Human review of the 859 auto-drafted construction decisions.csv rows

[rybesh]

Summary

sources/construction/decisions.csv was last hand-curated on 2026-07-08, against the candidate pool as it existed then. The 2026-07-23 rebase (#86, commit 47740f7) re-mined candidates.csv/occurrences.csv from the full per-record zip MARC — which now includes the non-held Z39.50 harvest — and that surfaced 859 heading clusters that had no decision.

Because a book only gets an ab:constructedUsing edge when one of its headings falls in an include=y cluster, this stale gap meant only 390 of the 3,567 books with candidate headings were linked. The join was correct; the curation was simply behind the data.

Commit 7c1a21b (on feat/construction-vocab, PR #103) fills the gap: it extends decisions.csv from 435 → 1,294 rows so every current candidate cluster now has a decision. The 859 new rows are an auto-generated best-guess DRAFT and need human review — that's what this issue tracks.

Verified rebuild impact of the draft as committed:

• construction-methods.ttl: 100 → 176 concepts (valid Turtle)
• ab:constructedUsing: 390 → 606 books, 980 edges

What the draft did

Each new row carries a note column explaining its auto-call. The bias is deliberate: ambiguous → include=n, because a wrong y silently adds a junk concept to the published SKOS scheme, whereas a wrong n just leaves a book unlinked (the status quo we're improving on).

Tier	note contains	Rows	Call
Reconcile to existing concept	reconciled to existing concept	113	y, reuses an existing conceptId
New construction concept	NEW … concept (best-guess)	87	y, new conceptId + guessed category
Genre/subject noise	genre/subject heading	227	n
Unclassified	unclassified	432	n

The columns you're reviewing

cluster,labels,include,conceptId,category,prefLabel,note

• cluster — the reconciliation key (a $0 authority URI, or <scheme>|<normalized-label>). Do not edit. It's the join key to occurrences.csv.
• labels — read-only: the distinct raw MARC heading strings in this cluster. This is what you're judging.
• include — y or n. The gate.
• conceptId — for y rows only: the SKOS concept this heading maps to. Reuse an existing id to merge variants across vocabularies; invent a new slug for a genuinely new concept.
• category — for y rows only. Must be exactly one of: material, technique, binding-format, production-printing.
• prefLabel — for y rows only: the human-facing label for a new concept. Ignored when conceptId reuses an existing concept (the existing prefLabel wins).
• note — the auto-rationale. Delete it once you've reviewed the row so remaining notes = remaining work.

How to review

Work tier by tier (grep the note column). Only the existing 435 rows are trusted; every row with an auto: note is a draft.

1. reconciled to existing concept (113, y) — quickest. Confirm the heading in labels really is the concept named by conceptId. Most are safe (screen prints. → screen-printing), but the plural/qualifier matcher can overreach — check the VERIFY ones. If wrong, either repoint conceptId or flip to n.

2. NEW … concept (best-guess) (87, y) — needs the most attention. For each, decide:

   • Is it actually a construction method/material/binding/printing technique? If not → n, clear the other columns.
   • Is category right (one of the four)?
   • Is conceptId/prefLabel clean, and does it duplicate an existing concept that should be merged instead? Known cases to fix: MARC typos Accordian/Acordion fold should fold into accordion-fold; long AAT parentheticals like Perfect binding (process need a clean label; borderline art-media headings (drawings, gouaches) were split — decide whether they belong.

3. genre/subject heading (227, n). Skim to confirm none is actually a construction method. These are things like artists' books, poetry, limited editions, exhibition catalogs — correctly excluded.

4. unclassified (432, n) — the long tail. Mostly genuine noise (foreign-language genre terms, short stories, containers), but this is where a real construction heading could hide, so it's worth a scan. Anything real → set include=y + conceptId + category (+ prefLabel if new).

To exclude a row: set include=n and blank out conceptId, category, prefLabel (matches the existing convention; the build only reads y rows).

Rebuild & sanity-check after editing

# regenerate the SKOS scheme + the per-book edges
make -C sources construction-methods.ttl
make -B graph/artists-books.ttl

# how many books are now linked?
tools/jena/bin/arq --data=graph/artists-books.ttl -q <<'SPARQL'
PREFIX ab: <https://dkglab.github.io/ns/artists-books/>
SELECT (COUNT(DISTINCT ?b) AS ?books) (COUNT(*) AS ?edges) WHERE { ?b ab:constructedUsing ?c }
SPARQL

# validate the scheme
tools/jena/bin/riot --validate sources/construction-methods.ttl

Integrity rules the build assumes (worth a final check): every y row has a non-empty conceptId, a prefLabel, and a category in the four-value set; every n row leaves those three blank.

Definition of done

Every auto: note has been reviewed and removed, and decisions.csv reflects deliberate human calls for all 1,294 clusters.

Follow-up to #103 (commit 7c1a21b).

[rybesh]

While wiring up the per-method pages for #103 (each construction method now links to a page listing the books that use it), the linking surfaced a batch of near-duplicate concepts that should be merged during this review. Because each surviving concept gets its own page, duplicates split one real method across several thin pages, so this is worth catching here.

Clusters found among the 176 concepts actually used by books (book counts in parens):

Clear duplicates — same method, spelling/label/suffix variants (should merge):

• Accordion fold — accordion-fold "Accordion fold format" (121) + accordian-fold-format-binding "Accordian fold format (Binding)" (1) + acordion-fold-format-binding "Acordion fold format (binding)" (1) — two are just misspellings.
• Handmade paper — handmade-paper "Handmade paper" (30) + paper-handmade "Paper, Handmade" (3) — label word-order variant.
• Non-adhesive bindings — non-adhesive-bindings (3) + non-adhesive-bindings-binding "Non-adhesive bindings (Binding)" (1) — redundant "(Binding)" suffix.
• Clamshell boxes — clamshell-boxes (1) + clamshell-boxes-binding "Clamshell boxes (Binding)" (1) — redundant suffix.
• Cotton fiber paper — cotton-fiber-papers "Cotton fiber papers" (1) + cotton-fibers-paper "Cotton fibers (Paper)" (1).

Judgment calls — possibly a material-vs-binding distinction worth keeping, or possibly a split to merge:

• Cloth — cloth "Cloth" (4, Materials) vs cloth-bindings "Cloth bindings" (2, Binding and format).
• Paper — paper "Paper" (15, Materials) vs paper-bindings-binding "Paper bindings (Binding)" (1, Binding and format).

For each cluster, the merge is a decisions.csv edit: point the variant clusters at the same concept id / preferred label so build-scheme.rq collapses them (the variants become skos:altLabels). The "(Binding)"-suffixed rows in particular look like auto-drafted duplicates of an already-curated concept.

This is only the set that at least one book links to today; the full decisions.csv review may turn up more among the un-used concepts.

[beeyop]

@rybesh all decisions are reviewed but not sanity checked. I have to stop for the day, but the branch construction-decision-update should have it all.

[beeyop]

@rybesh Construction methods are updated after reviewing the comparison to BARD. I have pushed the changes to the construction-decision-update branch. I am getting an error when I try to sanity check: bash: tools/jena/bin/riot: No such file or directory

[rybesh]

@beeyop Try running make tools/jena/bin/riot. If that fails, you may need to edit tools/jena/Makefile and edit line 4 to JENA_VERSION := 6.2.0 (they just released a new version).

[beeyop]

@rybesh Editing the makefile got me past step one but I'm hitting an error on step 2 of sanity checking now: No query string or query file