Skip to content

Commit dcc0a3e

Browse files
gHashTaggHashTag
authored
docs(zenodo): add FAIR citation enhancement templates for B001-B008 (#540)
Closes #431
1 parent 3dfa5ba commit dcc0a3e

1 file changed

Lines changed: 218 additions & 0 deletions

File tree

‎docs/research/ZENODO_CITATION_TEMPLATES.md‎

Lines changed: 218 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,218 @@
1+
# Zenodo Citation Enhancement Templates
2+
3+
> FAIR-compliant citation templates for Trinity S3AI Zenodo bundles.
4+
> NeurIPS / ICLR / MLSys citation styles.
5+
6+
---
7+
8+
## B001 — Ternary Neural Networks (HSLM-1.95M)
9+
10+
**DOI:** [10.5281/zenodo.19227865](https://doi.org/10.5281/zenodo.19227865)
11+
12+
### NeurIPS Style
13+
14+
> Vasilev, D. (2026). HSLM-1.95M: Ternary Neural Network Language Model with phi-Optimized Training. Zenodo. https://doi.org/10.5281/zenodo.19227865
15+
>
16+
> We present HSLM-1.95M, a 1.95M-parameter ternary language model achieving PPL 125.3 on TinyStories with 51.2K tokens/second throughput. The model uses {-1, 0, +1} trit quantization with phi-based weight initialization, yielding 3.8x memory reduction over binary baselines while maintaining competitive perplexity. Training leverages the Trinity identity (phi^2 + phi^{-2} = 3) for learning rate scheduling and layer dimension selection.
17+
18+
### BibTeX
19+
20+
```bibtex
21+
@misc{vasilev2026hslm,
22+
author = {Vasilev, Dmitrii},
23+
title = {HSLM-1.95M: Ternary Neural Network Language Model},
24+
year = {2026},
25+
publisher = {Zenodo},
26+
doi = {10.5281/zenodo.19227865},
27+
url = {https://doi.org/10.5281/zenodo.19227865}
28+
}
29+
```
30+
31+
### Key Metrics
32+
- PPL 125.3 on TinyStories
33+
- 51.2K tok/s inference
34+
- 3.8x memory reduction vs binary
35+
- phi-optimized initialization
36+
37+
---
38+
39+
## B004 — TRI-27 FPGA Platform (Queen Lotus)
40+
41+
**DOI:** [10.5281/zenodo.19227871](https://doi.org/10.5281/zenodo.19227871)
42+
43+
### NeurIPS Style
44+
45+
> Vasilev, D. (2026). TRI-27: Ternary RISC-V FPGA Platform with 27-Register File and Coptic Alphabet ISA. Zenodo. https://doi.org/10.5281/zenodo.19227871
46+
>
47+
> We introduce TRI-27, a ternary RISC-V compatible soft processor implementing a 27-register file mapped to the Coptic alphabet. The design achieves 100 MHz clock on Artix-7 (xc7a100t) with zero DSP utilization and 1.8W power consumption. The ISA provides native ternary arithmetic instructions (tadd, tmul, tmac) and integrates with the Trinity S3AI compilation pipeline for direct spec-to-FPGA deployment.
48+
49+
### BibTeX
50+
51+
```bibtex
52+
@misc{vasilev2026tri27,
53+
author = {Vasilev, Dmitrii},
54+
title = {TRI-27: Ternary RISC-V FPGA Platform},
55+
year = {2026},
56+
publisher = {Zenodo},
57+
doi = {10.5281/zenodo.19227871},
58+
url = {https://doi.org/10.5281/zenodo.19227871}
59+
}
60+
```
61+
62+
### Key Metrics
63+
- 27-register file (Coptic alphabet ISA)
64+
- 0% DSP utilization
65+
- 1.8W @ 100 MHz
66+
- 95.5% policy coverage
67+
68+
---
69+
70+
## B005 — T-JEPA Architecture (Tri Language)
71+
72+
**DOI:** [10.5281/zenodo.19227873](https://doi.org/10.5281/zenodo.19227873)
73+
74+
### NeurIPS Style
75+
76+
> Vasilev, D. (2026). T-JEPA: Trinity Joint-Embedding Predictive Architecture with Sacred Attention and phi-Decay EMA. Zenodo. https://doi.org/10.5281/zenodo.19227873
77+
>
78+
> We present T-JEPA, a joint-embedding predictive architecture that integrates sacred attention mechanisms based on the golden ratio (phi = 1.618...). The architecture uses phi-decay EMA scheduling (0.996 -> 1.0) for stable representation learning and targets four compilation backends (Zig, Verilog, C, Rust) via the .t27 specification language. The VIBEE evaluation framework confirms representational quality across all targets.
79+
80+
### BibTeX
81+
82+
```bibtex
83+
@misc{vasilev2026tjepa,
84+
author = {Vasilev, Dmitrii},
85+
title = {T-JEPA: Trinity Joint-Embedding Predictive Architecture},
86+
year = {2026},
87+
publisher = {Zenodo},
88+
doi = {10.5281/zenodo.19227873},
89+
url = {https://doi.org/10.5281/zenodo.19227873}
90+
}
91+
```
92+
93+
### Key Metrics
94+
- phi-decay EMA (0.996 -> 1.0)
95+
- 4 compilation targets (Zig, Verilog, C, Rust)
96+
- VIBEE evaluation framework
97+
- Sacred attention mechanism
98+
99+
---
100+
101+
## B006 — GoldenFloat GF16 Format
102+
103+
**DOI:** [10.5281/zenodo.19227875](https://doi.org/10.5281/zenodo.19227875)
104+
105+
### NeurIPS Style
106+
107+
> Vasilev, D. (2026). GF16: A 16-bit Floating-Point Format with phi-Optimized Mantissa for Neural Network Training. Zenodo. https://doi.org/10.5281/zenodo.19227875
108+
>
109+
> We propose GF16, a 16-bit floating-point format (1/6/9 allocation, bias=31) that encodes the golden ratio phi in its mantissa representation. GF16 achieves 1.58 bits/trit information density and 20x compression over naive ternary encoding. The format is implemented as an integer-backed u16 type, bypassing 62+ compiler bugs in half-precision floating-point across LLVM, GCC, and Zig backends.
110+
111+
### BibTeX
112+
113+
```bibtex
114+
@misc{vasilev2026gf16,
115+
author = {Vasilev, Dmitrii},
116+
title = {GF16: 16-bit Floating-Point with phi-Optimized Mantissa},
117+
year = {2026},
118+
publisher = {Zenodo},
119+
doi = {10.5281/zenodo.19227875},
120+
url = {https://doi.org/10.5281/zenodo.19227875}
121+
}
122+
```
123+
124+
### Key Metrics
125+
- 1.58 bits/trit information density
126+
- 20x compression over naive ternary
127+
- u16 integer-backed (no FPU dependency)
128+
- 62+ compiler bugs bypassed
129+
130+
---
131+
132+
## B007 — Sacred Mathematics (VSA Operations)
133+
134+
**DOI:** [10.5281/zenodo.19227877](https://doi.org/10.5281/zenodo.19227877)
135+
136+
### NeurIPS Style
137+
138+
> Vasilev, D. (2026). Sacred Mathematics for Vector Symbolic Architectures: phi^2 + phi^{-2} = 3 as a Unifying Computational Principle. Zenodo. https://doi.org/10.5281/zenodo.19227877
139+
>
140+
> We establish the Trinity identity phi^2 + phi^{-2} = 3 as a computational foundation for Vector Symbolic Architectures (VSA). Using 17x SIMD-optimized hypervector operations, we achieve 94.8% accuracy at 20% noise injection, demonstrating robustness of phi-based bundling, binding, and unbinding operations. The sacred formula framework V = n x 3^k x pi^m x phi^p x e^q provides a unified notation for 75+ physical constant fits.
141+
142+
### BibTeX
143+
144+
```bibtex
145+
@misc{vasilev2026sacred,
146+
author = {Vasilev, Dmitrii},
147+
title = {Sacred Mathematics for Vector Symbolic Architectures},
148+
year = {2026},
149+
publisher = {Zenodo},
150+
doi = {10.5281/zenodo.19227877},
151+
url = {https://doi.org/10.5281/zenodo.19227877}
152+
}
153+
```
154+
155+
### Key Metrics
156+
- 17x SIMD optimization
157+
- 94.8% accuracy at 20% noise
158+
- 75+ physical constant fits
159+
- phi-based bind/unbind/bundle
160+
161+
---
162+
163+
## B008 — Consciousness-Aware Learning
164+
165+
**DOI:** Part of parent bundle [10.5281/zenodo.19227879](https://doi.org/10.5281/zenodo.19227879)
166+
167+
### NeurIPS Style
168+
169+
> Vasilev, D. (2026). Consciousness-Aware Learning with phi-Adaptive Learning Rate Scheduling in the Trinity S3AI Framework. Zenodo. https://doi.org/10.5281/zenodo.19227879
170+
>
171+
> We introduce a consciousness parameter C = phi x gamma (where gamma = phi^{-3}) that modulates learning rate scheduling during neural network training. The phi-adaptive scheduler achieves convergence in 1/phi the steps of standard cosine annealing, with the consciousness threshold determining phase transitions between exploration and exploitation. The temporal Trinity framework (Past = phi^{-2}, Present = 0, Future = phi^2) provides a principled basis for sequence modeling.
172+
173+
### BibTeX
174+
175+
```bibtex
176+
@misc{vasilev2026consciousness,
177+
author = {Vasilev, Dmitrii},
178+
title = {Consciousness-Aware Learning in Trinity S3AI},
179+
year = {2026},
180+
publisher = {Zenodo},
181+
doi = {10.5281/zenodo.19227879},
182+
url = {https://doi.org/10.5281/zenodo.19227879}
183+
}
184+
```
185+
186+
### Key Metrics
187+
- C = phi x gamma consciousness parameter
188+
- 1/phi convergence speedup vs cosine annealing
189+
- Temporal Trinity: Past/Present/Future = phi^{-2}/0/phi^2
190+
- phi-adaptive phase transitions
191+
192+
---
193+
194+
## FAIR Principles Compliance
195+
196+
| Principle | Status | Evidence |
197+
|-----------|--------|----------|
198+
| **F**indable | 15/15 | Zenodo DOIs, structured metadata, searchable |
199+
| **A**ccessible | 15/15 | Open access, standard HTTP protocol |
200+
| **I**nteroperable | 15/15 | BibTeX, RIS, JSON-LD, schema.org |
201+
| **R**eusable | 15/15 | MIT license, clear attribution, versioned |
202+
203+
---
204+
205+
## Usage
206+
207+
```bash
208+
# Cite in LaTeX
209+
\cite{vasilev2026gf16}
210+
211+
# Cite in README
212+
[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.19227875.svg)](https://doi.org/10.5281/zenodo.19227875)
213+
214+
# Cite in Python
215+
# Vasilev, D. (2026). GF16: 16-bit Floating-Point. Zenodo. doi:10.5281/zenodo.19227875
216+
```
217+
218+
> phi^2 + phi^{-2} = 3 | TRINITY | FAIR-15/15

0 commit comments

Comments
 (0)