Reviewed native spellings — 2026-10-09

Alpha.3 publication follow-up — 2026-10-09: these merged spellings are included in the verified sahajlipi@0.1.0-alpha.3 registry package and repository demo. Alpha.2 does not include this slice. The release record separates publication and fresh registry-installed checks from the original source-development measurements below. The later merged 34-form batch is outside alpha.3 and shipped in published alpha.4.

This follow-up adds twelve exact native-word preferences and the missing भाग्न alternative for bhagna, using the unchanged development references. It follows the earlier four-word update. The repository engine and Nepali-enabled browser fields use the same entries.

The new source ledger records the thirteen target IDs, frozen references, source checks on 2026-10-09, access limits and implementation order. These are source-assisted project preferences, with no independent human review or representative accuracy claim. Published usage attests a spelling in context; it does not establish the intended meaning of every isolated informal Roman input.

Exact preferences and evidence

The twelve single-reading entries use their existing frozen preferred references. The linked sources below were checked for the whole spelling; two checks are limited to indexed PDF excerpts, as recorded in the ledger.

Roman key	Alpha.2 output	New output	Source and access
imandar	इमन्दर	इमान्दार	Kathmandu Metropolitan City, full HTML text
sarasar	सरसर	सरासर	Nepal Kanoon Patrika, page 1654, full HTML text
sakos	सकोस	सकोस्	Kawasoti Agriculture, full HTML text, final virama
kathanak	कथनक	कथानक	Information and Broadcasting Department, full HTML text
arambha	अरम्भ	आरम्भ	Tourism Department, full HTML text
ekadhik	एकधिक	एकाधिक	IGNOU Nepali educational text, indexed excerpt; PDF not inspected
jaghanya	जघञ	जघन्य	Nepal Kanoon Patrika, page 10160, full HTML text
pukar	पुकर	पुकार	Nepal Police Aasara publication, indexed excerpt; PDF not inspected
niskanda	निस्कन्द	निस्कँदा	Makawanpur District Administration, full HTML text, chandrabindu
bora	बोर	बोरा	Nepal Kanoon Patrika, page 9177, full HTML text, sack sense
utthan	उत्थन	उत्थान	Nepal Kanoon Patrika, page 2438, full HTML text
samanjasya	समन्जस्य	सामञ्जस्य	Nepal Kanoon Patrika, page 2438, full HTML text, ञ्ज conjunct

bhagna keeps its existing displayed reading भग्न and now offers भाग्न as the second candidate. भग्न is attested in the Lumbini Public Service Commission syllabus, whose extracted PDF text was inspected; visual pages were not. The distinct infinitive भाग्न is attested in the Kawasoti Agriculture article, inspected as HTML text. The frozen reference order remains [भाग्न, भग्न] with no preferred reference. Engine order [भग्न, भाग्न] preserves the existing default and makes no frequency or unique-meaning claim.

Typing and customization

Use the complete listed lowercase key. Incidental capitals such as Imandar and Kathanak follow existing normalization; reserved Shift sounds remain explicit. For example, Sarasar, niSkanda and kaThanak retain their phonetic readings. samanjasya supplies the reviewed ञ्ज cluster as a whole-word preference; it does not change the ordinary ny token.

These exact entries retain priority in both full and half consonant modes. Attached native forms such as imandarko use ordinary conversion unless the host supplies a complete custom entry. Vowel lengths, explicit nasal marks, backtick/slash halants, joiners, recognized loanword suffixes, protected addresses and literal English mode keep their existing behavior. The typing reference lists the original four keys and this alpha.3 spelling slice separately.

From this source checkout:

import { convertWord, convertText, createEngine } from './src/index.js';

convertWord('niskanda');
// { text: 'निस्कँदा', candidates: ['निस्कँदा'], ambiguous: false }

convertWord('bhagna');
// { text: 'भग्न', candidates: ['भग्न', 'भाग्न'], ambiguous: true }

convertText('imandar niskanda sakos.');
// 'इमान्दार निस्कँदा सकोस्.'

const custom = createEngine({ entries: { bhagna: ['भाग्न', 'भग्न'] } });
custom.convertWord('bhagna').text; // 'भाग्न'

In browser fields, finish bhagna and choose भाग्न from the candidate dropdown or with Alt+2 while the word is active. Space commits the displayed choice. The repository demo and marked fields use the same engine; English mode keeps subsequent Roman input literal.

Recorded comparison

The machine report compares the published alpha.2 source commit 48a0321bed417b1e995f66adbcbf9ade2c82f1cf with this source update, unpublished at the recorded comparison. Both engines are measured on the same final seed fixture and the unchanged reviewed development references. Engine, fixture, tool, original batch and ledger hashes identify the comparison.

Metric	Alpha.2 source	Source update before publication
Same final seed contracts	160/174	174/174
Historical seed contracts	160/160	160/160
Reviewed word default matches any reference	16/74	28/74
Reviewed word cases with any reference in candidates	16/74	28/74
Reviewed word cases with all listed references in candidates	15/74	28/74
Individual reviewed word-reference coverage	17/79	30/79
Preferred reviewed word default	14/69	26/69
Exact complete reviewed sentences	0/12	0/12

The final seed has 174 contracts and two exploratory cases. Thirteen word contracts and one synthetic text guard are appended; all 162 historical rows remain byte-for-byte unchanged. The historical fixture stays at 160/160 for both engines. The new cases protect the chosen software behavior, including candidate coverage for bhagna.

The reviewed denominator stays at 74 admitted words / 79 listed word references and 12 admitted sentences / 13 sentence references, with 14 excluded cases. Twelve defaults improve; the bhagna default already matched a valid reference, and its additional candidate improves coverage. The targets guided implementation, so these gains describe development progress. 46 admitted word defaults and all 12 complete sentences still differ. No new independently reviewed natural-typing sample or held-out accuracy result is added.

The ledger defers mahila, shanta and angrejharuko, whose listed alternatives do not specify a uniquely preferred default, plus contextual names and broader compounds. The roadmap tracks the remaining spelling work and independent typing evidence. Earlier dated measurements, original cases, review decisions and human review sheets retain their original contents.

Software validation

The original source-update validation on 2026-10-09, before alpha.3 release preparation, passed 345/345 Node tests, 174/174 seed contracts, 66/66 desktop browser checks across Chromium, Firefox and WebKit, and 11/11 website checks. The website build and public-link checks also passed. These dated checks retain their measured source identities; fresh checks of the versioned alpha.3 candidate are recorded separately in the historical preparation section.

A local unpublished tarball from that original source update passed standalone JavaScript and strict TypeScript NodeNext/Bundler consumers. Its installed engine passed all 13 new entry checks and a mixed-text integration check. This does not republish the npm alpha.2 artifact. These software checks do not establish physical mobile/IME compatibility, independent human spelling review or representative language accuracy.

Reproduce

Prepare the pinned original benchmark/data/review-batch-001/cases.jsonl using the review-batch instructions. Its SHA-256 must be 2f92bcf5a16d35b6a31dca052e4e358ffe417b5dcec0e8cc704d17f8acbf23e3. Preserve any existing directory containing edited review sheets.

From a checkout containing this update:

SPELLING_BASELINE_DIR=$(mktemp -d)
git archive 48a0321bed417b1e995f66adbcbf9ade2c82f1cf \
  src package.json benchmark/cases.jsonl benchmark/run.js benchmark/reports \
  | tar -x -C "$SPELLING_BASELINE_DIR"
node benchmark/measure-reviewed-spelling.js "$SPELLING_BASELINE_DIR" \
  --cases benchmark/data/review-batch-001/cases.jsonl \
  --output benchmark/data/nepali-spelling-current.json

npm test
npm run benchmark -- --check
npm run benchmark:source-reviewed

Choose a new output path on each measurement; the tool refuses to overwrite an existing report. The comparison checks frozen source identities and exact Unicode strings. The final source hashes in the machine report distinguish this run from a later checkout with the same package version. The current-engine source-review command reports remaining development mismatches without making them a regression gate.

This is a repository update prepared for review. Package publication and website deployment remain separate actions; the released alpha.2 artifact retains its recorded behavior.