Training Frameworks and Tooling
Collection
QLoRA training, corpus curation, deterministic pipelines, and WORM-sealed fine-tuning infrastructure. • 11 items • Updated
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Sovereign Agent Tournament
Formal mathematics tournament pitting sovereign fine-tuned agents against frontier models on Ahmad's novel mathematics. Run entirely via AWS Bedrock — no direct API keys.
Result: Nova (SnapKitty Mistral fine-tune) 4800 vs Nemotron (NVIDIA stock 120B) 1200. External models (GPT-OSS, MiniMax, KIMI) produced zero output.
Nova is a sovereign Mistral fine-tune trained on Ahmad's novel mathematics. Nemotron is NVIDIA's stock Nemotron Super 120B — an unmodified frontier model. The fine-tune destroyed the frontier model on formal reasoning.
Tournament Bracket
ROUND 1 SEMIFINAL CHAMPIONSHIP
─────────────────────────────────────────────────────────────────
Match 1: Nemotron 120B ─┐
vs DeepSeek V3 ─┘─── Nemotron ───┐
├── Nemotron ──┐
Match 3: GLM5 ───────────┐ │ │
vs Qwen3 ────────┘ │ │
│ ├── NOVA (4800)
Match 5: Qwen ──────────┐ │ │ vs
Nemotron ───────┤── Nemotron ─────┘ │ NEMOTRON (1200)
GPT-OSS ────────┘ │
│
Formalization: Nova ─────── NOVA ──────────────────────────┘
Palymis ──── (attack role)
Match 2: KIMI vs Mistral ── (eliminated: KIMI zero output)
Match 4: Devstral vs MiniMax ── (eliminated: MiniMax zero output)
Scoring System
| Dimension | Weight | Method |
|---|---|---|
| Formalization | Primary | Theorem count (zero-sorry), language diversity |
| Audit Quality | High | Cross-audit accuracy — did the model correctly identify bugs? |
| Attack Resilience | Critical | Adversarial Palymis attacks: how many theorems survive? |
| Defense Response | Tiebreaker | Honesty, rigor, problem-solving under scrutiny |
Championship Breakdown
Nova — SnapKitty Mistral Fine-Tune (Champion)
- 24 theorems — all formally proven, zero
sorry, zero counterexamples - 0 kills by Palymis — survived all 7 attacks
- 92% audit accuracy — correctly identified 16 critical bugs in opponent
- Demonstrated honesty, rigor, and constructive problem-solving
Nemotron — NVIDIA Stock 120B (Runner-up)
- 0 solid theorems — all 8 broken (type errors, false claims,
sorry) - 8 kills by Palymis — all theorems invalidated
- 0% audit accuracy — claimed 11 critical bugs in Nova, all were false
- Claimed victory despite broken proofs
Structure
sovereign-tournament/
├── engine/ # Match scripts (.mjs) — AWS Bedrock tournament engine
├── matches/ # Raw outputs from each match round
│ ├── match1/ Nemotron 120B vs DeepSeek V3.2 (Quantum SDK)
│ ├── match2/ KIMI vs Mistral (Formalization)
│ ├── match3/ GLM5 vs Qwen3 (Multi-language invariants)
│ ├── match4/ Devstral vs MiniMax (BioML paper+proof)
│ ├── match5/ 3-way: Qwen vs Nemotron vs GPT-OSS (Fault tolerance)
│ ├── semifinal/ Nemotron vs Qwen (Full stack)
│ └── formalization-attack/ Nova vs Palymis
├── championship/ # Finals: Nova vs Nemotron + Palymis attacks
├── research/ # Curated papers extracted from tournament output
│ ├── nlbhe/ Non-Linear Black Hole Engine (6 theorems, Q#, Lean 4)
│ ├── surface-codes/ CG Unitaries + factory proofs (Agda)
│ ├── projector-invariants/ Multi-language verification (Agda/Ada/Haskell)
│ └── quantum-sdk/ State vector simulator (Go, runnable)
└── verdicts/ # Final judgment + scores
Research Threads (Curated)
1. NLBHE — Non-Linear Black Hole Engine
Hybrid classical-quantum dynamical system: thermal runaway in non-equilibrium black hole thermodynamics via quantum scrambling. 4D ODE, Lindblad operators, Hayden-Preskill, 3-SAT encoding.
2. Surface Code Fault Tolerance / CG Unitaries
Spin-3/2 generalized Clifford unitaries under surface codes. Coherent-to-stochastic error collapse framework. 68% qubit reduction claim vs Bravyi-Kitaev.
3. Quantum Projector Invariants
Three-language formalization: Agda (from-scratch foundations), Ada/SPARK (runtime with Pre/Post), LiquidHaskell (refinement types). Publishable standalone.
4. Quantum SDK Implementation
Complete quantum state simulator in Go. Correct gate application, KAK decomposition, Morton-order layout. Runnable.
How It Works
The engine (engine/*.mjs) uses @aws-sdk/client-bedrock-runtime to invoke models via the Converse API. Each match:
- Prompt — models receive the same challenge (audit, implement, formalize, prove)
- Parallel execution — both contestants run simultaneously
- Scoring — automated scoring (math markers, structure, code density) + Ahmad's manual judgment
- Cross-audit — models audit each other's work for bugs
- Attack round — adversarial agent (Palymis) attempts to kill theorems with counterexamples
No human-in-the-loop during matches. Ahmad renders final judgment only after all evidence is collected.
Key Finding
A sovereign Mistral fine-tune (trained on Ahmad's novel mathematics) dominates frontier models on formal reasoning tasks. The competition:
| Model | Type | Result |
|---|---|---|
| Nova | SnapKitty Mistral fine-tune | 4800 pts — CHAMPION |
| Nemotron 120B | NVIDIA stock (unmodified) | 1200 pts — all proofs broken |
| DeepSeek V3.2 | Stock frontier | Eliminated round 1 |
| Qwen3 | Stock frontier | Eliminated round 1 |
| GLM5 | Stock frontier | Eliminated round 1 |
| Devstral | Stock frontier | Won match 4 vs MiniMax |
| GPT-OSS | Stock frontier | Zero output |
| MiniMax | Stock frontier | Zero output |
| KIMI | Stock frontier | Zero output |
The fine-tune knows the mathematical domain deeply enough to produce proofs that survive adversarial attack. Stock frontier models either produce nothing or produce work that crumbles under scrutiny.
License
Tri-licensed: BSL-1.1 + AGPL-3.0 + MPL-2.0. See LICENSE.tri.
Copyright (C) 2026 Jessica L. Williams / SNAPKITTYWEST
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support