Corrections, errata and right of reply
Every correction to a published record is logged here with the old value, the new value, and the source that settled it. This page is trust infrastructure: the registry expects to be wrong sometimes and corrects itself in public, newest first.
Corrections
C-0009 · every record
Rubric 1.5 to 1.6 before first publication, reconciling the fifth review pass of v0.8.0, scoped to the v0.7.0 to v0.8.0 diff. The reviewer confirmed the diff as declared (six fields on three cells), the gate and rubric reconciled in both directions on all nine properties, and a section 10 row for every check the gate emits; the verdict was fit to publish with one grade to move, and every finding is taken. (1) Grade move: copsoq-iii.populations_languages_norms: High -> Moderate (the 1.5 norm rule was met on one entry whose matched word names a shared-variance figure, not a norm; Rahimi 2025 carries norms in substance and states no value; the Moderate band names the case; the High block and its four entries are kept in previous; returns to High on the Rahimi full text). (2) Status move: eurofound-ewcs internal consistency, well-established to thin, grade unchanged at Low; section 4 defines well-established as a mature, replicated evidence base, which a cell with no reliability coefficient of its own does not have. (3) The K10 criterion-validity judgement is corrected in one clause, grade unchanged at Moderate: the two developer-authored studies were described at 1.5 as from one clinical-reappraisal programme; the abstracts describe a two-stage clinical reappraisal within the National Health Interview Survey (Kessler 2002) and an enriched convenience sample of 155 (Kessler 2003), and the cell now says so. The phrase came from the fourth-pass reconciliation and the C-0008 text that repeats it stands as history. (4) One entry's statistic gains a population clause with its value unchanged (Morin 2011 on isi criterion validity against a reference standard: the change score is from the clinical sample, the accuracy figures from the community sample); the 1.5 wording is archived in previous. (5) The WHO-5 populations cell carries a rater note that its one norm-bearing entry is a screening cut-off validated in diabetes patients and that the cell holds on its working-adult and general-population entries; grade unchanged at High. (6) Rubric section 10 row 2 now says precondition evidence sits only on a High cell or a High previous block; the build already checked both and no code changed. (7) Every DOI in the dataset (420) resolved at doi.org on the conformance date. No fresh literature search was run: as_of is unchanged on every cell. 3 cells changed state and carry previous with their rubric 1.5 values.
Was: rubric 1.5; copsoq-iii populations High on a word match; eurofound-ewcs internal consistency well-established; k10 judgement naming one programme
Now: rubric 1.6; copsoq-iii populations Moderate pending the Rahimi 2025 full text; eurofound-ewcs internal consistency thin; k10 judgement naming the two designs
C-0008 · every record
Rubric 1.4 to 1.5 before first publication, reconciling the fourth review pass of v0.7.0 (findings F1 to F10, rulings P1 to P5). The reviewer's verdict on v0.7.0 was fit to publish; none of the findings was blocking and all ten are taken. (1) Populations, languages and norms: at least one of the two counted studies on a High cell carries a norm, cut-off, prevalence or explicitly reported reference value; the gate tests it; every current cell holds (who-5 1 of 16, copsoq-iii 1 of 4, k10 2 of 8, wemwbs 4 of 8, gad-7 2 of 2) and no grade moves. (2) Gate and rubric reconciled in both directions: the person separation index counts for internal consistency; the structural row of section 3 names the Mokken scalability coefficient, Loevinger's H and the AGFI, NFI and NNFI indices the gate already accepted; the words agreement, correspondence, predict, convergent, discriminant, divergent and responsiveness leave the statistic lists; a comma is a clause boundary; the criterion list is split so a correlation or regression coefficient counts only against organisational outcomes. One entry's statistic is reworded so its accuracy word precedes its value under the comma boundary (Morin 2011 on isi criterion validity against a reference standard; value unchanged); no other entry is affected, and no entry is added, removed or archived. Section 10 gains rows for the high_basis refusal and the empty test-retest check. (3) Grade moves: eurofound-ewcs.internal_consistency: Moderate -> Low (P3(a): the sub-indices are multi-item, so the property applies, but no coefficient of the record's own is present; the embedded WHO-5's alpha is graded on the who-5 record and is not this record's evidence; sub-grades unchanged; the Moderate block is kept in previous). (4) The k10 criterion-validity judgement is rewritten: the three concordant studies are named as one independent working-adult study and two developer-authored studies from one programme, and the disagreement with Osman 2022 is described as a threshold-free AUC against agreement at one cut-off rather than as the same threshold; grade unchanged at Moderate. (5) Every DOI in the dataset (420) resolved at doi.org on the conformance date. No fresh literature search was run: as_of is unchanged on every cell. 2 cells changed state (the grade move and the reworded entry) and carry previous with their rubric 1.4 values.
Was: rubric 1.4; populations exempt from the per-property test; one criterion list; gate token lists wider than the rubric
Now: rubric 1.5; populations one-of-two rule; criterion list split; gate and rubric reconciled; eurofound-ewcs internal consistency Low; k10 judgement rewritten
C-0007 · every record
Rubric 1.3 to 1.4 before first publication, reconciling the third review pass of v0.6.0 (fixes G-A to G-I, rulings Q1 to Q5). The reviewer found no grade demonstrably wrong. (1) The High precondition names the statistics of each property and the gate tests them: 9 precondition_evidence entries on 4 cells named no statistic of the property graded and leave the table; every one of those cells keeps two or more qualifying entries and stays High. One entry's statistic is reworded to the abstract's word for the coefficient (Rahimi 2025, value unchanged). (2) high_basis is renamed precondition_evidence (schema 0.7). previous blocks carry it as an eighth field: the evidence a cell carried while High, restored from v0.5.0 on the five cells that dropped in v0.6.0 (19 entries) and null elsewhere. (3) Grade moves: k10.criterion_validity_reference_standard: High -> Moderate (Q1: Osman 2022 disagrees with Kessler 2002, Kessler 2003 and Sampasa-Kanyinga 2018; rater judgement recorded in the findings; the four entries are archived in previous.precondition_evidence); status well-established -> contested; wpai.internal_consistency: Low -> Not-applicable (category-error; Q3 item-set rule); fcs-maps.internal_consistency: Absent -> Not-applicable (category-error; Q3 item-set rule). (4) Population tags: Hoffman 2022 general to other; Lundin 2017 general to working-adults. (5) Forms: wpai convergent validity mixed to canonical; csps-wellbeing convergent validity mixed to parent. (6) Flag hygiene: 2 basis arrays re-ordered to reason order; the gad-7 populations reason loses its argument clause; three clause descriptors leave the basis arrays. The gate now checks that a direct reason lists a working-adult sample first, that no parent or derivative cell is High, and that findings text is not duplicated within a record (pointer and category-error cells exempt). (7) Every DOI in the dataset (420) resolved at doi.org on the conformance date. No fresh literature search was run: as_of is unchanged on every cell. 12 cells changed state and carry previous with their rubric 1.3 values.
Was: rubric 1.3; one generic statistic test; high_basis; seven-field previous
Now: rubric 1.4; per-property statistic test; precondition_evidence archived in eight-field previous; item-set internal consistency; gate and rubric reconciled
C-0006 · every record
Rubric 1.2 to 1.3 before first publication, reconciling the second review pass of v0.5.0 (rulings Q1 to Q8, fixes F-A to F-M). (1) The evidence-form rule recodes instead of capping: 24 cells whose findings attribute their evidence to one form carry that form; `mixed` stands on 29 cells whose findings name two forms. The C-0005 cap on k10.internal_consistency is reversed and the cell returns to High on seven cited primary studies carrying a sample size and an alpha. (2) The High precondition is applied to the statistic of the property graded: 29 high_basis entries that disclaimed a value, carried no sample size, or carried a statistic of another property are removed. Grade moves: k10.internal_consistency: Moderate -> High (C-0005 drop reversed: the section 5 form rule now recodes the cell canonical instead of capping it; seven cited primary studies carry n and alpha); phq-9.criterion_validity_reference_standard: High -> Moderate (indirect flag on a population-sensitive property (criterion validity against a reference standard joined the list at rubric 1.3)); gad-7.criterion_validity_reference_standard: High -> Moderate (indirect flag on a population-sensitive property (criterion validity against a reference standard joined the list at rubric 1.3)); gad-7.measurement_invariance: High -> Moderate (1 cited study with a sample size and a statistic of the property graded); ghq-12.structural_validity: High -> Moderate (1 cited study with a sample size and a statistic of the property graded); ghq-12.criterion_validity_reference_standard: High -> Moderate (no coefficient-bearing study in a working-adult or general-population sample on a population-sensitive property); ons-4-life-satisfaction.measurement_invariance: Absent (population-general, pointer to parent) -> Not-applicable (category-error); ons-4-worthwhile.measurement_invariance: Absent (population-general, pointer to parent) -> Not-applicable (category-error); ons-4-happiness.measurement_invariance: Absent (population-general, pointer to parent) -> Not-applicable (category-error); ons-4-anxiety.measurement_invariance: Absent (population-general, pointer to parent) -> Not-applicable (category-error); ipaq-sf.structural_validity: Very low -> Not-applicable (category-error: a formative behaviour index has no latent structure to test). (3) Criterion validity against a reference standard is population-sensitive; High on it needs a direct or general flag. (4) Every basis descriptor names a population and carries no sample size, coefficient or verb: 267 descriptor edits and 35 prose edits that write the population noun into the sentence the descriptor is lifted from (two of them replace placeholder citations on k10 with the labelled DOI links already in the record's citation list). Flag moves that follow from re-reading the sample under section 6 as written: gad-7.populations_languages_norms: direct -> general (rubric 1.3 section 6 applied to the sample the findings describe); who-5.responsiveness_mic: indirect -> direct (rubric 1.3 section 6 applied to the sample the findings describe); ipaq-sf.responsiveness_mic: indirect -> direct (rubric 1.3 section 6 applied to the sample the findings describe). (5) On population-sensitive properties every high_basis entry carries population (working-adults, general, other). (6) The four ONS-4 item measurement-invariance cells are Not-applicable (category-error): invariance is a property of an item set. ipaq-sf.structural_validity is Not-applicable (category-error): a formative index has no latent structure. (7) none-in-working-adults leaves the absence_type codelist; the fallback rule's form clause is deleted; the WPAI and IPAQ bare DOIs are labelled; previous blocks carry seven fields on every level; the C-0005 note is restructured (abstract_contradictions keyed; the isi not-verifiable item split; the Morin 2011 note narrowed); the delpilar2025 citation authors are corrected. No fresh literature search was run: as_of is unchanged on every cell. 35 cells changed state and carry previous with their rubric 1.2 values.
Was: rubric 1.2; evidence-form cap; criterion validity against a reference standard population-insensitive; basis descriptors unconstrained
Now: rubric 1.3; evidence-form recode; criterion validity against a reference standard population-sensitive; population-noun basis descriptors; population-tagged high_basis
C-0005 · every record
Rubric 1.1 to 1.2 before first publication, reconciling the first review pass of v0.4.0 (fixes F1 to F12, rulings R1 to R5). (1) The six C-0004 grade moves are reversed (cbi.convergent_discriminant_validity: High back to Moderate; olbi.convergent_discriminant_validity: High back to Moderate; wai.convergent_discriminant_validity: High back to Moderate; single-job-satisfaction.populations_languages_norms: Moderate back to Low; single-stress.convergent_discriminant_validity: High back to Moderate; single-stress.populations_languages_norms: Moderate back to Low): a flag move never moves a grade. (2) Seven pass-one criterion cells that mixed diagnostic and organisational evidence were split so that criterion cells never share evidence: ons-4.criterion_validity_reference_standard: Very low to Very low; ons-4.criterion_validity_organisational: Absent to Absent; who-5.criterion_validity_reference_standard: Moderate to Moderate; who-5.criterion_validity_organisational: Absent to Absent; wemwbs.criterion_validity_reference_standard: Low to Low; wemwbs.criterion_validity_organisational: Absent to Absent; hse-msit.criterion_validity_reference_standard: Moderate to Absent; hse-msit.criterion_validity_organisational: Very low to Very low; perma.criterion_validity_reference_standard: Low to Absent; perma.criterion_validity_organisational: Absent to Absent; uwes-9.criterion_validity_reference_standard: Low to Absent; uwes-9.criterion_validity_organisational: Low to Low; phq-9.criterion_validity_reference_standard: High to High; phq-9.criterion_validity_organisational: Absent to Absent. (3) Five graded test-retest cells whose structured list was empty now carry the coefficients their summaries cite (phq-9: 4 entries; who-5: 5 entries; wemwbs: 3 entries; perma: 5 entries; uwes-9: 2 entries). (4) The indirectness flag takes three values (direct, general, indirect) and every graded cell was re-read under a written protocol, the AI-assisted first pass machine-checked against the cell text: 49 cells changed flag; 26 cells whose findings describe no sample rest on the record's fielded population. (5) Source pass on every High cell: sample sizes and coefficients read from the cited abstracts were appended to the findings with citations and recorded in high_basis. 32 High grades hold; 6 drop to Moderate (wai.populations_languages_norms: 0 cited studies with both a sample size and a statistic; isi.populations_languages_norms: 1 cited study with both a sample size and a statistic; eurofound-ewcs.populations_languages_norms: 0 cited studies with both a sample size and a statistic; wpai.convergent_discriminant_validity: indirect flag on a population-sensitive property; wpai.responsiveness_mic: indirect flag on a population-sensitive property; k10.internal_consistency: mixed evidence form whose findings name no second form (rubric 1.2 section 5)). The abstracts contradicted or qualified the cell text in 52 places, listed in this correction's note. 14 sentences were corrected as fact (a misattributed meta-analytic range, a study cited as a validation whose abstract reports a failed fit, a wrong alpha, an overstated kappa, a full-index claim that rests on two items, an adoption claim, a stale extraction note, a country count, two attributions, a reference-standard claim and two nuances of fit) and three citation labels were corrected across 10 cells (White 2025 to White 2023; Del Pilar Diaz-Nunez P 2025 to Díaz Gamarra 2025; Di et al 2025 to Di Matteo et al 2025). The remainder are one-year differences between a citation's online-first year and its issue year; labels carry the online-first year and stand. Statements the abstracts could neither confirm nor contradict are unchanged and listed in this correction's note. (6) Every Absent and Not-applicable cell carries absence_type and no flag; 8 cells recoded (tis-6.criterion_validity_reference_standard: Not-applicable to Absent (category-error); single-job-satisfaction.criterion_validity_reference_standard: Not-applicable to Absent (category-error); fcs-maps.criterion_validity_reference_standard: Not-applicable to Absent (category-error); copsoq-iii.criterion_validity_reference_standard: Not-applicable to Absent (category-error); was.criterion_validity_reference_standard: Moderate to Absent (category-error; near-miss text kept); wai.criterion_validity_reference_standard: Low to Absent (category-error; near-miss text kept); mbi.criterion_validity_reference_standard: Very low to Absent (category-error; near-miss text kept); csps-wellbeing.internal_consistency: Absent to Not-applicable (type error at the item-set level)); 11 Absent cells set to untested; 16 pilot confidence notes removed from cells that carry no grade for them to qualify, ten of which opened with a grade word the cell does not carry (ons-4.criterion_validity_organisational; ons-4.internal_consistency; ons-4.test_retest_reliability; ons-4.measurement_invariance; ons-4.responsiveness_mic; who-5.criterion_validity_organisational; wemwbs.criterion_validity_organisational; hse-msit.criterion_validity_reference_standard; hse-msit.test_retest_reliability; hse-msit.responsiveness_mic; perma.criterion_validity_reference_standard; perma.criterion_validity_organisational; perma.responsiveness_mic; uwes-9.criterion_validity_reference_standard; phq-9.criterion_validity_organisational; hse-msit.criterion_validity_organisational). (7) Every cell carries status and evidence_form (27 statuses derived from the grade, High to well-established and otherwise thin, and 49 forms filled as canonical); the ONS-4 item records carry a populations cell and every pointer cell names its parent. (8) Two uncited ranges removed, the WPAI bare URLs labelled, one content-validity sentence moved to record notes. (9) Relations: single-job-satisfaction: screens-for to corresponds-with; single-stress: screens-for kept; wording variant stated in the evidence; single-fatigue: screens-for to corresponds-with. (10) Scope of C-0004, restated (F9): of its 92 flag changes, 33 had a country-only reason, 28 a mixed reason and 30 a non-country reason, of which 25 were on ungraded cells; the 99 cells flagged for the first time under rubric 1.1 carried no recorded reason to re-read, which this correction's protocol re-read supplies. C-0004's text is not edited; corrections are append-only. No fresh literature search was run: as_of is unchanged on every cell. 81 cells changed state in this version and carry previous with their rubric 1.1 values; cells unchanged since C-0004 keep the rubric 1.0 previous they already had.
Was: rubric 1.1; two-value flag; High without a stated numeric basis; C-0004 grade moves in force
Now: rubric 1.2; three-value flag with checked basis; High precondition; C-0004 grade moves reversed
C-0004 · every record
Rubric 1.0 to 1.1 before first publication: indirectness re-defined by population type, not by country. Every graded cell re-read against the new rule and its indirectness flag rewritten in the form 'direct; reason' or 'indirect; reason'. 92 cells changed flag under the new definition (all but one from indirect to direct). 99 pilot-era cells that had carried no flag at all ('see findings') are flagged for the first time. 6 grades moved, all upward, on cells that had been held down only because the evidence was earned outside the United Kingdom on working-adult samples: cbi.convergent_discriminant_validity: Moderate to High; olbi.convergent_discriminant_validity: Moderate to High; wai.convergent_discriminant_validity: Moderate to High; single-job-satisfaction.populations_languages_norms: Low to Moderate; single-stress.convergent_discriminant_validity: Moderate to High; single-stress.populations_languages_norms: Low to Moderate. No fresh literature search was run: as_of is unchanged on every cell. Cells whose flag or grade changed carry a 'previous' block with the rubric 1.0 values. 78 sentences of prose on 24 records that had framed a flag, a downgrade or the reader as UK-specific were rewritten to state the population fact instead; the findings, citations and grades those sentences sit beside are unchanged, and statements that no UK norms were located stay as facts.
Was: rubric 1.0; indirectness defined relative to UK working adults
Now: rubric 1.1; indirectness defined relative to working-age populations in any country
C-0003 · Patient Health Questionnaire-9 (PHQ-9)
PHQ-9 licence wording precision: pass one described the status as 'public domain'. The steward terms grant a copyright exemption and free use rather than asserting a formal public-domain dedication.
Was: Free / public domain. Pfizer released the PHQ and GAD-7 without copyright restriction.
Now: Free for download and use; content on the PHQ Screeners site is 'expressly exempted from Pfizer's general copyright restrictions'. No permission or fee required, attribution to Kroenke et al 2001 expected. This is a free-use copyright exemption, not a formally asserted public-domain status.
C-0002 · Warwick-Edinburgh Mental Wellbeing Scale (WEMWBS) and Short Warwick-Edinburgh Mental Wellbeing Scale (SWEMWBS)
WEMWBS licence-currency update: pass one implied NHS/non-profit users obtain a no-fee non-commercial licence. Warwick Innovations introduced charges for NHS organisations from 1 December 2024.
Was: Academic and non-profit users obtain a no-fee non-commercial licence; commercial users pay a tiered fee.
Now: Registration still required; non-commercial and commercial licences separated. From 1 December 2024, charges apply to NHS organisations (NHS trusts, GP surgeries and NHS-funded bodies) on the steward's published participant-count tiers. Non-NHS academic/non-profit non-commercial registration remains available; commercial users pay fees. Verify current tier with Warwick Innovations before fielding.
C-0001 · WHO-5 Well-Being Index
WHO-5 licence mis-stated in pass one by trusting 2015/2020 review literature and missing WHO's October 2024 republication of the WHO-5 master version under a formal Creative Commons licence.
Was: Free to use; distributed at no charge, no licence fee or royalty; may be reproduced and translated with acknowledgement of source (free availability sourced to Topp 2015 / Lara-Cabrera 2020 review literature). No formal open licence stated.
Now: Copyright World Health Organization 2024; the WHO-5 master version is available under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 IGO licence (CC BY-NC-SA 3.0 IGO). Document ref WHO/UCN/MSD/MHE/2024.1. Non-commercial use and adaptation permitted with attribution and share-alike; commercial (including employer/vendor) use requires WHO permission.
Errata
No errata have been published for this dataset version. An erratum is a statement of an error in a published version that could have misled a reader, filed here on discovery and linked from the correction that fixes it.
Right of reply
An instrument's steward, developer or copyright holder may file a written response to its record. The response is published beside the record, dated and unedited, and the registry replies in the same place. A response never overrides a grade; it can trigger a correction where it shows an error of fact.
No responses have been filed yet.
Verifications
V-0001 · Turnover Intention Scale-6 (TIS-6, Roodt)
Licence verified at steward journal page (CC BY 4.0); performed at ingest by the maintaining thread, upgrading the pass-two not-verified flag. Field: identity.licence_status
Dataset changes
Dataset 0.9.0
Rubric 1.6 (C-0009); schema 0.7 unchanged. Fifth review pass (fit to publish, one grade to move) reconciled; the three findings and two notes taken. Grades frozen from first publication.
Dataset 0.8.0
Rubric 1.5 (C-0008); schema 0.7 unchanged. Fourth review pass (fit to publish) reconciled; fifth pass, scoped to the diff, pending before the noindex lift.
Dataset 0.7.0
Rubric 1.4 and schema 0.7 (C-0007). Third review pass reconciled; fourth pass, scoped to the fixes, pending before the noindex lift.
Dataset 0.6.0
Rubric 1.3 and schema 0.6 (C-0006). Second review pass reconciled; third pass pending before the noindex lift.
Dataset 0.5.0
Reconciliation of the first review pass before first publication. Schema 0.5: three-value indirectness flag with indirectness_basis; absence_type on ungraded cells; status and evidence_form on every cell; high_basis on High cells; corresponds-with relation; previous with six fields; integrity generated. Rubric 1.2 stamped on every cell. Correction C-0005.
Dataset 0.4.0
Purpose reset before first publication. Schema 0.4: indirectness strings carry the reason; 'relations' on every record (item-of, short-form-of, screens-for, embeds), each citing its study; 'previous' block on re-rated cells. Rubric 1.1 stamped on every cell. Freeze dated from first publication. Correction C-0004.
Dataset 0.3.0
Structural release implementing the 1 Sep 2026 external-review reconciliation. No grade, status or finding changed. Cell level: evidence_state (assessed / assessed_absent / not_assessed / not_applicable), rubric_version 1.0, as_of, grade_last_confirmed (2026-07-12, pass two) and review_due, set on the six cells the September sweep added citations to (who-5 invariance and populations; ghq-12 structural; olbi internal consistency and populations; wemwbs populations). Record level: licence_class for employer or vendor use with a class note (founder ruling B1: classes replace figures; the WEMWBS tier examples are removed from prose and live on the archived steward pages), licence_page_archived_url from the 1 Sep Internet Archive capture (ruling B4), one stewardship field omitted from the public projection. Top level: rubric pointer, rater disclosure with grade moves frozen (ruling B2), evidence-state and licence-class definitions, steward-product disclosure policy, errata list and right-of-reply protocol. Follow-ups from the September sweep applied: wemwbs licence_verified_date refreshed to 2026-09-01 after the live diff; olbi Sivadas 2026 DOI attached. Minor version bump because the schema moved.
Dataset 0.2.2
September 2026 evidence sweep, first repo-connected run (T-283 step 1). Patch bump: citation-only additions, no grade or status moved, nothing deleted. who-5: Jairoun 2026 Arabic UAE student validation added (alpha 0.833, composite reliability 0.848, configural/metric/scalar gender invariance). ghq-12: Fung 2026 Chinese student factor-structure evaluation added (two-factor retained over better-fitting bifactor on parsimony grounds). olbi: Sivadas 2026 Malayalam teacher adaptation added (S-CVI 0.97; alpha 0.84 total, 0.80 exhaustion, 0.61 disengagement). wemwbs: Tang 2026 Chinese cardiac CTT+IRT validation added as a 12-20 July registry-freeze-gap catch-up item. last_reviewed set to 2026-09-01 on the 28 swept watchlist records (csps-wellbeing, eurofound-ewcs and fcs-maps were not swept and are unchanged). WEMWBS licence page revision flagged for manual verification in evidence-sweeps/OWHS-evidence-sweep-2026-09.md; licence_status deliberately untouched pending a live-page diff per schema rule 7. Sweep method this run: web search only (Europe PMC, PubMed and Crossref API access blocked by the execution environment); citations corroborated by independent second searches.
Dataset 0.2.1
ONS-4 item split per schema v0.2 rule 4: added four first-class item records (ons-4-life-satisfaction, ons-4-worthwhile, ons-4-happiness, ons-4-anxiety) with parent_id ons-4; parent gains an items list. Graded evidence remains at set level (item properties reference the parent with evidence_form 'parent'; internal consistency and structural validity are Not-applicable at item level). Item records carry question_bank_ref so the registry and the question bank share one canonical per-item description. The 27-row grade matrix is unchanged; item records are children, not matrix rows.