Gita Tier-1 deep reader (chapter 2 proof): free dictionary-grounded hover+sound+etymology - #136
Merged
Conversation
…of of the free layer) The operator-chosen "Hybrid" plan's free layer, proven on BG chapter 2 (72 verses) at /gita/chapter/2: the deep interleaved reader with per-akshara sound + hover dictionary meanings + etymology mode, assembled mechanically from public-domain sources — NO LLM authored any gloss. Pipeline (chapter-parameterized, scales to all 18): - Sanskrit mūla: sa.wikisource (PD); the whitespace-token surface obeys the surface law by construction (2.50-2.72 matches the pilot bg2-source.ts byte-for-byte). - Padaccheda: sanskrit_parser (rule-based sandhi split, deterministic) — used only to find lemmata, never touches the surface. - Glosses: Monier-Williams via Cologne (PD), confidence-filtered — direct hits + clean ≤3-piece compounds kept; letter-name garbage, homograph stubs, and stem-mismatched particles BLANKED. No word shows a guessed meaning (99.9% naive "any-gloss" was a Goodhart trap; the filter makes "no gloss rather than a guess" literal). - Sound: the app's own devanagari romanizer — IAST = romanize(surface), so romanizationMatches holds by construction (100% gate pass, 0 guessed sounds). Coverage (this chapter): sound 100%, dictionary meaning ~70% confident / ~30% honestly blank. Content words genuinely useful; function words + Vedic-first MW sense-ordering are the visible mechanical wall (muni → "impulse") — exactly what the reserved Tier-2 alignment curation covers, and the famous 2.50-2.72 verses already carry curated pilot glosses. Route /gita/chapter/2 wired before the /gita catch-all; index card added. tsc clean; 32 gita tests + full liturgy suites green; build green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0117Ksuo6oeSBw4iTq3yzfYi
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
…g, lazy-loaded chunk P1 (wrong glosses on common words): the scorer picked any parse that hit Monier-Williams regardless of fit, so पार्थ mapped to a stray patronymic, एषा (pronoun "this") read as the verb-root "gliding, running", मुनि as "impulse". Added a curated OVERRIDE table (~40 particles/pronouns/names/ epithets) that takes precedence over the dictionary at both word and sandhi-piece level, and the scorer now counts override lemmas (so सञ्जय wins whole-word instead of splitting to san+jaya). Ambiguous non-override words prefer BLANK over a low-fit hit. Verified via the real module loader: पार्थ → "O son of Pṛthā — Arjuna", एषा → "this", केशव → "Keśava — Kṛṣṇa", न → "not". Coverage 70.1% → 83.5%, 0 garbage. P2 (cache poisoning): transient Cologne failures no longer write "" into mw-cache.json — only completed lookups cache, so a timeout retries. P2 (bundle bloat): the ~468KB chapter module is now code-split — the route is React.lazy + Suspense and the page dynamically imports the data. Build confirms chapter2-tier1 (260KB/40KB gz) + GitaChapter2Page (7KB) as separate chunks; the bg2t1 marker is absent from the main bundle. tsc clean; 7249 tests green; build green with the chunk split. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0117Ksuo6oeSBw4iTq3yzfYi
Owner
Author
|
Codex findings adjudicated & fixed: P1 — curated override table for particles/pronouns/names takes precedence over the dictionary (पार्थ, एषा, केशव, न all correct-or-blank now; coverage 70→83.5%, 0 garbage; verified via the real module loader); P2 — transient lookup failures no longer poison the cache; P2 — the 468KB chapter data is code-split out of the main bundle (separate chunks confirmed). 7249 tests green. |
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.
The Hybrid free layer proven on BG ch2 at /gita/chapter/2 — deep reader with per-akshara sound + Monier-Williams hover glosses + etymology, mechanically assembled from public-domain sources, no LLM-authored glosses. Sound 100% clean; dictionary meaning ~70% confident / ~30% honestly blank (no guessed meanings — resisted the 99.9% any-gloss Goodhart trap). For operator eyeball before scaling to all 18 chapters.
🤖 Generated with Claude Code
https://claude.ai/code/session_0117Ksuo6oeSBw4iTq3yzfYi