Every structured problem — handwriting, prime numbers, language, speech, software — has a grammar underneath it. The same five mathematical operations describe them all. NeuroNet builds AI that computes through named, readable operations instead of black-box weights — and works at every level of intelligence.
Standard neural networks are trained black boxes. Billions of floating-point numbers that somehow produce correct answers — with no explanation of why, no way to add knowledge without full retraining, no readable structure. We believe this is the wrong foundation for general intelligence.
"Models expressed as programs. Programs expressed as grammar. Grammar expressed as calculus."
— NeuroNet framework thesisBillions of unnamed weights. Black-box computation. Post-hoc explanation (LIME, SHAP). Full retraining for every knowledge update. No readable reasoning path.
Named mathematical operations. Readable reasoning paths. Intrinsic explainability. Grammar edits add new knowledge in one line. Zero black-box weights in the operations.
The key discovery: grammar is not just about language. The same structural rules that describe how Sanskrit words combine, how prime numbers are sieved, how agents compose actions — all reduce to the same five mathematical operations. Build AI on those five operations, and you get a system that is interpretable by construction.
Grammar is not unique to any one domain. The same structural symmetry — typed operations, composition rules, growing vocabulary, named roles — appears at every level of intelligence. We are building frameworks at each level.
Same mathematical core · Different application domain
↑ Five operations (Panchatattva) · Composition rules (Sandhi) · Growing grammar (Āgama) · Named roles (Kāraka) ↑
Grammar-typed neural operations for building and training AI models. Every weight is a named mathematical operation. Every decision has a readable reasoning path. Works for vision, mathematics, speech, language — any problem with structural grammar underneath. This is what we are building now.
Grammar-typed primitives for composing intelligent agents. Neuro, Flow, Plan, Memory, Effect, Budget — all typed, all composable, all following the same Sandhi composition rules. An agent program is a sentence in a formal grammar. Actively developed.
Grammar-typed rules for writing any software. The same five operations that describe neural computation and agent behaviour also describe general program structure — selection, flow, transformation, policy, and spatial context. This level is at the hypothesis stage — the mathematical foundation is there, the implementation is ahead. Research direction.
| Property | NeuroNet (L1) | NeuroLang (L2) | NeuroScript (L3) |
|---|---|---|---|
| Domain | AI model weights | Agent programs | Any software |
| Grammar unit | Akshara (math op) | Neuro (typed primitive) | Expression (typed rule) |
| Composition rule | Sandhi (layer ordering) | Flow / Plan | Syntactic grammar |
| Growth mechanism | Āgama (new op/proto) | New Neuro types | New grammar rules |
| Status | Active experiments | Deployed (Neos) | Research direction |
Every transformation in the universe — selection, flow, rate of change, optimal policy, spatial structure — belongs to one of five mathematical families. These map to the five classical elements (Panchatattva) found in multiple ancient traditions: Indian, Greek, Chinese, Japanese. The correspondence is not coincidental — it reflects a deep structure in how nature organises transformation.
This is not specific to Indian philosophy. Greek: Fire/Water/Earth/Air/Aether. Chinese: Wood/Fire/Earth/Metal/Water. Japanese: 五大 (godai): Chi/Sui/Ka/Fū/Kū. All five-element systems reflect the same underlying taxonomy of transformation types. The Panchatattva is the version we use because its mathematical mapping is most precise.
Classical regime (0D–2D) · Quantum boundary (3D) · Quantum regime (≥4D)
NeuroNet translates grammatical structure directly into model architecture. Every component has a Sanskrit name because every component maps precisely to a Sanskrit grammatical concept — not metaphorically, but structurally.
32 fixed mathematical operations — 6 Prithvi, 6 Jal (strides 1/2/4/8/16/32), 8 Agni (frequency bands), 6 Vayu, 6 Aakash. All matrices are frozen at initialization. Zero learned parameters in the operations themselves. Every operation has a name and a type.
A small learned selector that picks which operation to use for each input. Two variants: BornRouter (Linear projection, learned) and GrammarRouter (distance-to-prototype, zero weights). The model learns which operation, not what to compute.
Sanskrit Sandhi rules govern which sounds can combine. Our Sandhi constrains which operations are allowed in each layer. Layer 1 cannot use Prithvi (no selection before analysis). Layer 2 cannot use Aakash. Layer 3 allows everything. This forces coarse-to-fine processing.
Input (Vaikharī, raw pixels) → features (Madhyamā, 147-dim) → prototype space (Paśyantī) → class identity (Parā, argmax). This four-level compression maps exactly to Bhartṛhari's four levels of speech. Classification is compression running backward — from noise to essence.
We tested the framework on two completely different problems — visual pattern recognition and pure mathematics — to check if the grammar truly generalises. Both succeeded. Different operations dominated in each, confirming the framework adapts its grammar to the problem's true structure.
70,000 handwritten digits, 10 classes. The model learns to classify 0–9 using only named mathematical operations.
What the model discovered without being told:
Vayu (attention/gating) dominates — image recognition needs to gate which edge features to attend to. The model discovered this independently.
Version history:
Best accuracy. 117K params. HOG features (+2.83%) are the decisive ingredient.
Stage 1 = three pure math ops (threshold → pool → sobel). 33K params. No learned feature extraction.
26K params. Router replaced by distance-to-prototype — no learned routing weights at all. Dynamic prototype growth (āgama): 1→3 prototypes per class.
Given any integer, predict: prime or composite? No formula given. The model must discover the structure of primality from data — and the grammar grows as new primes are found.
What the model discovered — without being told:
Jal (finite differences) dominates at 83.7%. Prime number theory is fundamentally about how divisibility changes — gaps between primes, difference in residue patterns. The model found this structure independently.
How numbers are encoded (CRT fingerprint):
MNIST routes through Vayu (attention gating). Primes route through Jal (differences). Nobody told the model which to use — it discovered the mathematical nature of each problem independently. This is what "grammar as computation" means: the model finds the right calculus for the task, not just the right numerical weights.
Standard AI requires full retraining to incorporate new knowledge. NeuroNet does it the way a living grammar works — by adding one new rule. When the grammar grows, the model immediately starts using the new rule on the next input.
In Sanskrit grammar, āgama is the insertion of a new phoneme rule into the existing grammar to handle a new case. In NeuroNet, āgama is adding a new prime modulus, a new prototype slot, or a new operation variant. One insertion = one new rule. The grammar is always live.
Primes have infinitely many grammar rules — the sieve never ends. The NeuroNet grammar bank has 60 slots for prime mods; we've filled 15 so far. More rules → higher precision. The grammar grows as the world grows. This is how a living mind works, not a frozen model.
"Prime numbers have a fixed sort of operation with a growing grammar. If we solve this problem, we will know that we can do this for any problem."
— Gopal Kumar, May 2026The framework works wherever there is grammar — and every sufficiently structured domain has grammar. Language, mathematics, code, music, physics, logistics — all have typed primitives, composition rules, and hierarchical structure. NeuroNet can be applied to any of them.
Audio is waves. Agni (QFT) for frequency bands — the spectral structure of vowels and consonants. Jal (differences) for formant transitions between phonemes. Sandhi rules for which phoneme combinations are permitted. Grammar = phonological rules of the language.
The inverse of ASR. Parā (text class identity) → Paśyantī (phoneme sequence) → Madhyamā (acoustic features) → Vaikharī (audio waveform). Running the Vāk hierarchy forward. Grammar converts abstract text into physical sound.
Language has grammar at every level. Agni for token frequency patterns, Vayu for contextual attention (Transformers are Vayu-dominant), Aakash for syntactic tree structure. A grammar-typed LLM would be dramatically more sample-efficient — structure replaces scale.
Prime numbers are proof that algorithmic and mathematical problems have grammar. Sorting, scheduling, graph problems, cryptography — all have structural rules that map to the five Tattvas. Any algorithm with a rule-based structure can be expressed as a grammar and solved by NeuroNet.
Physical action is governed by Vayu (variational calculus — optimal policy) and Jal (discrete transitions between states). A robot's motion grammar is as structured as a language grammar. The Sandhi constraints = physical constraints on which movements can follow which.
Physics, chemistry, biology — all operate under structured rules. The grammar of physics is the symmetry group (Aakash). The grammar of chemistry is the periodic table (Prithvi selection + Jal differences). Grammar-typed AI for science could discover structure the way it discovered prime gaps.
Every structured domain — language, mathematics, code, physics, biology — has grammar underneath it. That grammar is always composed of the same five types of transformation: selection, flow, frequency, policy, and structure. A model that is built to match this grammar will learn more efficiently, remain interpretable, and generalise more broadly than one built without it.
All share the same five-operation grammar. NeuroNet is the first framework that makes this explicit in a neural architecture.
The goal is not AGI in the narrow sense of "passes every benchmark." The goal is Artificial Generative Intelligence — a system that can generate structured behaviour in any domain, because it has learned the grammar of structure itself.
| Property | Standard Deep Learning | NeuroNet Framework |
|---|---|---|
| Architecture | Fixed (CNN, Transformer...) | Grammar-typed, evolves with problem |
| Routing | Implicit (all weights learn everything) | Explicit named operations per input |
| Explanation | Post-hoc (LIME, SHAP, attention maps) | Intrinsic — readable reasoning path |
| Knowledge update | Full retraining required | One-line grammar edit, no retraining |
| Domain transfer | Domain-specific architectures | Same framework, different dominant Tattvas |
| Scale needed | Billions of parameters | 67K–117K for competitive accuracy |
| Trust / audit | Black box | Every decision traceable to named op |
"Grammar is not about Sanskrit specifically. Sanskrit is just where the grammar happened to be found. The same grammar governs all structured phenomena."
— Framework thesis, paraphrasing PāṇiniResearch in progress · Gopal Kumar · May 2026