Lab notebook for The Arabic Tongue: data audits, anomaly investigations, structural deltas. https://arabicjuthoor.com/feed.xml 2026-05-24T00:00:00Z Yassine Temessek © Temessek for Research, Publishing & Training https://arabicjuthoor.com/05-audits/lv1-overview-archived.md 2026-05-19T13:25:42Z 2026-05-19T13:25:42Z

**Project:** Juthoor Linguistic Genealogy **Module:** Juthoor-ArabicGenome-LV1 **Author:** Yassine Temessek

https://arabicjuthoor.com/05-audits/2026-05-24-aa-second-attack.md 2026-05-24T00:00:00Z 2026-05-24T00:00:00Z

**Date:** 2026-05-24 **Pool:** the 79 entries in the Afro-Asiatic roster marked PROVISIONAL or PARALLEL-STEMS (i.e. the ones the first pass filed as "different roots, can't claim a link") **Method:** re-attack each one with the full framework — *don't accept "different roots."* For each, ask: **is there a better Arabic word than the obvious one?** (the زنجبيل → زن+جبيل move, the مَجوس → فَعول move

https://arabicjuthoor.com/05-audits/2026-05-24-aa-round3-4-new-cognates.md 2026-05-24T00:00:00Z 2026-05-24T00:00:00Z

**Date:** 2026-05-24 (same-day continuation of the [second-attack audit](2026-05-24-aa-second-attack.md)) **What this round did:** (1) confirmed and promoted the tentative cracks from the second attack, (2) re-attacked more of the hard remainder, and (3) **mined the Egyptian lexicon for entirely new cognate pairs** the original 200-entry roster never contained. **Result:** Egyptian top-three-tier

https://arabicjuthoor.com/05-audits/2026-05-24-aa-pass2-audit.md 2026-05-24T00:00:00Z 2026-05-24T00:00:00Z

**Date:** 2026-05-24 **Pool:** the 84 PROVISIONAL entries in [`tier-a-afro-asiatic-cognates.md`](../04-cross-linguistic/tier-a-afro-asiatic-cognates.md) **Method:** the same blind-rescore methodology that took the Indo-European Pass-1 ≥0.65 pool of 770 candidates down to a calibrated 189 (see [`2026-05-21-opus-calibrated-770.md`](2026-05-21-opus-calibrated-770.md)), now applied to the AA Tier-A ro

https://arabicjuthoor.com/05-audits/2026-05-21-per-language-zscore.md 2026-05-21T00:00:00Z 2026-05-21T00:00:00Z

**Date:** 2026-05-21 · **Seed:** 20260521 · **Permutations:** 1000

https://arabicjuthoor.com/05-audits/2026-05-21-pass3-evaluation.md 2026-05-21T00:00:00Z 2026-05-21T00:00:00Z

The earlier sixteen-pair test confirmed the operative grammar is genuinely **generative at the binary level**: when given an unused two-letter combination, it can predict what Arabic *would* have meant there, and is right about half the time.

https://arabicjuthoor.com/05-audits/2026-05-21-opus-calibrated-770.md 2026-05-21T00:00:00Z 2026-05-21T00:00:00Z

Our first sweep produced 770 Arabic ↔ European cognates at score ≥ 0.65. A second look at a 150-pair sample showed the first pass was a little too generous near the borderline. Rather than adding a footnote, we **re-scored every one of the 770 pairs** under a stricter rubric — without showing the second rater the first score. The result: **189** still come out as likely (was 770), **120** as stron

https://arabicjuthoor.com/05-audits/2026-05-21-khshim-laws-audit.md 2026-05-21T00:00:00Z 2026-05-21T00:00:00Z

When Arabic and a European word share a meaning, their consonants rarely match letter-for-letter — they shift in predictable ways: ك becomes Q, ف becomes P, the throaty Arabic ع matches an H or a silent gap, and so on. Dr Ali Fahmi Khshim's *رحلة الكلمات* listed **nine such laws**. We took our 770 confirmed cognate pairs and counted how often each law actually fires in real data. The answer: all n

https://arabicjuthoor.com/05-audits/2026-05-21-irr-gemini-uncalibrated.md 2026-05-21T00:00:00Z 2026-05-21T00:00:00Z

**Date:** 2026-05-21 · **Sample size:** 150 pairs (stratified random from the 770-cognate pool at score ≥ 0.65)

https://arabicjuthoor.com/05-audits/2026-05-21-inter-rater-agreement.md 2026-05-21T00:00:00Z 2026-05-21T00:00:00Z

**Date:** 2026-05-21 · **Sample size:** 150 pairs (stratified random from the 770-cognate pool at score ≥ 0.65)

https://arabicjuthoor.com/05-audits/2026-05-20-tier-a-cross-language-audit.md 2026-05-20T00:00:00Z 2026-05-20T00:00:00Z

**Date:** 2026-05-20 **Source:** `Juthoor-CognateDiscovery-LV2/outputs/eye1_full_scale_*.jsonl` **Method:** for each of the 19 Tier-A Quranic-anchored Arabic roots, scan the per-language Eye-1 skeleton-match output for the highest-scoring target-lemma candidate. The skeleton match drops vowels and reduces both Arabic and target consonants to a comparable skeleton, then scores by Jaccard overlap an

https://arabicjuthoor.com/05-audits/2026-05-20-stat-test-robustness.md 2026-05-20T00:00:00Z 2026-05-20T00:00:00Z

**Date:** 2026-05-20 · **Seed:** 20260520 · **Permutations:** 1000 **Corpus:** 2285 records · **Split:** A=1142, B=1143

https://arabicjuthoor.com/05-audits/2026-05-20-pass2-evaluation.md 2026-05-20T00:00:00Z 2026-05-20T00:00:00Z

| # | Prediction | Pass condition | Actual | Verdict | |--:|-----------|---------------|--------|:------:| | P1 | ≥ 92 of 100 readable | ≥ 92 | **95 of 96 attested** (96/100 if you count the 2 dropped-as-unattested and 2 duplicates as un-readable; 95/96 = 99% of net new attested entries) | 🟢 **PASS** | | P2 | Path-mix within bands: Reduplication 15–45%, B-on-B 20–45%, T+A/stacked 20–50% | All thre

https://arabicjuthoor.com/05-audits/2026-05-20-cross-branch-eye2-audit.md 2026-05-20T00:00:00Z 2026-05-20T00:00:00Z

Earlier Tier-A work centred on Greek and Latin. The Eye-2 pipeline has since been run across **seven Indo-European branches** — covering Hellenic, Italic, two Celtic sub-branches, and all three Germanic sub-branches. This audit consolidates the results and asks: does the Arabic ↔ IE cognate signal survive outside Greek and Latin?

https://arabicjuthoor.com/05-audits/2026-05-20-coherent-gaps-held-out.md 2026-05-20T00:00:00Z 2026-05-20T00:00:00Z

Arabic could in principle make a root from any two letters out of 28. But ~140 two-letter pairs sit unused — no root in the standard lexicon starts with them. We picked **the 16 pairs the framework reads cleanly** and wrote down, in advance, what each one *should* mean if a root were ever built on it. Then we searched the classical dictionaries (Lisān al-ʿArab, Tāj al-ʿArūs) to see whether some fo

https://arabicjuthoor.com/05-audits/2026-05-19-statistical-predictive-test.md 2026-05-19T00:00:00Z 2026-05-19T00:00:00Z

Both were registered in [`docs/PROJECT_METRICS.md`](https://github.com/ustaz-dev/Juthoor-Linguistic-Genealogy/blob/main/docs/PROJECT_METRICS.md) **before** any statistical test was run — they are pre-registered, not post-hoc:

https://arabicjuthoor.com/05-audits/2026-05-19-non-quranic-extension-test.md 2026-05-19T00:00:00Z 2026-05-19T00:00:00Z

The eleven-mode operative grammar was developed and tested against Quran-anchored vocabulary. Does it also read **pre-Quranic** Arabic (Jahili poetry) and **post-Quranic** Arabic (modern MSA, including academy-coined neologisms)?

https://arabicjuthoor.com/05-audits/2026-05-11-jabal-data-anomalies.md 2026-05-11T00:00:00Z 2026-05-11T00:00:00Z

The Jabal lexicon contains 4 entries in the `binary_root` column that are not true binary nuclei. They are data-quality artefacts and should be filtered when computing nucleus statistics.

https://arabicjuthoor.com/05-audits/2026-05-10-not-earned-nuclei-deep-investigation.md 2026-05-10T00:00:00Z 2026-05-10T00:00:00Z

**Date:** 2026-05-10 **Audit item:** §7.1.1 follow-up — the 4 nuclei (`ثن`, `ضه`, `ظف`, `ون`) whose composed reading from consensus charges did not earn corroboration from Jabal's trilateral family. **Method:** manual reasoning (Claude Pass 2), grounded in the vault's consensus letter charges and Jabal's recorded family for each nucleus. The earlier codex pass diagnosed three of these as "low data

https://arabicjuthoor.com/05-audits/2026-05-09-7.2.5-fourteen-root-delta-resolved.md 2026-05-09T00:00:00Z 2026-05-09T00:00:00Z

**Date:** 2026-05-09 **Item closed:** [`our-contributions-and-roadmap.md`](../02-architecture/our-contributions-and-roadmap.md) §7.2.5 **Original framing:** «xlsx جبل (1,924 جذرًا) ومُخرَجات Codex (1,938 جذرًا = +14)» **Verdict:** original framing was **misleading**; both numbers were row counts, not unique-root counts. The real discrepancy is much larger and traces to two distinct bugs in Juthoor

https://arabicjuthoor.com/05-audits/2026-05-09-7.1.4-anbar-golden-rule-structural.md 2026-05-09T00:00:00Z 2026-05-09T00:00:00Z

**Date:** 2026-05-09 **Hypothesis tested:** if R = (a, b, c) is a trilateral root, the reverse R' = (c, b, a) carries an inverted or related meaning to R.

https://arabicjuthoor.com/05-audits/2026-05-09-7.1.4-anbar-golden-rule-graded.md 2026-05-09T00:00:00Z 2026-05-09T00:00:00Z

**Date:** 2026-05-09 **Hypothesis tested:** if R = (a, b, c) is a trilateral root, the reverse R' = (c, b, a) carries an inverted or related meaning to R. **Method:** structural pass over 1,955 trilateral roots in Jabal's lexicon → 155 distinct reverse-permutation pairs. Each pair graded against a 4-level scale: inverse / related / adjacent / unrelated.

https://arabicjuthoor.com/05-audits/2026-03-24-lv1-data-audit.md 2026-03-24T00:00:00Z 2026-03-24T00:00:00Z

**Date:** 2026-03-24 **Purpose:** Fact-check all claims about scholar data, letter coverage, and dataset counts against actual source files, NotebookLM extractions, and Jabal's xlsx. Reconcile discrepancies between this platform (Cowork/Claude) and the Codex platform.