文字背后，许多人的工作。

# Attributions

## Unicode Unihan character readings

- Author: Unicode, Inc.
- Source: https://www.unicode.org/Public/17.0.0/ucd/Unihan.zip (Unicode 17.0.0)
- License: Unicode License v3 (Unicode-3.0), https://www.unicode.org/license.txt

`Data/CharacterSource` retains the pinned kTGH, kMandarin, kHanyuPinyin,
kTGHZ2013 and kXHC1983 property values and source hashes. The independent
`Tools/BuildLexicon/build_characters.py` strips tone marks, merges all readings
and omits compatibility duplicates and syllabic readings that the existing
Xiaohe codec cannot encode. Common dictionary readings for these characters
are retained from the existing Rectfly source and amzxyz’s CC-BY-4.0 RIME-LMDG
`zi.dict.yaml`, commit `5568f955750ca4b58deaafd4a18e2c6cf62562c9`. The exact
reading supplement, upstream hash and CC-BY license are in `CharacterSource`. It produces `characters.tsv` and the
shared, mapped `characters.rflx`, ordered by the Wanxiang weights described below
and filtered by iOS system-font coverage. This source uses no Gboard or rime-ice data.
The full copyright and permission notice is bundled in `Unicode-LICENSE.txt`
in both targets and displayed in the App’s 致谢 page.


## Wanxiang high-frequency core and exact-only cold words

- Authors: amzxyz and RIME-LMDG contributors.
- Optional exact-code packs: mingren, diming, renming, shici, wuzhong, yaopin,
  yixue, huaxue, duoyin, cuoyin, fangyan, taifeng and yiren, all from full commit
  `e5cb7dbebeacb5ddad63289640940105ef32cf45`. Per-file download and derived-data
  SHA256 values are in `Data/WordSource/source.json`; used copies in `Data/WordSource/Packs`.
  Missing source frequencies use 1; the core/cold frequency units are unchanged.
  Abbreviation/T9 tables and mixed-script input rules are not included.
- Source: https://github.com/amzxyz/RIME-LMDG/tree/e5cb7db
- License: CC BY 4.0, https://creativecommons.org/licenses/by/4.0/.
- Changes: tone removal, Xiaohe canonicalization, word/pronunciation deduplication
  by source maximum, omission of malformed/unencodable rows, frequency selection
  for the core and compact exact-code indexing for cold words. No frequency
  rescaling or extra coefficients.
- Public, unencrypted used-data copy: `Data/WordSource/wanxiang-used.tsv`.
  `Data/Wanxiang-LICENSE.txt` is bundled/displayed. A public source copy is not
  an ETM exemption; release requires the channel checks in `Data/WordSource/README.md`.
- Owner userdb records are a separate imported private layer, not in this copy.
- Release attachment: `wanxiang-used-e5cb7db.tar.gz`, built and byte-verified by
  `Tools/BuildLexicon/package_wanxiang_source.py`. Publish it as a separately
  downloadable, unauthenticated release asset. If this source repository or its
  releases are private, publish the attachment on a public release mirror before
  distributing the App. See `Data/WordSource/README.md` for the release procedure.

## wanxiang RIME-LMDG (dictionaries + grammar model)

- Author: amzxyz
- License: CC-BY-4.0
- URL: https://github.com/amzxyz/RIME-LMDG

`Data/CharacterSource/wanxiang-zi.dict.yaml` preserves the character dictionary
from commit `e5cb7db` (amzxyz, CC BY 4.0). The derived `characters.tsv` and
`characters.rflx` remove tone marks, convert readings to Flypy codes, merge
same-character/same-code weights by maximum, and exclude characters without
iOS 26.5 system-font glyphs. Missing weights are zero; supplied weights are unchanged.
The full license is in `Data/CharacterSource/Wanxiang-LICENSE.txt`.

The optional grammar download is the unchanged `wanxiang-lts-zh-hans.gram` from
[amzxyz/rime-wanxiang-data LTS](https://github.com/amzxyz/rime-wanxiang-data/releases/tag/LTS).
The October 19, 2026 asset has 398,497,836 bytes and SHA-256
`9eedac1d836f3fdafb61b162035c8937253aedc106be40b2b957d9ab64e9ed89`.
The pinned fallback `wanxiang-lts-20260909.gram` is the unchanged September 9, 2026 LTS
model (asset 552602424; 420,339,756 bytes; SHA-256
`1eb651288b117e68374110b32796346e128be790e994559cb7c1e3b0fd63029d`).
The App downloads the upstream asset (or pinned `ann61c/rectfly` fallback) into the App Group;
the keyboard and App preview map that same file. It is not bundled. Attribution and the CC BY 4.0 license link are included in both bundles'
`Octagram-LICENSE.txt`.

## librime-octagram (grammar format and query defaults)

- Author: lotem
- License: BSD-3-Clause
- URL: https://github.com/lotem/librime-octagram

The plugin's source (`gram_db.h`, `gram_db.cc`) documents the on-disk gram
format (Rime MappedFile header + darts-clone double-array trie). The offline
extractor and Swift runtime reader implement that published layout.
`Shared/OctagramLanguageModel.swift` adapts the basic-Han, non-rear word-query
defaults from commit `bfb168ca33d8b372596fdf2007933f3da1cf360e`.
Copyright (c) 2014, RIME Developers. The full BSD-3-Clause license is retained
in `Data/Octagram-LICENSE.txt` and included in the App and keyboard bundles.

## English frequency dictionary

`Data/english_words.tsv` derives from Wolf Garbe's SymSpell frequency dictionary,
revision `c239062ae02961df18ab7da1671d01b4388204e0`, filtered to ASCII letters and
apostrophes and sorted by word. The upstream truncated `you'v` is corrected to
`you've`; original frequencies are unchanged.

- Source: https://github.com/wolfgarbe/SymSpell/blob/c239062ae02961df18ab7da1671d01b4388204e0/SymSpell/frequency_dictionary_en_82_765.txt
- SymSpell: MIT, copyright 2018 Wolf Garbe.
- Google Books Ngram frequencies: CC BY 3.0, https://creativecommons.org/licenses/by/3.0/.
- SCOWL vocabulary: https://wordlist.aspell.net/scowl_v1-readme/.
- Required notices are bundled in `Data/English-LICENSE.txt` in both targets.
- Rebuild: `Tools/BuildLexicon/build_english.py`; pinned source and output SHA-256
  values are checked. No SymSpell implementation or runtime dependency is used.

## English word-pair frequencies

`Data/english_bigrams.tsv` derives from the same pinned SymSpell distribution's
`frequency_bigramdictionary_en_243_342.txt` (MIT repository license retained in
`English-LICENSE.txt`). Both words must occur in the bundled English vocabulary;
only the eight highest-frequency continuations per preceding word are retained.
Counts are unchanged; ties use alphabetical order.

`Data/english_context.bin` retains all 242,342 vocabulary-matched pairs from
that same source for spelling-candidate ranking. Its 80-byte header records
format version, vocabulary hash and row count; sorted 16-byte little-endian
records contain a pair of vocabulary IDs packed into a 64-bit key and the
unchanged 64-bit count. It contains no additional corpus or learned user data.
SHA-256: `3fa9d6d400d29dd770f35bc60fd14f9cd66b711e6c6ab8c6f76d0dd1f6d0cb15`.

- Source: https://github.com/wolfgarbe/SymSpell/blob/c239062ae02961df18ab7da1671d01b4388204e0/SymSpell/frequency_bigramdictionary_en_243_342.txt
- Source SHA-256: `fd892a160184101dd7ae807ac5a302d01fcea1c47304181a8ed7ed9c94545bcd`.
- Rebuild with `build_english.py --bigrams-source PATH`; source/output hashes
  and 64,097 retained rows are checked.

This is a separate word-pair table; the unigram corpus description above is
not evidence that the word-pair counts came from that same corpus.

## English regional spelling variants



`Data/english_variants.tsv` is a modified subset of [VarCon 2020.12.07](https://wordlist.aspell.net/varcon-readme/),
pinned to en-wl/wordlist revision `b22230cc5250887737fdefe9ca4c9d9d01230eaa`.
It adds 875 missing preferred/equal American and British spellings from verified
clusters through SCOWL level 60. Archaic usage rows and rare/improper variant
tags are excluded. Each spelling maps to an already covered form; when several
forms are covered, the highest existing frequency wins, with alphabetical ties.

The runtime borrows that form's frequency prior and context evidence. These are
not independently measured counts for the added spelling. The original vocabulary
and context resource remain byte-for-byte unchanged, including their word IDs.

- Source: https://github.com/en-wl/wordlist/blob/b22230cc5250887737fdefe9ca4c9d9d01230eaa/varcon/varcon.txt
- Source SHA-256: `75af63da46ec12d7eb14b9f1ba8d3898d484dd6872755b73c921b215875a3629`.
- Output: 18,631 bytes; SHA-256 `36f5a606d2ef2aa53030363814752c8793fdd36f5c334f3222c375bd24dba615`.
- Kevin Atkinson, Benjamin Titze and Geoff Kuenning notices are retained in the
  App and extension's `English-LICENSE.txt`. VarCon is an archived source; it is
  not presented as the current ESDB dataset.
- Rebuild: `Tools/BuildLexicon/build_english.py SOURCE --variants-source VARCON_SOURCE`.

## English sentence starters and modern word continuations

- Authors: Tatoeba contributors; all named sentence owners from the export are
  listed in `Data/English-LICENSE.txt`, bundled in both the App and keyboard.
- Source and export license: https://tatoeba.org/en/downloads
- Snapshot: complete English detailed export, October 3, 2026,
  https://downloads.tatoeba.org/exports/per_language/eng/eng_sentences_detailed.tsv.bz2
- License: CC BY 2.0 FR, https://creativecommons.org/licenses/by/2.0/fr/
- Modified: deduplicated sentences, removed a disjoint 10% evaluation holdout,
  extracted and bounded sentence-start, word-pair and three-word counts;
  sentence texts are not distributed. The derived tables are
  `Data/english_starters.tsv`, `english_prose_pairs.tsv` and
  `english_prose_trigrams.tsv`.
- Offline rebuild: `Tools/BuildLexicon/build_english_prose.py`; the compressed
  source SHA-256 is pinned in the script. App Settings → About → Acknowledgments
  exposes this attribution, license link, modifications and complete notices.

## English recorded word evaluation

`RectflyUIKitTests/Fixtures/ITE` contains modified, selected records from
[ITE Typing dataset v1](https://zenodo.org/records/12528163), published by
Katri Leino / Aalto University under [CC BY 4.0](RectflyUIKitTests/Fixtures/ITE/LICENSE).
Authors of the accompanying 2024 article: Katri Leino, Markku Laine, Mikko Kurimo
and Antti Oulasvirta. The fixture README and manifest record selection, provenance
and limitations. The data is included only in the UIKit test bundle; reference
words and other keyboards' outputs are evaluation labels, never decoder input.
## Emoji and annotations

The keyboard bundles Unicode Emoji 17.0 fully-qualified data and an emoji-only
subset of CLDR 48 Chinese/English annotations, under Unicode License v3.
Sources, rebuild instructions and the complete license are in
[`Data/Emoji/README.md`](Data/Emoji/README.md) and
[`Data/Emoji/UNICODE-LICENSE.txt`](Data/Emoji/UNICODE-LICENSE.txt).

## English screen-title display metadata

`Data/english_screen_titles.tsv` is separate display metadata derived from the
public [Wikidata Query Service](https://query.wikidata.org/) (2026-10-07 UTC).
It retains 1,362 existing vocabulary words that are single-word film/television
titles with at least 20 sitelinks. It adds no words or private user data.
Wikidata structured data is [CC0](https://www.wikidata.org/wiki/Wikidata:Licensing).
The pinned snapshot, exact query, source IDs, SHA-256 values and offline builder
are documented in `Tools/BuildLexicon/EnglishScreenTitles/README.md`.

## English proper names

`Data/english_names.rfen` uses Wikidata structured data under
[CC0 1.0](https://creativecommons.org/publicdomain/zero/1.0/), pinned to the
2026-10-08 query snapshot, the fixed 2026-10-14 technology query,
2026-10-15 subclass coverage, and 2026-10-22 app/website/software/brand query.
The last query prefers English labels and falls back to Wikidata multilingual
(`mul`) labels. All snapshots contain structured data only. The
project-authored public-name supplement is also released under CC0 1.0; it
contains names only, with no user input context. See
[the source and build manifest](Tools/BuildLexicon/EnglishNames/README.md).

## American vocabulary

`Data/english_us.tsv` and `Data/english_us.rfen` are project-authored vocabulary,
curated for the owner’s Rime configuration (`en_us.dict.yaml`, 2026-09-19).
The 480 entries cover chain brands, stores, banks, insurers, government
abbreviations and place names. Display spelling, punctuation and explicit
input codes are preserved; no third-party dictionary is used.
Rebuild with `python3 Tools/BuildLexicon/build_english_us.py`.

## Poetry next-line pack

`Data/poetry.rfpo` contains 823 ancient works (all 822 source records plus 易水歌) and 24,410 adjacent
clause mappings. The source revision, original SHA-256 checksums, inclusion rules,
reviewed readings and build instructions are in [the manifest](Tools/BuildLexicon/Poetry/sources.json)
and [documentation](docs/poetry-pack.md). Only ancient body text, titles and
authors are retained. Changes: full-corpus import, editorial-note removal, traditional to
simplified conversion, reviewed pinyin and the documented 洒→酒 typo correction.
The independently transcribed public-domain 《易水歌》 is from 《史记·刺客列传》.
The 3-work test fixture remains separate from product resources.

Upstream: [chinese-poetry/chinese-poetry](https://github.com/chinese-poetry/chinese-poetry/tree/b8594f81a89752241442f2ce267d6f66f96704ee).
[MIT license at the pinned revision](https://github.com/chinese-poetry/chinese-poetry/blob/b8594f81a89752241442f2ce267d6f66f96704ee/LICENSE):

```text
The MIT License (MIT)

Copyright (c) 2016 JackeyGao

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
```

### English canonical casing

`Data/english_casing.tsv` and `Data/english_casing_context.tsv` derive case-form counts from the pinned Tatoeba
English detailed export (2026-10-03; byte-identical download on 2026-10-09),
under [CC BY 2.0 FR](https://creativecommons.org/licenses/by/2.0/fr/).
Source: <https://downloads.tatoeba.org/exports/per_language/eng/eng_sentences_detailed.tsv.bz2>.
Attribution: Tatoeba contributors; the full existing owner list and notices
remain in `Data/English-LICENSE.txt`. Deduplicated sentence-internal counts
exclude the same 10% holdout and are reduced to spellings/ambiguity flags and high-support previous-word pairs.
No source sentences are shipped. See `docs/ENGLISH_CASE.md` for the policy.

### English offensive-word admission

`Data/english_offensive.tsv` derives from David Sojevic's
[dsojevic/profanity-list](https://github.com/dsojevic/profanity-list), under MIT
(2021 David Sojevic; full notice in `Data/English-LICENSE.txt` and
`Tools/BuildLexicon/Offensive/LICENSE`).

- Commit: `c27924319aa9bd6f917e3782b4f4b6604a50b652`.
- Source: <https://raw.githubusercontent.com/dsojevic/profanity-list/c27924319aa9bd6f917e3782b4f4b6604a50b652/en.json>.
- Source SHA-256: `bcf8bfc09dd3f7480972cb041a4cba291a6ff16a6fbf29022d5d679e3c39502c`.
- Vocabulary SHA-256 (`Data/english_words.tsv`): `a91457abe5d196228c3f351239e98fe1cba1f206661c75c4075a972b363c5199`.
- Generated SHA-256: `5338b2b7190b6e82a4cb0fcb7620b6a084699bd27ada63b9cb52f8e2e02b377d`.
- Rebuild: `python3 Tools/BuildLexicon/build_english_offensive.py`.

The existing English source resources have no usable offensive-word flag.
This source provides severity metadata, unlike a flat blocklist. Apply the same
rule to every entry: retain upstream severity 2–4, exclude severity 1 (mild).
Of 434 source entries, 406 are retained and 28 excluded. Expand explicit aliases,
then materialize repeated-letter forms and common suffixes (`s/es/ed/ing/er/ers`)
only when present in the pinned existing English vocabulary, excluding each
entry's upstream exceptions from its expanded forms (`*` substitutes the matched
word). Explicit aliases remain intact. This removes 12 expanded forms, including
`spices`, `spiced`, `spicer`, `spicing`, `dicker`, `cocker` and `cockers`, and produces
756 unique whole words/phrases. Runtime matching uses whole tokens and phrases,
with punctuation (including straight/curly apostrophes) separating tokens, never
arbitrary substrings; `class`, `shell`, `Scunthorpe`, `cockpit` and `Dickinson`
remain available.

Ambiguity is not resolved by guessing intent or handpicking exceptions. Ordinary
or polysemous entries include `ass`, `bloody`, `Christ`, `hell`, `Jesus`, `snatch`,
`breasts`, `genitals`, `nipples`, `penis`, `rectum`, `semen`, `vagina` (excluded
by severity); `anal`, `bareback`, `bastard`, `bitch`, `chink`, `cock`, `collaring`,
`cornhole`, `Dick`, `domination`, `dyke`, `fag`, `fingering`, `grope`, `haji`,
`intercourse`, `jelly donut`, `lovemaking`, `nude`, `pansy`, `panties`, `pegging`,
`playboy`, `poof`, `rape`, `retard`, `scissoring`, `sex`, `shrimping`, `sissy`,
`skeet`, `snowballing`, `spastic`, `spunk`, `switch hitter`, `threesome`, `topless`,
`twink`, `vibrator`, `wet dream`, `xxx`, plus identity terms `cishet`, `enby`,
`transbian` (retained by the same severity rule). This is an admission policy for
unsolicited suggestions, not a judgment about a word or its speaker. Fully typed
literal text stays available unchanged, including its case.

All 28 excluded entry IDs: `anus`, `ass`, `bloody`, `bollocks`, `boob`, `breasts`, `bugger`, `christ`, `crap`, `damn`, `dingleberry`, `genitals`, `hell`, `jap`, `jesus`, `kraut`, `mayonnaise-monkey`, `nipples`, `penis`, `pubes`, `rectum`, `semen`, `snatch`, `sod-off`, `sodding`, `vagina`, `venus-mound`, `wog`.

Upstream severities are subjective; its README notes that many original entries
defaulted to severity 3. Consequently ordinary senses can lose suggestions, and
some offensive senses of mild entries remain. This finite resource also does not
cover every new spelling or phrase. The user-facing switch restores the original
candidate behavior completely. No ranking scores or evaluation scoring change.

## English acronym statistics

`Data/english_acronym_casing.tsv` derives from the
original Tatoeba English detailed export, retrieved 2026-10-16, under CC BY 2.0 FR
(https://creativecommons.org/licenses/by/2.0/fr/). Source:
https://downloads.tatoeba.org/exports/per_language/eng/eng_sentences_detailed.tsv.bz2
Compressed SHA-256: `1d8c40545b4af2c6531d1a8b64941fa9d0fa3ee5bcc11cb705c8d8c1fd930c84`.

The generator deduplicates normalized sentences, removes the deterministic 10%
holdout and segment-initial words, then selects repeated short abbreviation forms.
Training token counts remain in the private generation report; the bundled table
contains only keys and canonical forms.
No source sentences are bundled. Contributor attribution, including additional
export owners, is retained in `Data/English-LICENSE.txt` in both targets. Rebuild
with `Tools/BuildLexicon/build_english_acronyms.py`; see
`docs/evidence/english-acronyms.md` for thresholds and hashes.