0	—	absent rule.no_difficulty_term, rule.multiple_difficulty_terms; conformance.rule_violations=0

0 violations. The direct validator call also returned 0 failures (conformance.validator_failures=0, conformance.validator_failure_detail=). Missing qw_choices/qw_type meta: 0 (conformance.missing_meta=0).

Choice-count histogram over the 679 scanned questions:

choices	questions	Command key
2	91	choice_hist.2=91
4	585	choice_hist.4=585
5	2	choice_hist.5=2
8	1	choice_hist.8=1

Empty explanation is allowed by the validator; the rule is only a 4000-character cap. Current empty-explanation count: 40 (conformance.empty_explanation=40).

Duplicates

Normalisation for duplicate detection: use the question stem from post content when non-empty, otherwise post title; strip tags; lowercase; replace punctuation/non-alphanumeric runs with one space; collapse whitespace; trim. The local-corpus handoff separately proved question text lives in post_title and post_content duplicates it for ID 67; meta keys are qw_type, qw_choices, qw_correct_idx, qw_explanation, qw_lang, qw_source, and qw_opentdb_id, none of which holds the stem.

kind	groups	posts	10 largest groups / evidence	Command key
Exact normalised stem among published questions	1	2	2x [357/publish 1915/publish] "who painted the mona lisa"	dup.groups_publish_only=1; dup.group1=2x [357/publish 1915/publish] "who painted the mona lisa"
Exact normalised stem across published + draft questions	1	2	same group; there are no draft qw_question posts	dup.groups_total=1, conformance.draft=0
Published-vs-draft collisions	0	0	none	dup.groups_publish_vs_draft=0
Unique bench fingerprints in published + draft scan	678	—	679 scanned posts minus the one 2-post group	dup.unique_fingerprints=678

The local-corpus bench census agrees: duplicate group ID 357 Who painted the 'Mona Lisa'? equals ID 1915 Who painted the Mona Lisa? after quote stripping; publish distinct count is 678.

Historical near-duplicate evidence retained

The earlier page also listed a near-duplicate pair, 1574 Which compound is chiefly responsible for the heat of chili peppers? and 1591 Which compound is chiefly responsible for the pungent heat of hot chili.... That was a heuristic first-40-normalised-character check, not part of the exact-normalised duplicate requirement re-measured in this pass, so it remains historical evidence and is not counted in the exact collision table above.

Sufficiency against the default English progressive request

Default criteria measured through the deployed code path: GenerationCriteria::from_array(array('lang'=>'en','difficulty'=>'prog','count'=>10), ExposureSettings::get()) produced suff.criteria=source=dynamic lang=en difficulty=prog type=mixed count=10; default language is en (suff.default_lang=en) and default count is 10 (suff.count_default=10).

The code’s own progressive rule is not 3/3/3. QuestionSelector::capacity() computes progressive capacity as 2 * min(easy, harder) + (easy > harder ? 1 : 0), where harder is normal + hard (QuestionSelector.php:160-177). QuestionSelector::select() then requires easy >= ceil(count / 2) and harder >= floor(count / 2) (QuestionSelector.php:181-196). For a 10-question progressive English quiz, the real requirement is therefore 5 easy English questions and 5 normal-or-hard English questions under the selected subject.

Builder advertisement also uses that same rule: SubjectTaxonomy::can_satisfy() clones the criteria, sets the subject categories, calls QuestionSelector::capacity($probe), and requires capacity >= count (SubjectTaxonomy.php:395-428); builder_cards() skips cards unless pool >= 1 and can_satisfy(...) is true (SubjectTaxonomy.php:436-502).

Metric	Value	Command key
Subject terms considered	456	suff.subjects_advertised=456
Specific subject terms satisfying default English progressive 10	2	suff.subjects_can_satisfy=2
Satisfying specific subjects	general-knowledge, geography	suff.can_satisfy_slugs=general-knowledge, geography
Non-empty but insufficient subject pools	197	suff.nonempty_but_insufficient=197
Zero-capacity subject pools for the progressive criteria	257	suff.zero_capacity=257
Synthetic any card satisfies the same criteria	yes	suff.any_can_satisfy=yes
Synthetic any progressive capacity	438	suff.any_capacity=438

Top progressive capacities under the code rule (suff.top12_prog_capacity=...): geography 14, general-knowledge 13, science-nature 7, history 6, language 4, society 4, science 3, animals 3, geography-maps 3, geography-world-capitals 3, food-drink 2, film-tv 2.

Superseded 2026-08-19 06:45 UTC

The earlier text on this page used two hand-derived thresholds (3/3/3 and 1/1/1) and said history also qualified under the relaxed bar. That is now superseded. The deployed code’s rule is easy-vs-harder progressive capacity; under that rule history’s capacity is only 6, so the real specific-subject count is 2, not 3. Including the synthetic any card, the builder can advertise 3 cards: any, general-knowledge, and geography.

Local corpus inventory

Method from local-corpus handoff: stdlib XLSX reader /tmp/corpus-work/xlsx.py using zipfile + ElementTree, with sharedStrings indexed and t="s" resolved as an index; fingerprint is sha1(lower → strip punctuation → collapse whitespace → strip); fingerprints live in /tmp/corpus-fingerprints/*.txt. Bench dump command used by that handoff: scp /tmp/corpus-work/dump.php loca@…:/tmp/dump.php then ssh loca@178.18.250.170 'cd ~/dev/wrdp && sudo -n docker compose --profile cli run --rm -v /tmp/dump.php:/tmp/dump.php wpcli wp eval-file /tmp/dump.php'; it emitted 680 qw_question rows.

dataset	sheet/node	question column	records	non-empty	unique norm	∩ bench pub	% of set
04 QUESTIONS/Qwizz_Gold_Question_Corpus_240.xlsx	Question Bank	Question	240	240	240	65	27.1%
04 QUESTIONS/Qwizz_Gold_Question_Corpus_360.xlsx	Question Bank	Question	360	360	360	70	19.4%
04 QUESTIONS/Qwizz_Gold_Question_Corpus_504.xlsx	Question Bank	Question	504	504	504	70	13.9%
04 QUESTIONS/Qwizz_Gold_Question_Corpus_1008.xlsx	Question Bank	Question	1008	1008	1008	101	10.0%
04 QUESTIONS/Qwizz_10k_Content_Runway.xlsx	Question Bank (+ Production Queue: 8992 placeholder rows, empty Question)	Question	10000	1008	1008	101	10.0%
04 QUESTIONS/Qwizz_10k_Content_Runway_dfq.xlsx	byte-identical to above	Question	10000	1008	1008	101	10.0%
04 QUESTIONS/Qwizz_Wave_01_Editorial_Candidates_168 (1).xlsx	Question Bank	Question	168	168	165	19	11.5%
04 QUESTIONS/Qwizz_Wave_01_Editorial_Candidates_168.xlsx	byte-identical to above	Question	168	168	165	19	11.5%
04 QUESTIONS/Qs.json	questions[]	stem	32	32	32	2	6.2%
02 DOCS/questions.json	questions[]	question	229	229	229	205	89.5%
04 QUESTIONS/World Quiz Datase files 4 qwizz/questions.json	identical copy	question	229	229	229	205	89.5%
04 QUESTIONS/quiz-lib_claude-jam_files/questions.json	identical copy	question	229	229	229	205	89.5%
04 QUESTIONS/loca-dev-trivia-atlas/data/physical-sciences/easy.jsonl	jsonl	q	26	26	24	0	0.0%
04 QUESTIONS/loca-dev-trivia-atlas/meta/rejections.jsonl	jsonl	packet.q	33	26	26	0	0.0%
04 QUESTIONS/world_quiz_reference_demo_v1/data/items.jsonl	jsonl	prompt.stem	33	33	33	1	3.0%
01 APP/quizwizz-1.3.1-external-audit/…/assets/data/opentdb-local.json	json array	question_en	4196	4196	4192	37	0.9%

Superseded local inventory

The earlier local section on this page reported a 1,425-fingerprint local union and 324 bench overlaps. The local-corpus handoff supersedes that after fixing the XLSX shared-string reader and including additional corpus files: total distinct local is 5,701, local ∩ bench is 360, and grand distinct union is 6,019.

Overlap with the bench

set	unique fingerprints	Method / proof
Bench published posts	679	local-corpus bench dump and this page’s wp post list cross-check agree
Bench published distinct	678	one exact duplicate group, ID 357 vs ID 1915
Total distinct local	5,701	local-corpus fingerprint union using sha1(lower → strip punctuation → collapse whitespace → strip)
Local ∩ bench	360	local-corpus cross-match against the 678 distinct bench published stems
Bench questions in no local dataset	318	678 distinct bench stems minus 360 shared
Grand distinct union	6,019	5,701 local + 678 bench - 360 shared

Notable structure from the local-corpus handoff:

• Gold 360 ⊂ 504 ⊂ 1008, and Gold 1008 is identical to the 10k Runway Question Bank.
• Gold 240 carries into Gold 360 at 239/240; exactly one 240-era question was replaced or reworded.
• Byte-identical file pairs/triples: 10k_Content_Runway.xlsx equals _dfq.xlsx; both Wave 01 files are identical; the three questions.json copies are identical.
• Sixteen discovered dataset files collapse to thirteen real corpora after byte-identical duplicates.
• Wave 01 shares only one item with the Gold line and has three internal duplicates.
• Gold is almost disjoint from imported material: Gold ∩ OpenTDB pack = 2; Gold ∩ 229-pack = 1.
• The trivia-atlas material has not been imported: 0 bench overlap for easy.jsonl; easy.jsonl has 24 unique stems and rejections.jsonl has 26, with 2 shared.
• Qs.json needed in-memory repair only: line 287 contains one illegal JSON escape, a bare \l in \lim; source was untouched.
• No workbook carries a language column; Qs.json declares meta.languages = [en,es,fr,de,zh-Hans,pt-BR] but every stem is English.

Deltas this programme

Every bench number below was re-measured on 2026-08-19 at 06:45 UTC where possible; source-history items that cannot be re-derived from current database state are marked [cited] with the page that records them.

change	then	now	proof
Import blocker (chess vs games-chess)	packs died on slug mismatch	[cited] slug resolution fixed and recorded complete	completed task log 2026-08-18 completed row
Corpus import	269 published [cited]	679 published, index 679, drift 0	mounted audit keys count.qw_question.publish=679, index.rows=679, drift.published_not_indexed=0, drift.indexed_not_published=0; headline cross-check wp post list --post_type=qw_question --post_status=publish --format=count → 679
Five English packs	—	drafts 1556 / 1657 / 1758 / 1859 / 1960 still draft; five older drafts also remain	quiz.draft_ids=598,613,694,710,938,1556,1657,1758,1859,1960
Sufficiency gate / Wave 6 #12	guest cards 11 [cited]	code-derived default English progressive: 2 specific subjects (general-knowledge, geography) plus synthetic any; 197 non-empty subject pools still insufficient	sufficiency census; mounted audit keys suff.subjects_can_satisfy=2, suff.can_satisfy_slugs=general-knowledge, geography, suff.any_can_satisfy=yes, suff.nonempty_but_insufficient=197
B001 player review restore	in-process review length 58 [cited]	browser-verified anonymous quiz 413 result 13 / 58 (22%), 58 .qwizz-player__review-item, 0 console errors [cited]	play and print answers; screenshot 03 EVIDENCE/2026-08-19-quiz-413-playthrough-review.webp
Assumed reveal enum `instant	end	never`	written as the player contract
Answers operator UI	one hidden/preserved bit	two visible controls on the quiz metabox and the builder; sheet bit persists independently; player-review disabled under Forced site policy [cited]	metabox/browser proof from answers-switch workstream
German catalog	153 missing / 76 obsolete	1650 = 1650, missing 0 / obsolete 0 [cited]	task 2.1.1
Probe harness	1..46 / 46 / 0	no probe run by this bench-corpus pass; latest known post-install probe remains [cited]	coordinator-owned probe receipts, not this page’s bench audit command
Debug log	38,433	not re-measured by this pass	previous page evidence retained; no current claim

Open recommendations

Each open, one line, no schedule.

id	status	recommendation
QC-01	open	Merge published pair 357 / 1915 (Mona Lisa) — same English easy item, punctuation-only difference.
QC-02	open	Decide whether chili pair 1574 / 1591 is a rewrite to keep or a near-duplicate to drop; it is not part of the exact-normalised collision count.
QC-03	open	Fill or explicitly mark the 40 published questions with an empty explanation.
QC-04	open	English progressive pool is the bottleneck, not raw count: only general-knowledge and geography satisfy the real code rule for a specific subject; any satisfies through the aggregate pool.
QC-05	open	Gold 1008 and the 10k runway sheet are the same 1008 non-empty rows — do not count them twice in the fusion ledger.
QC-06	open	Treat doctrine files separately from inventory; the repaired local pass did not count qwizz-irresist.yaml as a question corpus.
QC-07	open	Drop byte-identical duplicate corpus files from the working tree so future globs do not double-count them.
QC-08	open	qw-svg-probe-4lhqftjo is a leftover probe subject with a zero pool — delete or quarantine.
QC-09	open	Owner decision UX-ANS-01 (may Forced honour per-quiz sheet off?) is still open; today’s REST already honours per-quiz '0' for the sheet flag.