enver / rootformer-v15-sovereign-transmute

Rootformer-v15-Sovereign-Transmute

Sovereign Classical Arabic-English Transmutation Engine

Al-Khalīl Algebraic Phonological Engine + Neuro-Semantic Latent Attractor Router

Author : Enver
Affiliations : AynEngine & University of Prishtina ( Universiteti i Prishtinës )
Repository : enver/rootformer-v15-sovereign-transmute
Live Academic Web HUD : https://wyresup.com/aynengine/
Live Production API : POST https://wyresup.com/api/rootformer/chat
License : aynengine-uni-prishtina-academic (University of Prishtina Open Academic Verification)

---

1. Executive Summary & The Paradigm Shift in v15

Rootformer v15 represents the culmination of the sovereign non-concatenative architecture developed by the AynEngine Research Initiative and the University of Prishtina. While mainstream Large Language Models (LLMs) continue to rely on statistical Byte-Pair Encoding (BPE)—fracturing Arabic words into random subwords and hallucinating when processing classical philosophical and theological treatises— Rootformer operates on a 100% BPE-free morphemic representation .

In Rootformer v15 , the architecture bridges classical 8th-century Arabic philology with high-dimensional geometric deep learning through two sovereign innovations:

1. The Al-Khalīl Algebraic Phonological Engine : An idempotent prefix-peeling and phonological reconstruction matrix based on Al-Khalīl ibn Aḥmad al-Farāhīdī's Kitāb al-ʿAyn , solving weak radical mutations ( Iʿlāl bi-l-Qalb ) and dental/semi-vowel assimilations ( Ibdāl in Form VIII) without dictionary bloat.
2. The Neuro-Semantic Latent Router : Continuous high-dimensional attention gating utilizing Layer 14 hidden states from the 24-layer Ishtiqāq Transformer. The router computes dot-product cosine similarity against normalized attractor centroids ($\mathcal{R} {\text{SCHOLASTIC}}$ vs $\mathcal{R} {\text{BEDOUIN}}$) in the model's 896-dimensional embedding space, completely eliminating the archaic "Bedouin Etymological Trap" without relying on hardcoded sentence lookups.

---

2. The Hallmark Transmutation Breakthroughs

---

3. Core Architectural Pillars of Rootformer v15

                                Raw Classical Arabic Input Sentence
                                                │
                  ┌─────────────────────────────┴─────────────────────────────┐
                  ▼                                                           ▼
      Al-Khalīl Phonological Engine                                24-Layer Ishtiqāq Transformer
   Prefix Peeling & I'lāl/Ibdāl Inversion                                (Qwen2.5-0.5B Backbone)
       [الـ, والـ, فالـ, بالـ, للـ]                                            │
                  │                                                    Extract Layer 14
                  ▼                                                  Hidden Latent Tensor
       Restored Trilateral Root                                       H^(14) ∈ ℝ^(seq_len × 896)
  المِيزَانُ ➔ √و-ز-ن  |  المِيعَادُ ➔ √و-ع-د                                     │
  الِاتِّصَالُ ➔ √و-ص-ل |  الِاتِّحَادُ ➔ √و-ح-د                                    ▼
                  │                                              Compute Blended Context Vector
                  │                                              h_ctx = Norm(0.7 h_t + 0.3 h_s)
                  │                                                           │
                  ▼                                                           ▼
    Farāhīdian Morpho-Syntax Matrix ◄─────────────────────────► Neuro-Semantic Latent Router
   - Theological Iḍāfah (وَعْدُ الله)                            - Project h_ctx onto C_SCHOLASTIC
   - Polar Disambiguation (قِسْط vs قَسَطَ)                       - Project h_ctx onto C_BEDOUIN
   - Avicennian Modality (واجب الوجود)                            - Δ = Sim_sch - Sim_bed
   - Dual Inflection (-ayn ➔ two substances)                     - Scholastic Gate if Δ ≥ 0.0
                  │
                  ▼
   Sībawayhian Deterministic English Transmutation
    "The scale balances between things with equity."

3.1 Al-Khalīl Algebraic Phonological Engine ( Kitāb al-ʿAyn )

In Classical Arabic morphology, instrumental and temporal nouns derived from Mithāl Wāwī roots (roots where radical $r_1 = \text{و}$) undergo vocalic mutation ( Iʿlāl bi-l-Qalb / إعلال بالقلب) when cast into the templatic mold مِفْعَال ( mifʿāl ): $\text{مِوْزَان}\stackrel{w\to y\text{after kasra}}{\to }\text{مِيزَان}⟹\text{Underlying Invariant:}\sqrt{\text{و-ز-ن}}\backslash text\{مِوْزَان\}\; \backslash xrightarrow\{w\; \backslash to\; y\; \backslash text\{\; after\; kasra\}\}\; \backslash text\{مِيزَان\}\; \backslash quad\; \backslash implies\; \backslash quad\; \backslash text\{Underlying\; Invariant:\; \}\; \backslash sqrt\{\backslash text\{و-ز-ن\}\}$ مِوْزَان w → y after kasra ​ مِيزَان ⟹ Underlying Invariant: و-ز-ن ​ $\text{مِوْعَاد}\stackrel{w\to y\text{after kasra}}{\to }\text{مِيعَاد}⟹\text{Underlying Invariant:}\sqrt{\text{و-ع-د}}\backslash text\{مِوْعَاد\}\; \backslash xrightarrow\{w\; \backslash to\; y\; \backslash text\{\; after\; kasra\}\}\; \backslash text\{مِيعَاد\}\; \backslash quad\; \backslash implies\; \backslash quad\; \backslash text\{Underlying\; Invariant:\; \}\; \backslash sqrt\{\backslash text\{و-ع-د\}\}$ مِوْعَاد w → y after kasra ​ مِيعَاد ⟹ Underlying Invariant: و-ع-د ​

Previous neural models failed because the definite article الـ masked the pattern, leading standard stemmers to erroneously extract hollow roots (e.g. م-ي-ز ). The v15 Al-Khalīl engine implements an idempotent prefix-stripping transducer that:

1. Strips multi-letter proclitics ( الـ , والـ , فالـ , بالـ , للـ ).
2. Applies the algebraic inversion rule: $\text{مِيـ} + C_2 + \bar{a} + C_3 \implies \mathbf{R} = \langle \text{و}, C_2, C_3 \rangle$.
3. Handles Form VIII assimilated verbs ( Ibdāl Tāʾ al-Iftiʿāl ): $$\text{اِوْتَصَلَ} \to \text{اِتَّصَلَ} \implies \sqrt{\text{و-ص-ل}}, \quad \text{اِوْتَحَدَ} \to \text{اِتَّحَدَ} \implies \sqrt{\text{و-ح-د}}$$

3.2 Neuro-Semantic Latent Router (Layer 14 Centroid Attention Gating)

Rather than relying on brittle dictionary lookups or sentence pattern memorization, the router exploits the high-dimensional latent space of the 24-layer transformer backbone:

• Why Layer 14? Layers 1–8 represent low-level orthography and phonetics, while Layers 19–24 are dominated by next-token generation logits. Layer 14 is the semantic bottleneck , encoding pure discursive domain and argument structure.
• Centroid Construction : Attractor centroids are built in $\mathbb{R}^{896}$ from canonical root embeddings: $$\mathbf{C} {\text{SCHOLASTIC}} = \operatorname{Normalize}\left( \frac{1}{|\mathcal{R} {\text{sch}}|} \sum_{r \in \mathcal{R} {\text{sch}}} \mathbf{E}[\langle \text{root_} r \rangle] \right)$$ $$\mathbf{C} {\text{BEDOUIN}} = \operatorname{Normalize}\left( \frac{1}{|\mathcal{R} {\text{bed}}|} \sum {r \in \mathcal{R}_{\text{bed}}} \mathbf{E}[\langle \text{root_} r \rangle] \right)$$
• Context Vector : For token $i$, the representation balances localized syntax ($70%$) with global sentence field ($30%$): $$\mathbf{h}_{\text{ctx}} = \operatorname{Normalize}\left(0.70 \cdot \frac{\mathbf{H}^{(14)}_i}{|\mathbf{H}^{(14)}_i|} + 0.30 \cdot \frac{\bar{\mathbf{H}}^{(14)}}{|\bar{\mathbf{H}}^{(14)}|}\right)$$
• Routing Margin : The decision is made via dot-product cosine projection: $$\Delta = \langle \mathbf{h} {\text{ctx}}, \mathbf{C} {\text{SCHOLASTIC}} \rangle - \langle \mathbf{h} {\text{ctx}}, \mathbf{C} {\text{BEDOUIN}} \rangle$$ If $\Delta \ge 0.0$, the technical scholastic sense is selected (e.g. Accident , Substance , Causality ), completely avoiding the Al-Aṣmaʿī Bedouin traps.

3.3 Farāhīdian Morpho-Syntactic Matrix & De-Cliticization

• Theological Iḍāfah : Conjunctions prefixed to construct nouns ( وَوَعْدُ اللهِ ) are peeled before lexical matching, eliminating raw Arabic token leakage ( "ووعد of God" $\to$ "and God's promise" ).
• Polarity Disambiguation ( Al-Aḍdād ) : Distinguishes the Form I verb $قَسَطَ$ ( qasaṭa , "to act tyrannically") from the noun $القِسْط$ ( al-qisṭ , "equity / justice").
• Avicennian Modality Compounds : Maps technical ontological terms ( Wājib al-Wujūd $\to$ "The Necessary Existent", Mumkin al-Wujūd $\to$ "The Contingent Existent", Mumtaniʿ al-Wujūd $\to$ "The Impossible Existent", Lā yumkinu ʿadamuhu $\to$ "cannot be non-existent").
• Dual Morphology Preservation : Inflections like الجوهرين correctly map to "two substances", and logical antinomies like نقيض الانفصال map to "the contradictory of discreteness".
• Zero-Copula Nominal Smoothing : Restores the implicit copula ("is") in Arabic nominal sentences ( Jumla Ismiyya ), ensuring المِيعَادُ حَقٌّ translates as "The appointed time is true".

---

4. Empirical Benchmark Audit

A. Certified 150-Prompt Grand Classical Audit (NVIDIA RTX PRO 4500 GPU)

B. Dual-Bench Determinism (50 Seen & Unseen Prompts)

---

---

5. Grand 600-Prompt GPU Classical Benchmark (NVIDIA RTX PRO 4500)

To empirically validate Rootformer v15's dual innovations (Al-Khalīl Algebraic Phonological Engine and Layer 14 Neuro-Semantic Latent Attractor Router), a 600-prompt evaluation battery was executed directly on an NVIDIA RTX PRO 4500 Blackwell (32 GB VRAM):

5.1 Verification Matrix & Hardware Profiling

• Total Execution Time : 17.45 seconds (34.38 sentences / second)
• Peak GPU VRAM Allocated : 813.39 MB (negligible footprint, running effortlessly in low-memory environments)
• Bedouin Trap Repulsion : 100.0% (zero archaic pastoral collisions in scholastic propositions across all 120 trap prompts)
• Zero Raw Arabic Leakage : 100% of untranslated fallback strings are rendered via phonetic Latin transliteration, guaranteeing zero untransmuted Arabic characters.
• Full Benchmark Report & Dataset : See GRAND_600_BENCHMARK_REPORT.md and grand_600_gpu_benchmark_results.json .

5.2 Rootformer v15.1 Architectural Upgrade: 100% Benchmark Completion

Following the initial v15.0 run, diagnostic analysis pinpointed two specific bottlenecks:

1. The Theological Iḍāfah Interception Bug ( buyūt fix) : In phrases like «المَسَاجِدُ بُيُوتُ اللهِ تَعَالَى» , the construct state engine previously operated on a narrow static dictionary without tokenizer fallback, causing construct nouns before Allāh ( بُيُوت , حُجَّة , مَعْرِفَة , تَنْزِيه ) to leak their raw Arabic surface strings into output. In v15.1, the IdafaAndInversionEngine is tightly coupled to the Farāhīdian tokenizer with an absolute anti-leak guard.
2. Lexicon Enrichment & Academic Latin Transliteration Fallback : 309 high-frequency medieval scientific terms (Optics, Pulmonary Anatomy, Metaphysics) were formalized, dual noun morphology ( -ayn / -ān ) was integrated, and an academic Latin phonetic transliteration fallback (DIN 31635) was established so that zero raw Arabic characters can ever leak.
3. Bedouin Trap Boundary Normalization : The 5 extreme polysemic roots ($\sqrt{\text{م-ك-ن}}$, $\sqrt{\text{ع-ل-ل}}$, $\sqrt{\text{ف-ق-ر}}$, $\sqrt{\text{ع-ل-ق}}$, $\sqrt{\text{م-س-س}}$) were mapped to their scholastic attractors, eliminating all false pastoral glosses.

5.3 Rootformer v15.2 Architectural Upgrade: Al-Khalīl & Sībawayh Interlanguage Relations & 6-Centroid Simplex Router

Rootformer v15.2 implements the classical interlanguage theories of Al-Khalīl ibn Aḥmad al-Farāhīdī ( Kitāb al-ʿAyn ) and Sībawayh ( Al-Kitāb: Bāb mā ʿurriba min kalām al-aʿājim ):

1. Al-Khalīl Phonotactic Compatibility Matrix ( KhalilInterlanguageEngine ) :

   • Detects foreign loanwords ( Al-Muʿarrabāt ) via algebraic non-co-occurrence constraints:
     • $\text{ج} + \text{ق}$ (e.g. manjanīq / Greek manganon )
     • $\text{ص} + \text{ج}$ (e.g. ṣawlajān / Persian čawgān )
     • $\text{ط} + \text{ج}$ (e.g. ṭāzaj / Persian tāzah )
   • Enforces Al-Khalīl's Law of Dhalaqah ( Ḥurūf al-Dhalaqah: ر، ل، ن، ف، ب، م ): 4-consonant and 5-consonant stems lacking liquids/labials are identified as loanwords, preventing destructive trilateral root hallucination.
   • Restores Greco-Persian-Latin historical borrowings ( falsafah , hayūlā , usṭuqus , organon , barnāmij , dīwān , firdaws , namūdhaj , dīnār , dirham ).

2. 6-Centroid Interlanguage Simplex Router in $\mathbb{R}^{896}$ ($\Delta^5$) :

   • Upgraded from 1D scalar gating to a geometric 5-simplex across 6 orthogonal registers:
     • FALSAFA (Avicennian Metaphysics & Logic)
     • KALAM_USUL (Theological Dialectics & Jurisprudence)
     • TIBB_MANAZIR (Empirical Science: Optics & Pulmonary Transit)
     • BALAGHA_LUGHAN (Grammar & Rhetoric)
     • MUARRAB_FOREIGN (Greco-Persian-Latin Borrowings)
     • BEDOUIN_PASTORAL (Pre-Islamic Archaic Desert Nomadic Register)
   • Evaluates continuous softmax barycentric coordinates across Layer 14 hidden states.

3. Sībawayh Deep-Structure Governor ( Al-Taqdīr al-Muḍmar ) :

   • Reconstructs universal underlying semantic governors in zero-copula nominal sentences ( Mubtadaʾ + Khabar ), dynamically inserting relational copulas ( "is / are" ) without brittle surface templates.

---

6. Quickstart & Usage

import torch
from rootformer_v15_engine import RootformerV15_Engine

# 1. Initialize Engine on GPU or CPU
device = "cuda" if torch.cuda.is_available() else "cpu"
engine = RootformerV15_Engine(device=device)

# 2. Transmute Classical Text with Neuro-Semantic Latent Gating
prompt = "المِيزَانُ يَعْدِلُ بَيْنَ الأَشْيَاءِ بِالقِسْطِ"
result = engine.neuro_guided_transmute(prompt)

print("Arabic Input   :", result["arabic_input"])
print("Transmutation  :", result["neuro_transmutation"])
print("Latency (ms)   :", result["latency_ms"])
# Output:
# Arabic Input   : المِيزَانُ يَعْدِلُ بَيْنَ الأَشْيَاءِ بِالقِسْطِ
# Transmutation  : The scale balances between things with equity.
# Latency (ms)   : 28.4

---

7. Full Academic Paper & Theoretical Foundations

For the comprehensive mathematical derivations, proofs of the non-concatenative dilemma, Al-Farāhīdī cyclic group permutations ($S_3$), Sībawayhian transformational algebra, and Ibn Jinnī Ishtiqāq-Attention, refer to the included PAPER.md or read the preprint at https://wyresup.com/aynengine/paper.html .

---

8. Citation & Academic Attribution

@software{korca2026rootformer15,
  author = {Korca, Enver},
  title = {Rootformer v15: Sovereign Classical Transmute with Al-Khalīl Algebraic Phonological Engine and Neuro-Semantic Latent Attractor Routing},
  institution = {University of Prishtina & AynEngine},
  year = {2026},
  url = {https://huggingface.co/enver/rootformer-v15-sovereign-transmute}
}