Skip to content

Bundled words: the extraction tool, seven curated lists, attribution (GDR-0028) - #72

Open
ChaliceForAuri wants to merge 1 commit into
icloud-syncfrom
bundled-words
Open

ChaliceForAuri wants to merge 1 commit into
icloud-syncfrom
bundled-words

Conversation

@ChaliceForAuri

@ChaliceForAuri ChaliceForAuri commented Oct 19, 2026 •

Copy link
Copy Markdown
Owner

Stacked on #71. The content side of the word of the day (GDR-0027 property 4, GDR-0028), everything short of running the extract.

The pipeline

  • tools/extract-words.swift streams kaikki.org's English Wiktionary extract (one JSON object per line) from stdin, keeps only the terms the lists name, prefers a plain sense over an obsolete or rare one, takes the US pronunciation, cuts the origin at a sentence end, and writes packages/TimbreCore/Sources/TimbreCore/Resources/bundled-words.json in BundledList's own shape. It reports coverage per list and names every term it could not find. Proven on a sample dump.
  • Seven curated lists in packages/TimbreCore/Lists, 518 terms: general English (tier two), software and product, healthcare, law, finance, design, phrases and idioms. lists.json names each list's industries, which the profile matches.
  • BundledLists loads the resource once; every pick on the Mac, the iPhone and the Watch passes it. The resource ships empty until the extract runs, so the app shows no bundled word yet and nothing is promised in copy.
  • A bundled word you then say in a dictation joins your Library with the extract's meaning, labelled bundled, and is met. Tested.
  • Attribution: About the words in Settings › Library and in You; LICENSE-DATA.md carries CC BY-SA 4.0 for the data and the script, separate from the app's code (GDR-0005).

One honest measurement

The example writer's finance case ("accrual") is refused by Apple's on-device guardrail every time: Response may contain sensitive or unsafe content. The card shows no example rather than a bad one (GDR-0027). The corpus now marks the case expectRefusal, so the harness stays green and a draft appearing would be news.

To finish

Running the extract needs kaikki.org's English dump (a multi-gigabyte download). With your go-ahead I run it, commit the resulting resource, and the word of the day has 500 words to fall back on.

Verified

Core 100 tests; both apps and the site build; lint clean; sync:docs regenerated.

🤖 Generated with Claude Code

@vercel

vercel Bot commented Oct 19, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment
Project Deployment Actions Updated
timbre Ignored Ignored Preview Oct 21, 2026 8:59pm UTC

Request Review

…(GDR-0027, GDR-0028)

tools/extract-words.swift streams kaikki.org's English Wiktionary extract
from stdin, keeps the terms the lists name, picks the plain sense over
an obsolete or rare one, the US pronunciation, the origin cut at a
sentence end, and writes TimbreCore's bundled-words.json in BundledList's
own shape. It reports what each list found and names every term it did
not, so a gap is a curation decision. Run on our side, never in the app.

packages/TimbreCore/Lists: general English (163 tier-two words), software
and product (94), healthcare (51), law (55), finance (50), design (47),
phrases and idioms (58), with lists.json naming each list's industries.
The resource ships empty until the extract is run; the app then never
shows a bundled word. LICENSE-DATA.md carries CC BY-SA 4.0 for the data
and the script, separate from the app's code (GDR-0005).

BundledLists loads the resource once; every pick on the Mac, the iPhone
and the Watch passes it. A bundled word met in a dictation joins the
Library with the extract's meaning, labelled bundled, and is met
(tested). About the words, in Settings › Library and You, attributes
Wiktionary and the licence and says which lines are the model's.

The example writer's finance case is a refusal by Apple's guardrail
("Response may contain sensitive or unsafe content"), measured 0 of 5;
the corpus records it as expected, so a draft appearing would be news.

Core: 100 tests. Both apps and the site build. Lint clean.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@ChaliceForAuri ChaliceForAuri changed the title Bundled words: the extraction tool, seven curated lists, attribution (GDR-0026) Bundled words: the extraction tool, seven curated lists, attribution (GDR-0028) Oct 21, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant