test: the reading pass over test/h* - #731
Merged
Merged
Conversation
11 files read; changes in 8. haryanvi-alias, hebrewNeural and hakka are clean.
A HEADER THAT STATES THE OPPOSITE OF WHAT THE ENGINE DOES. hebrew.test.ts said
unvocalized restoration "is deferred". It is not: hebrew.ts:9 says "PHASE 2
(done, hebrewNeural.ts)", he-tagger.int8.onnx is committed, neuralRegistry.ts
wires it as "the NAKDAN", and test/hebrewNeural.test.ts asserts it restoring
שלום עולם → ʃalom ʔolam. Same class as afrikaans's "DEFERRED: … numbers" from
the a* pass — the header describing a state the file beside it disproves.
The same sentence was also unreadable, nesting three em-dashes: "(unvocalized
restoration — a neural nakdan — is deferred — the Arabic-diacritizer
analogue)".
TWO REFEREE SCORES THAT MEANT THE OPPOSITE OF WHAT THEY LOOKED LIKE, now
stated rather than left to inference:
- hmong reported "100% on-referee". The maps were FIT to that referee and RPA
is deterministic, so the figure is meaningless; the held-out 5-fold CV is
the only honest one. That disclosure was present but buried behind the
number.
- haitian reported 97.7% folded — while the eval backbone STRIPS the nasal
tilde, so the score says nothing about whether ã ɛ̃ ɔ̃ are right, which is
the whole nasal-vowel rule the file exists to test.
Also: haitian ended on the dangling "single source." that asturian and bavarian
both had; two chronology notes ("The 3rd creole in the fleet", "the fourth
language, and the first outside the Latin script"); hungarian's "This used to
assert `undefined` above 100" restated as why a range cap would be wrong; and
a hausa comment naming a trap ("the B.C.kafin dedupe") that its assertions do
not test — replaced with the one they do.
225 files / 3118 tests unchanged, typecheck clean. Comment-only.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
11 files read, changes in 8.
haryanvi-alias,hebrewNeuralandhakkaare clean.A header that states the opposite of what the engine does
hebrew.test.tssaid unvocalized restoration "is deferred". It isn't:src/languages/hebrew/hebrew.ts:9—PHASE 2 (done, hebrewNeural.ts): unvocalized restorationhe-tagger.int8.onnxis committedneuralRegistry.ts:34wires it asthe NAKDAN — restores niqqud on bare Hebrewtest/hebrewNeural.test.tsasserts it working:שלום עולם→ʃalom ʔolamSame class as
afrikaans'sDEFERRED: … numbersfrom thea*pass — a header describing a state that the file sitting next to it disproves.The same sentence was also unreadable, nesting three em-dashes: "(unvocalized restoration — a neural nakdan — is deferred — the Arabic-diacritizer analogue)".
Two referee scores that meant the opposite of what they looked like
Both had the caveat present but buried behind the number:
100% on-referee. The maps were fit to that referee and RPA is deterministic, so the figure is meaningless — the held-out 5-fold CV is the only honest generalisation. A reader skimming sees a perfect score.97.7% FOLDED— while the eval backbone strips the nasal tilde. The score therefore says nothing about whetherã ɛ̃ ɔ̃are right, which is precisely the nasal-vowel rule the file exists to test.Also
haitianended on the danglingsingle source.— the third instance, afterasturian(a*) andbavarian(b*).The 3rd creole in the fleet (after Kabuverdianu)andthe fourth language, and the first outside the Latin script.hungarian'sNO RANGE CAP any more. This used to assert undefined above 100restated as why a cap would be wrong: assertingundefinedabove a bound pins a table size, preserving a limitation rather than a behaviour.hausacomment naming a trap its assertions don't test (the B.C.kafin dedupe) — replaced with the one they do (kashimust not be doubled).Verification
Comment-only — zero non-comment lines changed. 225 files / 3118 tests passed, typecheck clean.
🤖 Generated with Claude Code
https://claude.ai/code/session_012Lc3WnUgogC7okV7n53vjr