coralForge — Sovereign Structure Prediction Engine

📐 Architecture-ready Individual AlphaFold primitives are implemented and validated (154 checks, 1e-10 tolerance vs NumPy). The end-to-end structure prediction pipeline is designed but not yet wired — see helixVision for the full pipeline status.

What It Is

coralForge is the structure prediction engine substrate — pure Rust f64 implementations of the mathematical primitives that AlphaFold2 and AlphaFold3 use for protein structure prediction. It is the computational core that 🧬👁️ helixVision builds upon.

Repository: syntheticChemistry/coralForge (archived — moving to sporeGarden/helixVision) License: ⚖️🔓 scyBorg (AGPL-3.0-or-later + ORC + CC-BY-SA 4.0)

The Isomorphism Theorem

Every “novel” operation in AlphaFold decomposes to six fundamental primitives:

AlphaFold’s Evoformer, IPA (Invariant Point Attention), diffusion module, pairformer, and confidence heads all decompose to compositions of these six primitives. No new mathematical primitive is required.

This is the isomorphism proof: structure prediction is not a new kind of computation. It is a new composition of existing computations. The same 🐟⚡ barraCuda WGSL shaders that compute molecular dynamics forces also compute attention scores for protein structure prediction.

Dual Pipeline Architecture

AlphaFold2 Path

MSA (Multiple Sequence Alignment)
    -> Evoformer (48 blocks of row/column attention + triangle updates)
    -> Structure Module (IPA iterations -> 3D coordinates)
    -> Confidence (pLDDT, PAE, pTM)

AlphaFold3 Path

Input Embedder (no MSA required for small molecules)
    -> Pairformer (pair representation attention)
    -> Diffusion Module (iterative coordinate refinement)
    -> Confidence (pLDDT, pDE, pAE)

Both paths share the six primitives. Both run in pure Rust f64. Both produce bit-reproducible results across architectures.

Validation

154 checks across three tiers:

Cross-tier validation: every Rust result is compared to the Python reference. Every GPU result is compared to the CPU Rust result. The chain is: NumPy -> Rust -> WGSL -> silicon.

LTEE Application

The most compelling application of coralForge is not protein engineering — it is evolutionary biology:

1. Take frozen stocks from Lenski’s Long-Term Evolution Experiment (LTEE)
2. Sequence the ancestral and evolved strains ( 💧♨️ wetSpring pipeline)
3. Predict protein structures for ancestral and evolved variants (coralForge)
4. Compare structural changes to fitness trajectories ( 🔥♨️ hotSpring)
5. Identify which structural innovations correspond to adaptive events

This connects computational constrained evolution (the thesis) to biological constrained evolution (the LTEE) — both producing innovation under pressure.

What Makes It Sovereign

coralForge is not an AlphaFold clone. It is a proof that structure prediction decomposes to six primitives — and those primitives can run on sovereign hardware, in pure Rust, with full precision, producing science that belongs to the scientist who asked the question.