# Knowledge v2 merge report (2026-10-01)

Scope of every number is stated where it appears. "Scratch merge" = `data/datasets/knowledge-src/merged-v2/` (base `APS-WSI/knowledge` + `final-ebr/` + `tcga-bracs/` with all audit fixes). Nothing was committed or pushed. No Python file or test was edited.

## 1. Status (read this first)

- Scratch merge: `KnowledgeBase.validate()` returns **no problems**; 189 passages, 346 edges, 88 sources.
- **NOT applied to `APS-WSI/knowledge/`.** The files were copied in, `pytest APS-WSI/tests` gave **312 passed, 1 failed**, so per the rule the backup was restored (verified identical with `diff -r`; baseline 313 passed again). The failure is a pinned expectation in an existing test, see section 8.
- Real dir is unchanged: 43 passages, validate clean, 313 tests pass.
- To apply once the owner accepts the one-line test conflict: see section 9 (apply and rollback commands).

## 2. Counts (scope: scratch merge)

| Item | Base | + final-ebr | + tcga-bracs | Merged |
|---|---|---|---|---|
| Passages | 43 | 72 | 74 (73 staged + 1 new) | **189** |
| Edges (facts) | 72 | | | **346** (has_feature 201, distinguished_from 85, grade_range 26, lacks_feature 19, variant_of 15) |
| Sources | 7 | 43 new records (incl. 1 licence-page record) + 2 overrides of base records | 38 (incl. 1 licence-page record; 3 base ids used by its passages were missing from its source file and are in the merged file) | **88** |
| Gold questions | 28 | 79 (72 staged + 7 proposed) | 72 | **179** (tcga-nsclc 24, bracs-fine 62, bracs-coarse 4, ebr-coarse30 89) |
| Passages by dataset (a passage can serve several) | | | | ebr-coarse30 92, bracs-fine 57, bracs-coarse 56, tcga-nsclc 40 |
| Passages with he_observable=false | | | | 40 of 189 |

Sources by tier (all 88 records, scope: sources.json):

| Tier | Count | Meaning | Passages whose `source_id` is of this tier |
|---|---|---|---|
| T1 | 2 | guideline / society / government: `iccr-lung`, `pmc10764572` (Japanese Breast Cancer Society guideline) | 6 |
| T2 | 47 | peer-reviewed journal article (16 reviews, 13 case reports, 18 original/other studies incl. both BRACS records) | 117 |
| T3 | 39 | Libre Pathology (30 pages + 2 licence-page records), Wikipedia (7) | 66 |
| T4 | 0 | | 0 |

Passages whose every named source (`source_id` plus sources named in the locator) is T3: **61** of 189. Tier fields (`tier`, `tier_label`, `legitimacy`, plus `journal`, `year`, `article_type` for papers) are stored per source in `sources.json`; the extra fields are accepted by the schema (verified: validate clean), so the module was not touched. `source-tiers.json` is a standalone copy of the same table (one line per source, includes journal, year, article type, licence and the legitimacy note). Journal, year and publication type of all 38 PMC records come from the Europe PMC REST API (fetched 2026-10-01); T1/T3 notes are from the records. Not verified: impact factors or indexing beyond PubMed Central/Europe PMC.

Tiering caveats: (a) 13 T2 sources are case reports and pmc11089008 is a brief report (Letter) on 7 cases, so T2 does not mean strong evidence; (b) both BRACS records are T2 as a peer-reviewed dataset paper but the cited text is the arXiv preprint; (c) no WHO Classification, NCBI Bookshelf or CAP protocol text is among the sources (T1 has only ICCR lung and the Japanese guideline), so every WHO CNS5, WHO breast and WHO thoracic statement is second-hand through reviews or the ICCR checklist; (d) Libre pages with a revision dated 2020 or earlier are flagged in their legitimacy note as predating WHO 2021.

## 3. Single-tier (T3-only) facts

**97 of 346 edges** have support only from T3 passages (same fact triple/feature appears in no T1 or T2 passage). They carry `"single_tier": "T3"` inside the fact in the corpus files; the full list with passage ids is `single-tier-facts.json`. By relation: has_feature 58, distinguished_from 21, grade_range 8, variant_of 6, lacks_feature 4. Scope: edge-level, by fact triple plus feature; support counted from `source_id` and sources named in the locator, not from names in free text except one passage (lung-luad-vs-lusc-location-clinical, set by hand).

**Classes whose discriminating (`distinguished_from`) edges rest on T3 only:**

| Class (dataset) | distinguished_from edges | of which T3-only | Which edges |
|---|---|---|---|
| Anaplastic meningioma (ebr-coarse30) | 2 | 2 | 20 or more mitoses per 10 HPF; carcinoma- or sarcoma-like morphology (Libre meningioma page) |
| Atypical meningioma (ebr-coarse30) | 2 | 2 | 4 or more mitoses per 10 HPF; brain invasion (Libre) |
| Anaplastic astrocytoma IDH-wt, legacy (ebr-coarse30) | 2 | 1 | no necrosis or microvascular proliferation versus GBM (Libre); the other edge comes from T2 |
| Pilocytic astrocytoma | 5 | 2 | low cellularity against medulloblastoma and ependymoma (Libre) |
| Glioblastoma IDH-wt | 6 | 3 | border and processes against ependymoma and metastasis (Libre); no necrosis rule against the legacy AA label (Libre) |
| Haemangioblastoma | 2 | 1 | EMA-negative, NSE-positive against metastasis (Libre) |
| ADH versus DCIS (bracs-fine) | extent edges | both extent edges in Libre/Wikipedia passages | a T2 passage lists the size rules but, by owner default, as "sources differ" |
| LUAD (tcga-nsclc) | 4 | 1 | vacuolated cytoplasm against LUSC (Libre); LUAD sub-pattern contrasts (papillary vs micropapillary, acinar vs lepidic, solid-mucin) are Libre/Wikipedia |
| PB versus IC (bracs-fine) | 3 (PB) | 1 | lobular arrangement and two cell layers in sclerosing adenosis (Libre) |

Classes with many T3-only has_feature facts (descriptive, not discriminating): meningothelial meningioma 5 of 6 facts, ependymoma and anaplastic ependymoma 5 of 7, UDH 6 of 13, ADH 5 of 20, PB 5 of 10, medulloblastoma 4 of 6, metastasis 4 of 6. **No dataset label rests on T3 only**: each of the 42 labels has at least one T2 passage (scope: label to entity-cited passages in the dataset). The weakest are Haemangioma (Libre plus 1 T2 case report), Lipoma (1 T2), Haemangiopericytoma (1 T2 plus Libre), PB (2 T2 against 5 T3 named sources) and meningothelial meningioma (2 T2 against 3 T3).

The BRACS "expert definition" statements for N, PB, UDH, FEA, ADH, DCIS and IC are T2 (the BRACS paper) but are annotation-protocol text of the dataset authors, not an external guideline.

## 4. Coverage per dataset and class (rule: covered = at least 2 passages citing the entity in that dataset, including one differential; partial = 1 passage or no differential; uncovered = none)

"Differential" = a passage with topic `differential` or with a `distinguished_from` fact on the entity. The rule is mechanical: a passage that only states a limitation ("H&E cannot separate") counts, so read it with section 7. `label_spaces.json` keeps the recorded value; it was not promoted by me (only the staged label patch was applied: Haemangioblastoma, Haemangioma, Haemangiopericytoma, Lipoma uncovered to partial; Anaplastic astrocytoma IDH-wildtype entity repointed to the legacy entity). Passages count includes passages where the class is only a comparator, so it overstates dedicated coverage.

| Dataset | Label | Recorded | By rule | Passages | Differential passages | Sources named (T1/T2/T3) | distinguished_from edges (T3-only) |
|---|---|---|---|---|---|---|---|
| tcga-nsclc | LUAD | covered | covered | 28 | 9 | 12 (1/8/3) | 4 (1) |
| tcga-nsclc | LUSC | covered | covered | 21 | 10 | 9 (1/4/4) | 5 (0) |
| bracs-fine | N | covered | covered | 5 | 2 | 5 (0/3/2) | 1 (0) |
| bracs-fine | PB | covered | covered | 7 | 3 | 7 (0/2/5) | 3 (1) |
| bracs-fine | UDH | covered | covered | 9 | 5 | 10 (1/7/2) | 2 (0) |
| bracs-fine | FEA | covered | covered | 11 | 5 | 12 (1/8/3) | 4 (0) |
| bracs-fine | ADH | covered | covered | 17 | 9 | 16 (1/10/5) | 4 (1) |
| bracs-fine | DCIS | covered | covered | 16 | 9 | 18 (1/8/9) | 7 (1) |
| bracs-fine | IC | covered | covered | 8 | 5 | 10 (0/5/5) | 4 (0) |
| bracs-coarse | benign | covered | via members (all covered); group entity cited by 1 passage | 1 | 0 | | |
| bracs-coarse | atypical | covered | via members (all covered); group entity cited by 1 passage | 1 | 0 | | |
| bracs-coarse | malignant | covered | via members (all covered); group entity cited by 1 passage | 1 | 0 | | |
| ebr-coarse30 | Glioblastoma IDH-wildtype | covered | covered | 12 | 9 | 10 (0/7/3) | 6 (3) |
| ebr-coarse30 | Gliosarcoma | partial | covered | 4 | 1 | 3 (0/2/1) | 2 (0) |
| ebr-coarse30 | Transitional meningioma | partial | covered | 5 | 1 | 3 (0/2/1) | 1 (0) |
| ebr-coarse30 | Anaplastic meningioma | partial | covered | 6 | 2 | 4 (0/3/1) | 2 (2) |
| ebr-coarse30 | Pituitary adenoma | partial | covered | 5 | 2 | 4 (0/2/2) | 1 (0) |
| ebr-coarse30 | Oligodendroglioma IDH-mutant and 1p/19q codeleted | covered | covered | 8 | 5 | 5 (0/4/1) | 4 (0) |
| ebr-coarse30 | Haemangioma | partial | covered | 3 | 1 | 3 (0/1/2) | 0 (0) |
| ebr-coarse30 | Ganglioglioma | partial | covered | 5 | 2 | 3 (0/2/1) | 1 (0) |
| ebr-coarse30 | Pilocytic astrocytoma | partial | covered | 6 | 5 | 4 (0/3/1) | 5 (2) |
| ebr-coarse30 | Schwannoma | partial | partial | 3 | 0 | 3 (0/2/1) | 0 (0) |
| ebr-coarse30 | Anaplastic oligodendroglioma IDH-mutant and 1p/19q codeleted | covered | covered | 8 | 5 | 5 (0/4/1) | 4 (0) |
| ebr-coarse30 | Fibrous meningioma | partial | covered | 6 | 2 | 4 (0/3/1) | 2 (0) |
| ebr-coarse30 | Anaplastic astrocytoma IDH-wildtype | partial | covered | 6 | 3 | 5 (0/4/1) | 2 (1) |
| ebr-coarse30 | Angiomatous meningioma | partial | covered | 6 | 2 | 4 (0/3/1) | 0 (0) |
| ebr-coarse30 | Haemangioblastoma | partial | covered | 6 | 4 | 5 (0/4/1) | 2 (1) |
| ebr-coarse30 | Adamantinomatous craniopharyngioma | partial | covered | 5 | 2 | 4 (0/3/1) | 0 (0) |
| ebr-coarse30 | Anaplastic astrocytoma IDH-mutant | covered | covered | 7 | 2 | 4 (0/4/0) | 1 (0) |
| ebr-coarse30 | Glioblastoma IDH-mutant | covered | covered | 7 | 2 | 4 (0/4/0) | 1 (0) |
| ebr-coarse30 | Diffuse astrocytoma IDH-mutant | covered | covered | 7 | 2 | 4 (0/4/0) | 1 (0) |
| ebr-coarse30 | Ependymoma | partial | covered | 8 | 4 | 6 (0/3/3) | 3 (1) |
| ebr-coarse30 | Anaplastic ependymoma | partial | covered | 8 | 4 | 6 (0/3/3) | 3 (1) |
| ebr-coarse30 | Atypical meningioma | partial | covered | 6 | 3 | 4 (0/3/1) | 2 (2) |
| ebr-coarse30 | Metastatic tumours | partial | covered | 11 | 8 | 10 (0/6/4) | 2 (1) |
| ebr-coarse30 | Meningothelial meningioma | partial | partial | 6 | 0 | 5 (0/2/3) | 0 (0) |
| ebr-coarse30 | Langerhans cell histiocytosis | partial | covered | 5 | 2 | 4 (0/3/1) | 0 (0) |
| ebr-coarse30 | Diffuse large B-cell lymphoma of the CNS | partial | covered | 5 | 3 | 4 (0/3/1) | 1 (0) |
| ebr-coarse30 | Secretory meningioma | partial | covered | 5 | 1 | 3 (0/2/1) | 0 (0) |
| ebr-coarse30 | Haemangiopericytoma | partial | covered | 4 | 1 | 2 (0/1/1) | 3 (0) |
| ebr-coarse30 | Lipoma | partial | covered | 2 | 1 | 2 (0/1/1) | 0 (0) |
| ebr-coarse30 | Medulloblastoma non-WNT/non-SHH | partial | covered | 6 | 2 | 4 (0/2/2) | 0 (0) |

Result by rule (scope: 42 label rows incl. 3 coarse group entities): uncovered 0; partial only Schwannoma and Meningothelial meningioma (no differential passage; both have limitations in text) plus the 3 coarse group entities; every other label is covered by the rule. Recorded values differ: 24 EBR labels remain recorded "partial" because the owner has not decided to promote them.

`distinguished_from` edges, merged (scope: scratch merge, edges counted per passage so duplicated facts across passages count more than once): 85 across 34 class pairs. Weak pairs unchanged from the audit: no edge for PB vs UDH, PB vs ADH (the BRACS benign-to-atypical boundary), N vs UDH, FEA vs DCIS, ADH vs IC, N vs DCIS.

## 5. Sanity (scope: top-3 BM25 retrieval over the scratch merge; the questions were written by the same authors as the passages, so this is a smoke test of wording, not an independent evaluation)

| Dataset | Existing gold (original file) | Staged / proposed new gold | Total |
|---|---|---|---|
| tcga-nsclc | 6 / 8 | 16 / 16 | 22 / 24 |
| bracs-fine | 8 / 10 | 51 / 52 | 59 / 62 |
| bracs-coarse | none | 4 / 4 | 4 / 4 |
| ebr-coarse30 | 10 / 10 | 79 / 79 (72 staged + 7 proposed) | 89 / 89 |

Reference: untouched `APS-WSI/knowledge` gives tcga-nsclc 6 / 8, bracs-fine 9 / 10, ebr-coarse30 10 / 10 (43 passages). Misses in the merge, all listed:

1. tcga-nsclc "which features separate squamous from adenocarcinoma of the lung" (expect lung-luad-vs-lusc, lung-lusc-cytomorphology): **already missed on the untouched base**.
2. tcga-nsclc "how is lung adenocarcinoma graded by pattern" (expect lung-luad-grading): **already missed on the untouched base**.
3. bracs-fine "difference between atypical ductal hyperplasia and ductal carcinoma in situ" (expect breast-adh-vs-dcis, breast-hard-boundaries): **already missed on the untouched base**.
4. bracs-fine "why are atypical lesions clinically important" (expect breast-atypical-significance): **new regression** (was a hit on the base; top 3 is now breast-fea-vs-adh-expert-disagreement, breast-coarse-groups-bracs, breast-ic-bracs-definition). The audit documented this (rank 3 to 4, margin 0.19 BM25) and offered to widen the expect with breast-fea-vs-adh-expert-disagreement. I did **not** widen it: the audit called it an owner option, not a proven alternative.
5. bracs-fine (staged) "How is nuclear grade of DCIS judged against red blood cells?" (expect breast-dcis-libre-features-grading): miss in the full merge only. A scratch merge without the EBR passages (base + tcga-bracs) hits it, so the cause is the shared BM25 statistics: the EBR vascular passages contain "red blood cells" and lower those terms' weight (idf) and shift the average length. Not widened, no audit covers it. (The base passage breast-dcis-grades also mentions nuclei relative to an erythrocyte, but it is not in the top 3 either.)

Gold changes made (every one):

- ebr-coarse30: question "How to separate a glial tumour from a metastatic carcinoma at high power" gained expect ['cns-metastasis-libre-histology']. Reason: the rule is also in that passage, which the sanity check ranks first; both are correct hits
- ebr-coarse30: NEW question "Why can a tumour that looks like grade 2 astrocytoma on H&E be grade 4" (from `final-ebr/gold.proposed-changes.json`)
- ebr-coarse30: NEW question "Molecular glioblastoma without necrosis or microvascular proliferation TERT EGFR" (from `final-ebr/gold.proposed-changes.json`)
- ebr-coarse30: NEW question "IDH wildtype grade 2 or 3 astrocytoma outcome compared with glioblastoma" (from `final-ebr/gold.proposed-changes.json`)
- ebr-coarse30: NEW question "Does medulloblastoma histology determine the molecular group" (from `final-ebr/gold.proposed-changes.json`)
- ebr-coarse30: NEW question "Langerhans cell nuclear grooves eosinophils and overlapping histiocytic lesions" (from `final-ebr/gold.proposed-changes.json`)
- ebr-coarse30: NEW question "Can a metastasis infiltrate brain beyond a well-demarcated border" (from `final-ebr/gold.proposed-changes.json`)
- ebr-coarse30: NEW question "Perinuclear halo and chicken wire vessels in a tumour that is not oligodendroglioma" (from `final-ebr/gold.proposed-changes.json`)
- bracs-fine: question "Do usual hyperplasia and flat epithelial atypia differ molecularly or by basal cytokeratin staining?" gained expect ['breast-udh-vs-fea-ihc-molecular']. Reason: The IHC and molecular contrast now lives in its own he_observable=false passage.

Not applied: the two optional tcga-bracs expect additions (FEA cytologic features adds breast-fea; atypical vs malignant adds breast-coarse-groups-bracs) because both questions already hit; the widening for "why are atypical lesions clinically important"; nothing else was widened or removed. Staged gold otherwise unedited.

## 6. Fixes applied (every one)

Applied from the audits and owner approval (log of each step in this section; ids are passage or source ids):

- **tcga-bracs passage patches: 33** from `tcga-bracs/audit/fixes.proposed.json` (the audit report counts 15 faithfulness/locator/flag fixes and 17-18 wording-only; the file holds 33 patch objects and I applied all 33 as written, field by field via `replace`): breast-udh-vs-fea-indirect, breast-dcis-vs-ic-myoepithelium, breast-ic-bracs-definition, breast-n-bracs-paper-tissue, breast-n-vs-pb-ratio-and-glands, breast-pb-sclerosing-adenosis-libre, breast-ic-wikipedia-nst, breast-fea-interobserver-variability, breast-adh-interobserver-variability, breast-adh-vs-dcis-size-rules-disagree, breast-dcis-vs-ic-confined-to-basement-membrane, breast-dcis-low-grade-vs-high-grade-pathway, breast-adh-vs-dcis-same-biology, breast-dcis-grading-variability, lung-luad-libre-differential, lung-luad-vs-lusc-libre-bridges-mucin, lung-luad-wikipedia-precursors, lung-luad-patterns-who-2021-wikipedia, lung-luad-solid-mucin-lineage, lung-luad-pattern-interobserver-variability, lung-lusc-libre-microscopy, lung-lusc-keratinization-grading-cohort, breast-pb-fibroadenoma-libre-wikipedia, breast-udh-ck56-er-pattern, breast-fea-bracs-and-expert-definition, breast-fea-proliferation-and-risk, breast-adh-libre-features-size, breast-adh-wikipedia-size-criteria, breast-adh-review-morphology, breast-fea-vs-adh-architecture, breast-adh-vs-dcis-libre-table, breast-dcis-subtypes-architecture, breast-ic-libre-invasion-criteria.
- **1 new passage:** breast-udh-vs-fea-ihc-molecular (he_observable=false), inserted after breast-udh-vs-fea-indirect (now an H&E passage labelled as inference).
- **final-ebr:** its 72 corrected passages were taken as delivered (24 audit edits, 7 depth passages, namespaced `cns:` features, dangling entities and facts removed, 10 first authors added, 7 new PMC sources). Its gold change and label patch applied as in section 5 and section 4.
- **Legacy entity link:** `cns:anaplastic-astrocytoma-idh-wt-legacy` added to the entity lists of the old base passages cns-legacy-names and cns-gbm-definition.
- **Sources:** pmc7698816 corrected to Weng CF, Cancers 2020;12(11):3426, CC BY 4.0 (was Moreira); pmc12308192 licence confirmed CC BY 4.0 and first author Patel S added (Europe PMC); the three source ids missing from the tcga-bracs source file (pmc12308192, pmc7698816, pmc8833401) are present; all 32 Libre records (30 pages + 2 licence-page records) carry `CC BY-NC-SA 3.0 (Copyrights page says NC SA 3.0)` plus a `license_note` about the footer saying 4.0; raw licence strings of 16 PMC records normalised to `CC BY x.y (licence statement read in the article text 2026-10-01)` where the snapshot gives a version (otherwise "version not stated"); snapshot paths made repo-relative.
- **BRACS licences, both recorded** in `bracs-paper` and `bracs-paper-arxiv`: arXiv non-exclusive distribution licence for the arXiv version, CC BY-NC 4.0 for the journal version (Database 2022, baac093, PMC9575967).
- **Four stretched rubric facts (KONGMING-REVIEW.md, P0-3), passage side:**
  1. Secretory meningioma, psammoma: cns-secretory-histology now says the article calls the intracytoplasmic structures "pseudopsammoma bodies", a different structure from psammoma bodies, and that rarity of psammoma bodies is what the case report names as defining.
  2. Secretory, no atypia: same passage now scopes "absence of cellular atypia" to the single case report and says it is not a stand-alone exclusion rule.
  3. Fibrous meningioma whorls: cns-fibrous-libre-histology now quotes Libre "few to none" and states that occasional whorls do not exclude the subtype; cns-fibrous-vs-transitional-histology marked a single case report.
  4. FEA columnar cells: no passage rests the claim on a paper title (it is in breast-fea-libre-features from Libre and in the 2022 review passage as CCC/CCH with atypia), so **text left unchanged**; the Libre-only edge breast:FEA has_feature breast:columnar-cells is flagged single-tier T3. The rubric files themselves (`verified/rubric.json`) were not edited, those four downgrades remain to be made there.
- **Facts changed for an owner default:** breast-adh-vs-dcis-size-rules-disagree: both facts used `breast:size-under-2mm` / `breast:two-complete-spaces-or-over-2mm`, which encode one threshold as the rule; both now use `breast:extent-size-rule-sources-differ`.
- **Single-tier marks:** 97 facts tagged `single_tier: "T3"`.

## 7. Decisions taken by default (owner defaults from the task) and weak spots

Defaults used:
- ADH size rule: recorded as "sources differ", no single primary rule (passage lists Libre, Wikipedia, BRACS, 2018, 2012, 2022 reviews and the Brazilian series; fact feature neutral).
- UDH versus FEA: corrected H&E passage plus a separate he_observable=false IHC/molecular passage; the H&E contrast stays labelled as inference.
- LUAD growth patterns (lepidic, acinar, papillary, micropapillary, solid) stay passage entities only, not labels.
- `cns:anaplastic-astrocytoma-idh-wt-legacy` kept as an entity; legacy label points to it (coverage partial).
- pmc7698816 author Weng (Cancers 2020, CC BY 4.0). Libre licence string as above. Both BRACS licences recorded.
- Libre paraphrase is allowed (non-commercial, paraphrase only); Wikipedia share-alike text only paraphrased.

Weak spots (scope: this corpus; none was fixed, all are limits of the sources read):

1. **Haemangioblastoma versus metastatic RCC is IHC only**; no H&E rule in any source (stated in the passage). Only one distinguishing edge is T3-only (EMA-negative, NSE-positive in Libre); the others are T2.
2. **Lipoma**: Libre plus one autopsy case report; not distinguishable from normal fat on H&E per the source; no EBR comparator class. 2 passages, 1 T2 source.
3. **LCH**: no H&E differential against an EBR class; the "differential" passage states the limitation. Libre page is a 2016 revision; one T2 source is a Papanicolaou cytology case.
4. **Haemangioma**: whether "Haemangioma" in EBRAINS means cavernous venous malformation is an assumption stated in the passage; 3 passages, 1 T2 source (case report).
5. **Single-source or thin classes** (sources named, scope: entity-cited passages, includes comparator mentions): haemangiopericytoma 2 sources, lipoma 2, gliosarcoma 3, transitional 3, ganglioglioma 3, schwannoma 3, secretory 3, haemangioma 3; oligodendroglioma definition, grading and molecular passages rest on PMC10216527 alone. pmc10216527 and pmc9427889 (the old base) were not re-verified.
6. **T3-only classes**: no label rests on T3 alone, but the meningioma grade thresholds (atypical and anaplastic mitosis counts) and the "anaplastic astrocytoma IDH-wt is mostly underdiagnosed GBM" claim are Libre only; ependymoma, medulloblastoma, meningothelial and PB descriptive facts lean on Libre (30 of 88 sources are Libre pages, several predate WHO 2021).
7. **No WHO or Bookshelf source**: authoritative definitions for UDH, FEA, ADH and the lung patterns are second-hand.
8. **BRACS benign-to-atypical boundary** has no discriminating edges (PB vs UDH, PB vs ADH, N vs UDH, FEA vs DCIS, ADH vs IC, N vs DCIS).
9. UDH versus FEA has no head-to-head source (inference by design); ADH/DCIS size rules disagree between sources; Libre disagrees with itself on ADH ducts and nucleoli.
10. The merged tcga-nsclc sanity is 6 / 8 on the existing questions and the three base misses are not fixed; the TTF-1/Napsin A single-cohort percentages are labelled single-cohort in the text.
11. No retrieval was run against the real encoded case bank; no source outside `snapshots/` was checked; the per-sentence evidence logs in the audits are pointers, not full re-verification of every quote.

## 8. Test conflict (why the merge is not live)

`APS-WSI/tests/test_knowledge_base.py::test_coverage_reports_gaps` asserts `"Lipoma" in kb.coverage("ebr-coarse30")["uncovered"]`. The owner-approved label patch makes Lipoma `partial` (2 passages exist), so the test fails on the merged data. Everything else passes (312 of 313; 116 subtests). Per the instruction the tests were not edited and the backup was restored. Options for the owner: (a) change that test to a label that stays uncovered (none remains in ebr-coarse30 under the staged patch: all 4 formerly uncovered labels are now partial), or assert the partial list; (b) keep Lipoma uncovered in `label_spaces.json` and apply the merge otherwise, which contradicts the 2 Lipoma passages and the proofread. The corpus is otherwise compatible: test_verify_wsi and the other knowledge tests pass on the merged data.

## 9. Apply and rollback commands

Merged files are in `data/datasets/knowledge-src/merged-v2/` (sources.json, label_spaces.json, gold_questions.json, corpus/*.json; also source-tiers.json and single-tier-facts.json which are not read by the module). Backup: `data/datasets/knowledge-src/backup-APS-WSI-knowledge-2026-10-01/` (verified identical to the original with `diff -r`).

Apply (after the test decision):

```
cd /home/than-minh/Project/Thesis
M=data/datasets/knowledge-src/merged-v2
cp $M/sources.json $M/label_spaces.json $M/gold_questions.json APS-WSI/knowledge/
cp $M/corpus/*.json APS-WSI/knowledge/corpus/
```

Rollback (exact):

```
cd /home/than-minh/Project/Thesis && rm -rf APS-WSI/knowledge && cp -a data/datasets/knowledge-src/backup-APS-WSI-knowledge-2026-10-01 APS-WSI/knowledge && diff -r APS-WSI/knowledge data/datasets/knowledge-src/backup-APS-WSI-knowledge-2026-10-01 && echo RESTORED
```

Validate: `/home/than-minh/miniconda3/envs/thesis/bin/python APS-WSI/knowledge_base.py validate` (it reads the real dir only; the scratch merge was validated through `KnowledgeBase(Path(...))`).