synthfuse 0.1.0

ALCHEM-J: operator-fusion language for swarm-RL-numeric hybrids

Synth-Fuse ⚡ ALCHEM-J

“Every algorithm is a plugin; every fusion is a pipeline; every pipeline is a JAX transform.”

[

Synth-Fuse is a JAX-native fusion engine that composes swarm intelligence, reinforcement learning, and numerical algorithms into a single, differentiable, hardware-scalable execution graph—written as one-line spells.

---

⚡ 30-Second Demo

pip install synthfuse
alj
>>> (𝕀 ⊗ 𝕃(alpha=1.5) ⊗ ℂ(r=3.8))
>>> run 100

Stage	Primitive	Meaning
𝕀	ISO/RIME	Swarm exploration
𝕃	Lévy	Heavy-tailed jumps
ℂ	Chaos	Adaptive perturbation

One-liner Benchmarks

# Continuous optimisation
uv run sfbench "(𝕀 ⊗ 𝕃(alpha=1.5))" --bench rastrigin --dims 1000 --steps 5000

# Combinatorial
uv run sfbench "𝕎 ⊗ ℍ" --bench tsp-200 --pop 256

# Live telemetry
uv run sfmonitor --recipe fql_rime --steps 500

🧪 Pre-Built Recipes (import → JIT)

from synthfuse.recipes import fql_rime, mrbmo_ppo, ns2uo, ntep, stcl

step, state = fql_rime.make(dims=1000, pop=512)

Recipe	Spell	Use-Case
FQL-RIME	(𝔽𝕃 ⊗ 𝕃 ⊗ ℝ𝔽)	Flow-guided Lévy + PPO
MRBMO-PPO	(𝕊𝕄 ∘ 𝜑 ⊗ ℝ𝕄)	Siege-elite PPO
NS²UO	(𝕊𝕨 ⊗ 𝕆𝕊)	Neuro-Swarm-to-Universal-Opt
NTEP	(𝕎 ⊗ 𝕀𝕋)	Neural Tool-Embedding Protocol
STCL	(𝕊𝕋 ⊗ 𝕎)	Semantic-Thermo Compression Loop

📦 Install

# stable
pip install synthfuse

# dev speed-run
curl -LsSf https://astral.sh/uv/install.sh | sh
uv pip install synthfuse[dev]

🪄 Spell Language (ALCHEM-J)

Symbol	Primitive	Meaning
𝕀	ISO/RIME	Swarm update
ℝ	PPO/A2C/DQN	RL policy
𝕃	Lévy	Heavy-tailed noise
𝕊	SVD-UKF	Low-rank stabiliser
𝕎	Weierstrass	Semantic gravity field
ℂ	Chaos	Adaptive perturbation
𝜑	Meta-gradient	Natural-gradient correction
Compose via pure combinators:

• ⊗ – sequential fusion

• ⊕ – parallel fusion

• ∘ – conditional fusion

Example spell:

"(𝕎 ⊗ 𝕊𝕋 ⊗ ℍ)(sigma=0.7, halt=cos<0.01)"

🧰 CLI Tools

Tool	Command	Purpose
REPL	alj	Interactive spell casting
Benchmark	sfbench <spell>	Standardised perf tests
Monitor	sfmonitor --recipe <name>	Live telemetry dashboard
Visualise	sfviz "<spell>" -f svg	Spell → SVG / HTML

🏗️ Adding Your Own Primitive

# alchemj/plugins/my_step.py
from synthfuse.alchemj.registry import register
import jax
import jax.numpy as jnp

@register("𝕏")
def my_step(key: jax.Array, state: PyTree, params: dict) -> PyTree:
    """One-sentence purpose. Failure: alpha<0 → divergence; caught by CI."""
    alpha = params.get("alpha", 1.0)
    return state.replace(x=state.x + alpha * jax.random.normal(key, state.x.shape))

Micro-bench in tests/test_my_step.py → open PR → fast-merge.

---

🧪 Emergent Behaviours (observed & bounded)

Behaviour	Mechanism	Control Signal
Spontaneous decentralisation	consensus dominance	modularity Q ≥ 0.3
Crosstalk-free routing	zeta pole separation	σ_zeta gradient
Thermal self-balancing	Hamiltonian heat term	κ ≤ max-current-density
Topology-safe compression	semantic load Λ ≥ τ	rollback if Λ < τ
Single-shot convergence	Weierstrass smoothing	σ_min clamp

🔐 Security & Safety

• Pure functions only – no side-effects inside step.

• Sandboxed execution – external I/O outside JIT.

• Checkpoint every 15 min – rollback < 30 s.

• Formal verification hooks exported for external provers.

📜 Citation

@software{synthfuse2026,
  author = {Jimenez, J. Roberto and K2, Kimi},
  title = {Synth-Fuse: A Modular Fusion Library for Hybrid Swarm–RL–Numerical Intelligence},
  url = {https://github.com/deskiziarecords/synthfuse},
  version = {0.1.0-alpha},
  date = {2026-01-17},
  license = {Apache-2.0}
}

---

Support Me

If you would like to support the development of these resources, consider contributing towards helping me get some gear for continued improvement or simply treating me to a coffee. Your support means a lot!" buy me a coffee

---

🤝 Contributing

We treat PRs like spells: pure, composable, JIT-ready.
Read CONTRIBUTING.md → open issue/PR → fast-merge.

---

💬 Community

• Matrix: #synthfuse:matrix.org

• Discussions: GitHub Discussions tab

• Email: [e-mail] | [e-mail]

Welcome to the fusion – cast your spell.