Dr. Diehl 2026-10-05: NK species audit, CL issue drafts, group 1 ILC hierarchy proposal

.claude/skills/gap-issue-filing/SKILL.md

@@ -23,8 +23,13 @@ how many logged gaps still lack a tracking issue.
 
 - `reports/gaps.tsv` — the source rows. Columns: `soulcap_abbreviation`,
   `soulcap_full_name`, `gap_type`, `description`, `related_cl_ids`,
-  `evidence`, `status` (`open` / `filed_as_issue`), `repo_issue`,
-  `date_logged`.
+  `evidence`, `status` (`open` / `filed_as_issue` / `draft_for_review` /
+  `on_hold`), `repo_issue`, `date_logged`. `draft_for_review` rows are CL
+  issue drafts waiting for Alex (and sometimes Dr. Diehl) to approve them;
+  change them to `open` only after that approval. `on_hold` rows wait on an
+  open question named in their description. Never file either directly.
+  `withdrawn` (the proposal was dropped) and `not_a_cl_issue` (the problem
+  turned out not to be in CL) are closed: never file them.
 - `.github/ISSUE_TEMPLATE/mapping_gap.yml` — the template whose fields map
   directly onto `gaps.tsv` columns: `soulcap_abbreviation` → title, `gap_type`
   → the dropdown, `description` + `related_cl_ids` + `evidence` → the
@@ -38,7 +43,7 @@ how many logged gaps still lack a tracking issue.
 ## Step 1 — find what's actually unfiled
 
 Read `reports/gaps.tsv` and list every row where `status` is `open` (not
-`filed_as_issue`). Note two rows sharing a `soulcap_abbreviation` group (e.g.
+`filed_as_issue`, `draft_for_review`, `on_hold`, `withdrawn` or `not_a_cl_issue`). Note two rows sharing a `soulcap_abbreviation` group (e.g.
 `NK1 / NK3`) file as **one** issue, matching how the row itself is structured
 — don't split a single row into multiple issues.
 

.claude/skills/mapping-audit/SKILL.md

@@ -71,10 +71,22 @@ is to question it, not just print it — for each row, ask:
    fixes changes the term's markers/definition/label out from under a mapping
    made before the fix landed.
 3. **Does literature actually back a species-specific claim the mapping
-   relies on?** Cross-check against `reports/pro_marker_species_support.tsv`
-   — if a mapping leans on a marker that TSV records as
-   `insufficient_evidence` or a different species than SOULCAP's row assumes,
-   that's a real finding, not noise.
+   relies on?** Cross-check against `reports/pro_marker_species_support.tsv`,
+   using `human_status` (SOULCAP rows are human). The four values mean
+   different things, so report them separately and never merge them:
+   - `contradicted`: a real finding. Evidence goes against the CL assertion
+     (`soulcap-audit` raises `species_marker_contradicted`).
+   - `no_evidence_found`: searched, nothing found. Note it as an evidence gap,
+     **not** a contradiction.
+   - `not_applicable`: the species question doesn't apply (term or protein
+     limited to the other species, or a human-only marker such as HLA-DRA or
+     NCAM1/CD56 on the mouse side). **Never a failure.**
+   - blank / `not_assessed`: not checked yet; say so.
+
+   Use `pro_species_support.species_status(row, "human")` to read a row: it
+   fills a blank with `not_applicable` when a rule applies and never
+   overrides a recorded value. The legacy `species_support` column
+   (`insufficient_evidence` etc.) is superseded by these columns.
 4. **Was it marked `uncertain` for a reason that's since been resolved (or
    gotten worse)?** Re-read the `note` field and check whether the ambiguity
    it describes is still accurate.

.claude/skills/pro-marker-species-support/SKILL.md

@@ -120,12 +120,16 @@ own discipline here.
 
 `reports/pro_marker_species_support.tsv` has one row per directly asserted
 (CL term, PRO marker) pair for every CL term in
-`mappings/curated_mappings.tsv` (87 pairs as of 2026-09-28). The question is
+`mappings/curated_mappings.tsv` (87 pairs as of 2026-09-28), plus every term
+under group 1 ILC (`CL:0001067`, NK cells included) from the October 2026 NK
+audit (`audit_batch = nk_2026-10`; see `reports/nk_marker_audit_2026-10.md`).
+Extra terms can be added the same way: scope them from Ubergraph and pull
+their labels from `cl_pro_relationships.tsv`. The question is
 Dr. Diehl's: **is each marker, as CL asserts it, valid in human, mouse, or
 both?**
 
 **Never edit the first 10 (legacy) columns** of an existing row. They record
-the original human-only check from issue #11. All new work goes in the 7
+the original human-only check from issue #11. All new work goes in the
 columns after them. **Where the new columns disagree with the legacy
 `species_support` value, the new columns take precedence** (e.g. plasmablast
 "lacks CD138", where the legacy value still says `human`).
@@ -134,11 +138,43 @@ columns after them. **Where the new columns disagree with the legacy
 |---|---|
 | `cl_id`, `cl_label`, `pr_id`, `pr_label`, `cd_synonym` | From `cl_pro_relationships.tsv` — pull them programmatically, don't retype them. |
 | `species_support`, `evidence_source`, `citation`, `quote`, `source_ref` | **Legacy** (issue #11 human-only check). Read-only. |
-| `species_scope` | `human`, `mouse`, `both` or `unresolved` (blank = not yet assessed). |
-| `human_citation`, `human_quote` | Citation (author, year, PMID, PMCID) and an exact quote for the human side. |
+| `species_scope` | `human`, `mouse`, `both` or `unresolved` (blank = not yet assessed). Derived from the two status columns. |
+| `human_status`, `mouse_status` | One value per species, from the controlled list below. |
+| `human_citation`, `human_quote` | Citation (author, year, PMID, PMCID) and an exact quote: the evidence behind `human_status`. For `contradicted`, this is the contradicting quote. |
 | `mouse_citation`, `mouse_quote` | The same for mouse. |
+| `counter_citation`, `counter_quote` | For a `contradicted` verdict, the strongest quote on the other side (supporting the assertion), so both sides are recorded. Blank if none was found. |
 | `note` | Free text explaining the verdict, e.g. "not tested in mouse, no reagent", "ortholog exists but not expressed", "contradicted in mouse", or "species from PRO label". |
 | `checked_on` | Date the row was last assessed. |
+| `audit_batch` | Which audit last set the row, e.g. `nk_2026-10`. Blank for earlier rows. |
+
+Status values for `human_status` / `mouse_status` (Dr. Diehl, 2026-10-05).
+Keep these four apart; they mean different things:
+
+| Value | Meaning |
+|---|---|
+| `supported` | An exact quote supports the assertion as CL states it. |
+| `contradicted` | An exact quote goes against it. Put any supporting quote in `counter_*`. |
+| `no_evidence_found` | Searched, found nothing. **Not** a contradiction. Say in `note` what was searched. |
+| `not_applicable` | The species question doesn't apply: the CL term is limited to the other species (", human" / ", mouse"), the PRO term is the other species' protein, or the marker is human-only. Human-only markers so far: HLA-DRA (mouse ortholog H2-Ea is a different protein) and NCAM1/CD56 for NK cells. |
+| *(blank)* | Not assessed. |
+
+The same vocabulary is in code: `pro_species_support.SPECIES_STATUSES`,
+`HUMAN_ONLY_MARKERS`, `not_applicable_reason()`, `species_status()` and
+`species_scope()`. `soulcap-audit` uses them. It reports `contradicted` rows
+as findings and counts every status per species in `audit_summary.md`. Add
+a new human-only marker to `HUMAN_ONLY_MARKERS` (with a test), not only to a
+note. `tests/test_pro_species_support.py` checks that every recorded value is
+in the vocabulary, that `supported`/`contradicted` have a quote, and that a
+species which can't apply is never marked anything but `not_applicable`.
+
+Derive `species_scope` from them: `both` if both are `supported`; `human`
+if human is `supported` and mouse is `not_applicable` or `contradicted`
+(likewise `mouse`); otherwise `unresolved`.
+
+Lineage-negative markers (lacks CD3/CD19/CD20/CD14 on non-T, non-B,
+non-myeloid cells) are generally safe across species (Dr. Diehl), but still
+look for mouse evidence and record `no_evidence_found` if there is none. Never
+mark them `supported` on expert opinion alone; put the opinion in `note`.
 
 Rules for `species_scope`:
 

.gitattributes

@@ -0,0 +1,2 @@
+# The benchmark reserve test hashes these files; keep LF on every OS.
+mappings/*.tsv text eol=lf

.gitignore

@@ -49,3 +49,6 @@ htmlcov/
 *.swp
 .idea/
 .vscode/
+
+# Library copies of paywalled papers, for local quote checks only (never commit)
+literature/paywalled/

CLAUDE.md

@@ -75,6 +75,7 @@ AST in `marker_syntax.py`, compiled/evaluated by `phenotype.py`.
 | `pyproject.toml` | Project metadata, deps, and the CLI entry points. |
 | `.mcp.json` | MCP servers available in this project (see below). |
 | `.env` | `ASTA_API_KEY=...` (gitignored; required for the Asta MCP server). |
+| `.gitattributes` | Forces LF line endings on `mappings/*.tsv`, whose hashes `tests/test_benchmark_reserve.py` checks (Windows `core.autocrlf` would otherwise break it). |
 
 When you create a new artifact, put it in the established location and link it
 from here and the README so locations stay consistent. Do not scatter

ROADMAP.md

@@ -141,6 +141,18 @@ Mgr` tab are covered — B cell, DC, ILC, iNKT, NK cell, T cell, gdT, MAIT
 category, not a cell type). Still missing: monocyte, neutrophil, basophil,
 mast cell, granulocyte, MDSC.
 
+**Human/mouse marker validity (Dr. Diehl, 2026-09-28 and 2026-10-05):**
+[reports/pro_marker_species_support.tsv](reports/pro_marker_species_support.tsv)
+records, per CL marker assertion, a `human_status` and `mouse_status`. The
+values are `supported` / `contradicted` / `no_evidence_found` /
+`not_applicable`, each verdict backed by a quote checked against the paper.
+- **Done (2026-10-05):** every marker on the group 1 ILC / NK subtree
+  (66 pairs on 28 terms) and on plasmablast (4 pairs). See
+  [nk_marker_audit_2026-10.md](reports/nk_marker_audit_2026-10.md) and
+  [plasmablast_marker_audit_2026-10.md](reports/plasmablast_marker_audit_2026-10.md).
+- **Remaining:** the other pairs on mapped CL terms (blank status =
+  not assessed).
+
 ---
 
 ## Milestone 3 — Audit curated CL mappings ⛔ (blocked)
@@ -163,6 +175,14 @@ against current CL marker axioms and literature, catching drift — without
 needing the sheet columns. It cannot capture curator *intent*, so this
 milestone stays ⛔ for the full deliverable until the sheet changes.
 
+**Update (2026-10-05):** `soulcap-audit` now keeps `contradicted`,
+`no_evidence_found` and `not_applicable` species statuses apart. Only
+`contradicted` raises a finding, and human-only markers on the mouse side are
+`not_applicable`, not failures. The mapping-audit skill uses the same
+vocabulary. 6 contradicted marker assertions are currently in the review
+queue ([audit_summary.md](reports/audit_summary.md)). This is still the
+evidence half of M3; the blocker above is unchanged.
+
 ---
 
 ## Milestone 4 — Candidate improved CL mappings 🟡 (80/127 entities with proposals)
@@ -189,6 +209,14 @@ proposal. The SSSOM export is generated from this registry. All proposals are
 `needs_review`, not confirmed equivalences. The gap log includes family-level
 entries and is not a one-to-one accounting of the 47 entities without proposals.
 
+**CL structure proposal (2026-10-05, design only):**
+[docs/group1_ilc_species_hierarchy_proposal.md](docs/group1_ilc_species_hierarchy_proposal.md)
+proposes a parallel human/mouse hierarchy under group 1 ILC → NK.
+- 6 new grouping terms, plus 2 recommended (uterine NK, human; group 1 ILC,
+  human = #14) and 1 optional.
+- Awaiting Dr. Diehl's decisions ([docs/diehl_review_2026-10.md](docs/diehl_review_2026-10.md)).
+  No CL edits are made here; approved changes go to David's team as issues.
+
 ---
 
 ## CL term corrections
@@ -201,7 +229,15 @@ upstream to `obophenotype/cell-ontology`.
 | Issue | Terms | Status |
 |-------|-------|--------|
 | [#13 NKT cell naming](https://github.com/Cellular-Semantics/SOULCAP_CL_Mapping/issues/13) | CL:4052055 → "CD56-positive T cell, human"; CL:0000814 → "iNKT cell" | Filed upstream: [CL#3663](https://github.com/obophenotype/cell-ontology/issues/3663), [CL#3664](https://github.com/obophenotype/cell-ontology/issues/3664) |
-| [#14 Missing human-specific ILC1 term](https://github.com/Cellular-Semantics/SOULCAP_CL_Mapping/issues/14) | CL:0001067 "group 1 innate lymphoid cell" — no `, human` variant, unlike group 2 (CL:0001081) and group 3 (CL:0001078) | Open — not yet filed upstream |
+| [#14 Missing human-specific ILC1 term](https://github.com/Cellular-Semantics/SOULCAP_CL_Mapping/issues/14) | CL:0001067 "group 1 innate lymphoid cell" — no `, human` variant, unlike group 2 (CL:0001081) and group 3 (CL:0001078) | Open — not yet filed upstream. Recommended as NEW:8 in the group 1 ILC hierarchy proposal |
+| B1 "lacks HLA-DRA" on CD56-bright/dim NK (draft) | CL:0000938, CL:0000939 | Draft in [reports/gaps.tsv](reports/gaps.tsv), awaiting Dr. Diehl ([review](docs/diehl_review_2026-10.md)) |
+| B3a plasmablast "lacks CD138" (draft) | CL:0000980 | Draft, awaiting Dr. Diehl |
+| B3b plasmablast markers to a species child (draft) | CL:0000980 | Draft, awaiting Dr. Diehl (human vs mammal-level child) |
+| B6a–d smaller fixes (drafts) | CL:0002345 (DX5+); CL:0002348/0002349 (wrong PMID); CL:0002439–0002448 ("T cell" wording); CL:0002343 (galectin-1 relation) | Drafts, awaiting Dr. Diehl |
+| C1 liver-resident NK human sentences to a human child (draft) | CL:4047101 | Draft, awaiting Dr. Diehl |
+| Pre-NK markers (on hold) | CL:0000937 | On hold pending Dr. Diehl |
+| Protégé NCAM1 query returns too few classes | — | Not filed; repro note in [docs/protege_ncam1_query_repro.md](docs/protege_ncam1_query_repro.md) for Dr. Diehl to reproduce |
+| ~~Add CD56 to generic NK~~ | CL:0000623 | Withdrawn 2026-10-05 (generic NK keeps only lack statements) |
 
 ---
 
@@ -214,6 +250,14 @@ upstream to `obophenotype/cell-ontology`.
 - **Data hygiene.** The errors in `reports/marker_string_issues.md` (spaced
   marker names, unbalanced brackets, etc.) should be fixed in the Google Sheet
   before relying on automated parsing for M1/M3/M4.
+- **Matcher evaluation fingerprints break on Windows checkouts** (noted
+  2026-10-05, not fixed yet). `soulcap-evaluate` fingerprints source files by
+  hashing their raw bytes (`evaluation.py`). With `core.autocrlf=true`, a
+  Windows checkout has CRLF line endings, so the hashes differ from a Linux
+  run and the dashboard shows the evaluation as `stale_or_unverified`. The
+  same problem in `mappings/*.tsv` was fixed with a `.gitattributes` LF rule.
+  Fix later: normalise line endings before hashing, or extend the LF rule to
+  the fingerprinted files.
 - **Folders to standardise:**
   - `marker_mappings/` — curated mapping CSVs (tracked).
   - `literature/` — M2 literature reports.

docs/README.md

@@ -9,6 +9,9 @@ pages hold the full detail.
 | [pipeline.md](pipeline.md) | Each pipeline stage and its command: sheet sync, CL→PRO axioms, candidate matching, OAK lexical search, SSSOM export, ROBOT QC |
 | [marker_resolution.md](marker_resolution.md) | Token → PRO resolution rules, policies, lexical cache |
 | [matching_and_evaluation.md](matching_and_evaluation.md) | Audit dashboard, matcher evaluation, regression triage and follow-up |
+| [diehl_review_2026-10.md](diehl_review_2026-10.md) | CL issue drafts from the October 2026 NK and plasmablast audit, for Dr. Diehl to approve before anything is filed |
+| [group1_ilc_species_hierarchy_proposal.md](group1_ilc_species_hierarchy_proposal.md) | Design proposal for a parallel human / mouse hierarchy under group 1 ILC and NK cell (Dr. Diehl, 2026-10-05): before/after trees, new grouping terms, terms to decide. Per-term changes in the matching [.tsv](group1_ilc_species_hierarchy_proposal.tsv) |
+| [protege_ncam1_query_repro.md](protege_ncam1_query_repro.md) | Repro note for the Protégé NCAM1 (CD56) query that returned too few classes (Dr. Diehl, 2026-10-05): expected classes, likely causes, what to record |
 | [planning/](planning/) | Sprint backlog and semester plan (hand-maintained); agenda for the [12 Oct 2026 meeting](planning/2026-10-12_meeting_summary.md); [human/mouse marker pilot summary](planning/2026-10-02_diehl_summary.md) for Dr. Diehl |
 
 Related specs that live at the top level: [MARKER_SYNTAX.md](../MARKER_SYNTAX.md)