NSAIF is an open-source research framework for composing neural representations, symbolic rules, causal graphs, adaptive model topology, and multiple forms of memory behind one inspectable Python API.
Important: NSAIF is an early research prototype. Mathematical claims and benchmark results must be independently reviewed and reproduced before they are treated as established scientific results.
Input -> Perception -> Cognitive Bridge -> Reasoning -> Action
| | |
+------ Adaptive Topology ------+
|
Episodic + Semantic + Procedural Memory
The implementation contains:
nsaif.core: fuzzy operators, symbolic tensors, and gradient routing.nsaif.reasoning: knowledge graphs, causal interventions, modal facts, and metacognitive confidence checks.nsaif.topology: auditable MDL-inspired grow/prune policies.nsaif.memory: episodic, semantic, and procedural stores with unified recall.nsaif.perception: lightweight vision, language, and multimodal adapters.nsaif.training: hybrid losses, curriculum scheduling, and a generic trainer.
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
python -m pip install -e ".[dev]"
python -m pytestfrom nsaif import NSAIFConfig, NSAIFModel
config = NSAIFConfig.from_preset("nsaif-small")
model = NSAIFModel.from_config(config)
output = model.reason(
"If all mammals breathe air, and dolphins are mammals, "
"do dolphins breathe air?",
return_proof=True,
)
print(output.answer) # Yes
print(output.proof_trace) # Auditable logical steps
print(output.confidence) # Confidence estimateThe same API accepts explicit facts and rules:
model.learn_fact("dolphin", "is_a", "mammal")
model.learn_rule("mammal", "breathes", "air")
result = model.query("dolphin", "breathes", "air")nsaif/
|-- README.md
|-- LICENSE
|-- CONTRIBUTING.md
|-- CITATION.cff
|-- pyproject.toml
|-- nsaif/
| |-- core/
| |-- reasoning/
| |-- topology/
| |-- memory/
| |-- perception/
| `-- training/
|-- benchmarks/
| |-- arc_challenge/
| |-- gsm8k/
| |-- clutrr/
| `-- custom_nsaif_bench/
|-- docs/
| |-- theorems/
| |-- tutorials/
| `-- api/
|-- experiments/
|-- configs/
`-- tests/
CognitiveBridge maps dense values to interpretable symbolic labels and can
reconstruct a dense representation from symbolic activations. With PyTorch
installed, callers may attach the bridge to a differentiable model; the core
package intentionally remains usable without heavyweight dependencies.
from nsaif.core import CognitiveBridge
bridge = CognitiveBridge(
neural_dim=4,
symbolic_vocab=["cold", "warm", "bright", "dark"],
norm_type="godel",
)
symbols, trace = bridge.encode([0.1, 0.9, 0.7, 0.2])
print(symbols.top_symbols(2))from nsaif.reasoning import CausalGraph
graph = CausalGraph()
graph.add_edge("treatment", "outcome", weight=0.8)
estimate = graph.do(treatment=1.0).query("outcome")This module is a transparent educational intervention engine, not a substitute for validated statistical causal inference in medical, legal, financial, or other high-stakes settings.
The topology manager records every decision and evaluates a simplified minimum description length objective:
objective = data_loss + mdl_lambda * structural_complexity
Adapters can connect these decisions to real PyTorch modules. The base package does not silently mutate arbitrary production models.
Each benchmark directory contains a reproducible harness and result schema. No headline scores are shipped as verified results. Run the harnesses, record the environment and seed, and publish generated JSON artifacts with any claim.
python -m benchmarks.run_all --output artifacts/benchmark-results.jsonVersion 0.1.0 provides a tested reference architecture and extension points.
It is not a trained 7B model, a proof assistant, or a clinically validated
decision system. The nsaif-7b-reasoning preset is a configuration description
for downstream integrations; it does not download or bundle model weights.
- Differentiable grounding with formally measured approximation error.
- Verified rule backends using Lean, Coq, or SMT solvers.
- Distributed knowledge graphs and multi-agent memory.
- Continual topology adaptation with stability guarantees.
- Reproducible comparisons on ARC, GSM8K, CLUTRR, and causal benchmarks.
See CITATION.cff. A BibTeX example is also available in
docs/citation.bib.
Read CONTRIBUTING.md before opening a pull request and SECURITY.md before reporting a vulnerability.
Every contribution supports compute infrastructure, dataset access, reproducible experiments, documentation, and open publication.
Select the account matching your transfer currency. Please verify every field before confirming a payment with your bank.
EUR - SEPA / International SWIFT
| Transfer field | Details |
|---|---|
| Beneficiary | Ciprian Stefan Plesca |
| IBAN | BE83 9679 1975 8915 |
| SWIFT / BIC | TRWIBEB1XXX |
| Bank | Wise |
| Bank address | Rue du Trone 100, 3rd floor, Brussels, 1050, Belgium |
Recommended use: IBAN for SEPA transfers; IBAN and SWIFT/BIC for international transfers.
GBP - United Kingdom / International SWIFT
| Transfer field | Details |
|---|---|
| Beneficiary | Ciprian Stefan Plesca |
| Account number | 92055372 |
| Sort code | 23-14-70 |
| IBAN | GB68 TRWI 2314 7092 0553 72 |
| SWIFT / BIC | TRWIGB2LXXX |
| Bank | Wise Payments Limited |
| Bank address | 1st Floor, Worship Square, 65 Clifton Street, London, EC2A 4JE, United Kingdom |
Recommended use: account number and sort code for UK transfers; IBAN and SWIFT/BIC for international transfers.
USD - ACH / Wire / International SWIFT
| Transfer field | Details |
|---|---|
| Beneficiary | Ciprian Stefan Plesca |
| Account type | Checking |
| Routing number | 026073150 - ACH and wire |
| Account number | 8314225367 |
| SWIFT / BIC | CMFGUS33 |
| Bank | Community Federal Savings Bank |
| Bank address | 89-16 Jamaica Ave, Woodhaven, NY 11421, United States |
Recommended use: routing and account numbers for US transfers; SWIFT/BIC for international transfers.
Compute time | Dataset access | Reproducible science | Open publication
Thank you for helping NSAIF remain independent, transparent, and accessible.
NSAIF is released under the MIT License.