RAIL SYSTEM — ledatic.org

COMPILER

Rail compiles itself. The compiler is a Rail program.

source

.rail

▶

lexer

tokens

▶

parser

AST

▶

codegen

ARM64

▶

link

binary

-- Rail compiles itself in one command
$ ./rail_native self

=== Self-compilation ===
  Source: 103,583 chars
  Compiled to assembly
  as: OK
  ld: OK
  Binary: /tmp/rail_self

ARM64 Native

Direct assembly generation. No LLVM, no GCC. Pure Rail → machine code. 297K binary, zero dependencies.

Metal GPU

Automatic GPU dispatch. Pure arithmetic maps route to Metal compute shaders. map f data just works.

Linux ARM64

Cross-compile to bare Linux ELF. Syscall-based libc, no glibc. 74K static binaries for Pi Zero.

FLEET

Three machines. One dashboard. Every node reports to the compiler.

MINI

CPU: Apple M4 Pro

RAM: 24 GB unified

Role: Compiler · Inference · Training

Agent: fleet_agent.rail — pure Rail

AIR

CPU: Apple M1

RAM: 8 GB unified

Role: Relay · Future Pi gateway

Agent: fleet_agent — compiled Rail binary

RAZER

CPU: AMD Ryzen

GPU: RTX 3070 · 8 GB VRAM

Disk: 936 GB

Role: CUDA training · Large model LoRA

Agent: fleet_agent.py — Python (x86_64)

PI ZERO PENDING

CPU: ARM Cortex-A53

Role: Edge compute · 74K Rail binaries

Status: Awaiting USB data cable

FLYWHEEL

The compiler is the oracle. Generate → Compile → Harvest → Train → Repeat.

COMPILER
ORACLE

GENERATE
code

COMPILE
verify

HARVEST
data

TRAIN
model

4.5K

verified examples

difficulty levels

92%

peak benchmark

trained models

0.5B

Smallest model. 16 LoRA layers, 1024 seq. Scores 24% baseline. Fast iteration.

1.5B

Primary training target. 16 layers, 1024 seq. Running on Mac Mini Metal.

4B

Needs Razer3070 for full training. 4 layers on Mac (OOM). CUDA unlocks 16 layers.

9B

Code brain. Inference only on 24GB. Reached level 25. Generates training data for smaller models.

LANGUAGE

Pure functional. Pattern matching. Foreign function interface. Compiles to bare metal.

-- factorial with tail-call optimization
fact n = if n <= 1 then 1 else n * fact (n - 1)

-- HTTP server in 4 lines
import "stdlib/http.rail"
handle req = respond 200 "hello world"
main = serve 8080 handle

-- GPU dispatch — map routes to Metal automatically
main = let r = map (\x -> x * 3 + 1) (range 1000000)
print (show (head r))

json

Parse & emit JSON

http

Client & server

sqlite

Database bindings

regex

POSIX regex

socket

BSD sockets

base64

Encode & decode