Helix

DNA-encoded algorithm storage and execution system.

Programs are compiled to an ACGT nucleotide sequence, stored in aqueous/encapsulated DNA, retrieved, and executed on a software virtual machine.

.hx source
    │
    ▼ helix-asm (parse + resolve + encode)
raw codons
    │
    ▼ helix-bio (GC/homopolymer/repeat optimization)
bio-safe codons
    │
    ▼ helix-ecc (SYNC + Reed-Solomon + TMR)
DNA string  ──────────────────────────── aqueous/encapsulated storage
    │
    ▼ helix-ecc (decode: TMR vote + RS correction + SYNC resync)
raw codons
    │
    ▼ helix-vm (decode + execute)
result

1. Codon degeneracy — opcodes 0x00–0x0F have synonymous codons
2. SYNC markers — 4-codon pattern inserted every 8 instructions for frame sync
3. Reed-Solomon — RS(10,8) over GF(2⁶), corrects 1 symbol error per block
4. TMR — Triple Modular Redundancy with majority vote

The bio optimizer swaps synonymous opcode codons to satisfy these constraints.

# Build everything
cargo build --workspace -j4

# Assemble a program (outputs DNA string)
cargo run --bin helix-asm -- examples/fibonacci.hx

# Assemble and run (pipe DNA into the VM)
cargo run --bin helix-asm -- examples/fibonacci.hx --no-tmr 2>/dev/null | \
  cargo run --bin helix-vm --

# Simulate DNA storage errors
cargo run --bin helix-sim -- \
  "$(cargo run --bin helix-asm -- examples/gcd.hx --no-tmr 2>/dev/null)" \
  --sub-rate 0.005 --runs 50

# Run model weights demo
cargo run --example weights --manifest-path helix-vm/Cargo.toml

Programs in examples/ use the Helix assembly language (.hx):

.program "my_program"
.version 1
.memory 256          ; optional: memory words to allocate
.data
arr: .words 1, 2, 3  ; data section (loaded at addr 0)
.text
.entry main          ; entry point label

main:
    IMML R1, 42      ; R1 = 42 (low byte)
    IMMH R1, 0       ; R1 high byte = 0
    HALT

Branches use labels — the assembler resolves offsets automatically:

loop:
    CMP  R1, R2
    BGE  done        ; branch to label
    INC  R1, R1
    JMP  loop
done:
    HALT

The helix-vm/examples/weights.rs demo shows how neural network weights can be stored as DNA:

• Quantize f32 weights → 6-bit values (0–63, one value per codon)
• Encode as 3 nucleotides per weight
• 20 weights → 60 nucleotides
• MSE from 6-bit quantization: ~0.0001

cargo run --example weights --manifest-path helix-vm/Cargo.toml

# All tests
cargo test --workspace -j4

# Specific crate
cargo test -p helix-vm -j4

# Integration tests only
cargo test --test programs -p helix-vm -j4

Test coverage:

See SPEC.md for the full technical specification.