feat(ko): uniform plain-다 ending-monotony hot signal — KO×GPT catch 45→82.5% (5.2.0)#498
Merged
Conversation
Adds a KO-only per-paragraph deterministic hot signal (koreanEndingMonotony): fires when declarative -다 endings dominate (ratio >= 0.6, count >= 2) AND burstiness CV is below the low band AND the paragraph has >= 20 tokens. Unlike the standard burstiness trigger it does not require 3 sentences, so it catches short, length-uniform AI Korean the band gate skipped, while the -다 + low-CV conjuncts spare formal human Korean (varied lengths -> high CV) and conversational Korean (요/습니다), and the 20-token floor spares terse snippets. Implemented as a first-class signal (not the advisory koPostEditese payload), wired into the per-paragraph hot OR in src/features/index.js, mirrored in playground/analyzer.js for browser parity, and counted in rebaseline-score's trigger_counts. KO rebaseline (n=380): KO×GPT catch 45.0->82.5%, recall 59.2->70.8%, F1 0.644->0.716, FPR 12.8->14.0% (within published CI); EN unchanged. Documented in core/stylometry.md and SKILL.md with calibration.
Minor bump for the KO uniform plain-다 ending-monotony hot signal. KO-only deterministic stylometry addition; en/zh/ja byte-identical; FP within published tolerance. Syncs all version surfaces and adds the CHANGELOG 5.2.0 entry.
Contributor
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Adds a KO-only deterministic hot signal — uniform plain-다 register — that closes the biggest measured detection gap (KO×GPT recall 44%). Bumps to 5.2.0 (minor).
A paragraph is hot when declarative
-다endings dominate (ratio ≥ 0.6 and count ≥ 2) and sentence lengths are uniform (burstiness CV below the low band) and the paragraph has ≥ 20 tokens. Unlike the standard burstiness trigger it does not require 3 sentences, so it catches short, length-uniform AI Korean the band gate skipped — while the-다+ low-CV conjuncts spare formal human Korean (same-다, but varied sentence lengths → high CV) and conversational Korean (요/습니다), and the 20-token floor spares terse snippets.Implemented as a first-class signal (not the advisory
koPostEditesepayload), wired into the per-paragraph hot OR insrc/features/index.js, mirrored inplayground/analyzer.jsfor browser parity, and counted inrebaseline-scoretrigger counts.Measured impact (KO rebaseline manifest, n=380, deterministic analyzer)
15 new true positives for 3 new false positives (5:1), so precision rose too. EN/ZH/JA are byte-identical (KO-only). The frozen public claim manifests and the headline catch/FP claim are refreshed on the next dedicated rebaseline pass, not here.
How it was derived
Root cause: all 22 missed KO×GPT samples were single-paragraph 2–3 sentence snippets, and KO detection relied entirely on burstiness (which needs ≥3 sentences); MATTR/lexicon/ko-diagnostics fire ~0% for KO. The
-다-monotony + low-CV separator was found by length-matched comparison of missed-AI vs 250 human controls, and the 20-token floor + low-CV conjunct were added after the signal over-fired on terse toy fixtures.Verification
npm test— 787 pass / 0 fail (5 new ending-monotony unit tests incl. precision guards)npm run benchmark— 49-fixture suite still 100% (natural KO fixtures stay cold)npm run lint— syntax OK, cspell 0 issuesnpm run release:check— OK for 5.2.0npm run check:no-private-assets— OKDocumented in
core/stylometry.md(hot rule + calibration + failure mode) andSKILL.md.