PhonoSemantics

Official code repository for:

Kumar, A. (2026). Phonosemantic Grounding: Sanskrit as a Formalized Case of Motivated Sign Structure for Interpretable AI (Version 2.0). Zenodo. DOI: 10.5281/zenodo.19564026

Kumar, A. (2026). DDIN: Devavāṇī-Derived Interpretable Network — Sequential Neural ODEs for Phonosemantic Grounding. arXiv:2604.XXXXX.

Kumar, A. (2026). PhonosemanticMeta Benchmark: Evaluating AI Vector Grounding via the Phonosemantic Manifold. Kaggle Competition + arXiv.

---

What this is

This repository contains all experimental code, the phonosemantic embedding layer, and the figure generation script associated with the papers above.

The research program explores two complementary tracks:

1. Phonosemantic Grounding: The paper proposes that the articulatory anatomy of speech production — the five loci of the vocal tract, manner of constriction, phonation type, and somatic resonance — constitutes a physically real coordinate system for AI semantic representations. Sanskrit, with its rigorously formalized phonological system, serves as the proof-of-concept case.

2. DDIN (Devavāṇī-Derived Interpretable Network): A parallel research track exploring whether sequential phoneme encoding through neural ODEs can extract semantic structure from Sanskrit verbal roots without backpropagation or dense connectivity. The "Receiver Model" (W=0) architecture proves that semantics can emerge from heterogeneous neuron physics alone.

---

Repository contents

The main repository contains the core scripts for Phase 1-3 validation. The /Experiments folder contains all exploratory tracking from Phase 4 (Formant Grounding), Phase 5 (Neuromorphic SNN constraints), and Phase 5B-6 (Sequential ODE + Architecture Ceiling).

File/Folder	What it does
phonosemantic_embedding.py	The core embedding layer. Replaces nn.Embedding with a 10-dimensional physically grounded coordinate system derived from articulatory anatomy. Importable module.
experiment_v3.py	Canonical root clustering experiment. 150 Sanskrit verbal roots, 5 phonological groups, 3 statistical tests (Wilcoxon, Mann-Whitney, multinomial classification). All semantic scores derived independently from Monier-Williams dictionary definitions. Reproduces the main results in Section 9 of the paper.
linear_probe.py	Linear probe comparison. Tests whether articulatory geometry carries semantic signal beyond phoneme identity alone. Phonosemantic coordinates vs one-hot PCA baseline vs random at equal dimensionality. Reproduces the +14pp result in Section 9.4.
blind_clustering.py	Blind clustering experiment. Automatic TF-IDF semantic clustering with permutation test — no manual axis labels. Reproduces the null result reported honestly in Section 9.5.
make_figure.py	Generates Figure 1: the phonosemantic manifold map (phonemes in locus × manner space with word trajectory overlays). Requires matplotlib.
phonosemantic_figure.png	Figure 1 from the paper — pre-generated for convenience.
/Experiments	Contains Phase 4 (Continuous Formant Grounding), Phase 5 (SNN simulations), and Phase 5B-6 (Sequential ODE breakthrough + architecture ceiling).
generate_submission_v21.py	Competition submission generator for Kaggle. Uses v21 sequential ODE model.
submission_v21.csv	Competition predictions: 42% accuracy, ARI = 0.0538

---

Installation

git clone https://github.com/HmbleCreator/PhonoSemantics.git
cd PhonoSemantics
pip install -r requirements.txt

---

Running the experiments

Run in this order — each script builds on the root data defined in experiment_v3.py:

# 1. Main clustering experiment (reproduces Section 9.1–9.3)
python experiment_v3.py

# 2. Linear probe (reproduces Section 9.4)
python linear_probe.py

# 3. Blind clustering with permutation test (reproduces Section 9.5)
python blind_clustering.py

# 4. Regenerate Figure 1
python make_figure.py

All scripts are self-contained and print results to stdout. No GPU required. Runs on CPU in under 2 minutes total.

---

Phase 4, 5 & 5B-6 Extensions (Version 2.0)

The Experiments/ directory contains the complete chronological chain of advanced empirical validation.

Phase 4: Continuous Formant Grounding

Script	What it does
ddin_exp15_formant_grounding.py	Continuous-Time Formant Grounding (Section 10). Employs real F1/F2 acoustic frequencies.
ddin_exp16_weighted_formant.py	Establishes the ARI = 0.0366 unsupervised baseline mapping.

Phase 5: Neuromorphic SNN Constraints

Script	What it does
ddin_exp17_snn_pynn.py	PyNN deployment. Identifies the Epileptiform Synchrony Limit (~500 Hz).
ddin_exp18_snn_inhibition.py	Tests if lateral inhibition prevents seizure state.
ddin_exp19_snn_wta.py	Global Winner-Take-All architecture.

Phase 5B-6: Sequential ODE Breakthrough & Architecture Ceiling

Script	Key Result
ddin_exp21_sequential_ode.py	ARI = 0.0591 — shatters static 0.037 ceiling (+61%)
ddin_exp22_grpo.py	GRPO with discrete ARI reward — flat gradient
ddin_exp24_grpo_supervised.py	Supervised centroid reward — no improvement
ddin_exp28b_structured_w.py	Structured W initialization — best ARI = 0.0592
ddin_exp29_two_layer.py	Two-layer hierarchy — +0.017 delta
ddin_exp30_contrastive.py	Contrastive GRPO — ARI = 0.0690
ddin_exp31_supervised_centroid.py	Supervised on Layer 2 — ARI = 0.0690

To reproduce the key experiments:

cd Experiments
python ddin_exp21_sequential_ode.py   # Best model: ARI = 0.0591
python ddin_exp29_two_layer.py        # Two-layer baseline

(Note: Phase 5 experiments require pyNN and Brian2 backend.)

---

Using the embedding layer

from phonosemantic_embedding import PhonosemantikEmbedding

# Initialize
embed = PhonosemantikEmbedding()

# Get 10D coordinate for a phonological group
vec = embed.get_group_vector('LABIAL')   # returns np.array shape (10,)

# Get centroid trajectory for a root
centroid = embed.get_root_centroid('LABIAL')  # mean over phoneme vectors

# Full coordinate breakdown
print(embed.describe('THROAT'))
# → locus: [1,0,0,0,0,0], manner: 0.30, phonation: [0.5, 0.40], resonance: 1.0

---

The coordinate system

Every phoneme maps to a 10-dimensional vector:

φ(p) = ( ℓ(p), α(p), β(p), ρ(p) )

Dimension	Description	Size
D1 — Articulation locus	Where in the vocal tract (throat, palate, cerebral, dental, labial, nasal)	6D
D2 — Articulation manner	Degree of closure (0 = full stop, 1 = fully open)	1D
D3 — Phonation type	(voicing, breath force)	2D
D4 — Somatic resonance	Primary body region of proprioceptive feedback (R1–R5, spinal axis)	1D

Total: 10 dimensions, fixed by articulatory anatomy. No statistical learning required for the coordinate system itself.

---

Key results

Experiment	Result
Axis clustering (Test 2, between-group)	5/5 groups significant, all p < 0.001
Multinomial classification	41.3% vs 20% chance, p ≈ 10⁻¹⁴
Linear probe — phonosemantic geometry	63.3% ± 10.5%
Linear probe — phoneme identity baseline	49.3% ± 6.8%
Linear probe — geometry advantage	+14.0 percentage points, p < 0.001
Blind TF-IDF clustering	ARI = 0.007, p = 0.143 (not significant — reported in full)
Phase 4: Acoustic Formant Grounding (Sec 10)	ARI = 0.0366 via unsupervised F1/F2 continuous clustering.
Phase 5: SNN Validation Limit (Sec 11)	Epileptiform Synchrony Limit identified at ~500 Hz when modeling temporal context extension.
Phase 5B: Sequential ODE (v21)	ARI = 0.0591 — sequential phoneme encoding via Neural ODE shatters static ceiling (+61% improvement).
Phase 5B-6: GRPO & Architecture (v22-v31)	All reward variants converge to ~0.06 ARI ceiling. Two-layer hierarchy adds +0.017 delta.
Kaggle Competition Submission	Accuracy: 42%, ARI: 0.0538 (v21 model) — submitted April 2026

Research Papers

Paper 1: Phonosemantic Grounding (Zenodo / CC BY 4.0)

Kumar, A. (2026). Phonosemantic Grounding: Sanskrit as a Formalized Case of Motivated Sign Structure for Interpretable AI (Version 2.0). Zenodo. DOI: 10.5281/zenodo.19564026

Core claim: The articulatory anatomy of speech production (locus, manner, phonation, resonance) constitutes a physically grounded coordinate system for AI semantic representations — no statistical learning required for the coordinate system itself.

Key results:

• 41.3% multinomial classification (vs 20% chance)
• +14pp linear probe advantage over phoneme identity baseline
• ARI = 0.0366 via F1/F2 formant clustering (Phase 4)
• Epileptiform Synchrony Limit at ~500 Hz (Phase 5 SNN)

Paper 2: DDIN — Sequential Neural ODEs for Phonosemantic Grounding (arXiv)

Kumar, A. (2026). DDIN: Devavāṇī-Derived Interpretable Network — Sequential Neural ODEs for Phonosemantic Grounding. arXiv:2604.XXXXX.

Core claim: Sequential phoneme encoding through neural ODEs can extract semantic structure from Sanskrit verbal roots without backpropagation or dense recurrent connectivity.

Key results:

• ARI = 0.0591 (v21 sequential ODE) — +61% over static baseline
• Receiver Model validated: W=0 (no recurrent weights) achieves ~0.06 ARI
• Two-layer hierarchy adds +0.017 delta
• All optimization strategies (GRPO, supervised, contrastive) converge to ~0.06 ceiling
• Precise architecture ceiling measurement for this task

Functional claim (publication-ready):

A zero-weight, 128-neuron sequential ODE reservoir achieves ARI ~0.06 on unsupervised semantic clustering of Sanskrit roots — the practical ceiling for this architecture.

Paper 3: PhonosemanticMeta Benchmark — Exposing the Grounding Gap in Frontier AI (Kaggle)

Kumar, A. (2026). PhonosemanticMeta Benchmark: Evaluating AI Vector Grounding via the Phonosemantic Manifold. Kaggle Competition + arXiv.

Core claim: The benchmark evaluates whether frontier AI systems possess intrinsic semantic grounding by testing them against a physically real coordinate system (Locus × Manner × Phonation × Resonance). By enforcing a strict Vector Grounding Score (VGS) threshold, it empirically exposes the "Grounding Gap" in modern LLMs.

The 8 Benchmark Tasks:

Task	Name	What it tests	VGS Result
T1	Axis Prediction	Semantic clustering of 150 Sanskrit roots into 5 phenomenological axes	Pass
T2	Phonological Siblings	Same-locus root similarity vs cross-locus	Pass
T3	Fabricated Roots	Novel root prediction from phoneme patterns	Pass
T4	Cross-Locus Distance	Semantic distance across articulation loci	Fail (positional bias)
T5	Rule Generalization	Paninian rule application to held-out roots	Pass
T6	Trajectories	Mapping word trajectories to phenomenological arcs	Fail (95% overconfidence, all "Arc A")
T7	Triplets	Harmonic coherence: anchor-to-locus matching	Fail (0% confidence collapse)
T8	Phonation	Mapping breath force to motor-unit recruitment	Fail (0% confidence collapse)

Key finding: Tasks 7 & 8 produce 0% confidence collapses in frontier models (Gemini 2.5 Flash). The model guesses the correct physical answer but reports zero internal certainty — proving the absence of bodily grounding in statistical embeddings.

VGS Metric: VGS = (Is_Correct) × (Confidence ≥ 70%)

---

Phase 5B-6: Sequential ODE & Architecture Ceiling (2026-04-16)

The /Experiments folder now contains the complete Phase 5B-6 experimental chain documenting the transition from static embeddings to sequential Neural ODE processing.

Key Experiments

Script	Description	Key Result
ddin_exp21_sequential_ode.py	Sequential phoneme encoding through 128-neuron ODE reservoir. W=0 (Receiver Model).	ARI = 0.0591 — shatters static 0.037 ceiling
ddin_exp22_grpo.py	GRPO optimization on α (decay) topology with discrete ARI reward.	ARI = 0.0411 — flat gradient (discrete metric)
ddin_exp24_grpo_supervised.py	Supervised centroid reward on α+β.	ARI = 0.0538 — no improvement
ddin_exp28b_structured_w.py	Structured W initialization (Dhātu-like clusters).	Best: ARI = 0.0592 (10 clusters)
ddin_exp29_two_layer.py	Two-layer hierarchical architecture (Layer 1: phoneme encoder, Layer 2: semantic organizer).	ARI = 0.0492 (+0.017 delta)
ddin_exp30_contrastive.py	Contrastive GRPO on two-layer architecture.	ARI = 0.0690
ddin_exp31_supervised_centroid.py	Supervised centroid on Layer 2.	ARI = 0.0690

Key Findings

1. Sequential ODE breakthrough (v21): ARI = 0.0591 — a +61% improvement over static embedding baseline
2. The 0.06 ceiling: All reward variants (unsupervised, supervised, contrastive) converge to ~0.06 regardless of:
   • Number of layers (1 or 2)
   • Reward type
   • W structure (zero, random, structured)
3. Two-layer adds signal: Exp 29 showed +0.017 delta from hierarchy
4. The ceiling is a feature: It's a precise measurement of single-layer heterogeneous ODE reservoir capacity on this specific task — not a universal limitation

The Publication Case

The Phase 5-6 investigation produces a complete, precise measurement:

• Exact ceiling mapped from every direction
• Architecture capabilities precisely quantified
• Hierarchical processing adds signal but insufficient
• Clean, negative-but-precise result is scientifically valuable

Functional claim (ready for publication):

A zero-weight, 128-neuron sequential ODE reservoir achieves ARI ~0.06 on unsupervised semantic clustering of Sanskrit roots — the practical ceiling for this architecture.

---

Citation

@article{kumar2026phonosemanticmeta,
  author    = {Kumar, Amit},
  title     = {PhonosemanticMeta Benchmark: Evaluating AI Vector Grounding
               via the Phonosemantic Manifold},
  journal   = {Kaggle Competition + arXiv preprint},
  year      = {2026},
  eprint    = {2604.XXXXX},
  archivePrefix = {arXiv}
}

@article{kumar2026ddin,
  author    = {Kumar, Amit},
  title     = {DDIN: Devavāṇī-Derived Interpretable Network — Sequential Neural
               ODEs for Phonosemantic Grounding},
  journal   = {arXiv preprint},
  year      = {2026},
  eprint    = {2604.XXXXX},
  archivePrefix = {arXiv}
}

@misc{kumar2026phonosemantic,
  author    = {Kumar, Amit},
  title     = {Phonosemantic Grounding: {Sanskrit} as a Formalized Case
               of Motivated Sign Structure for Interpretable {AI}},
  year      = {2026},
  publisher = {Zenodo},
  version   = {2.0},
  doi       = {10.5281/zenodo.19564026},
  url       = {https://doi.org/10.5281/zenodo.19564026}
}

---

License

Code: MIT License Paper: CC BY 4.0

---

Author

Amit Kumar — Independent Researcher, Bihar, India GitHub: @HmbleCreator