A numeric literal names its own type with a suffix after a required _: 255_u8, -128_i8, 1.5_f32, 0xFF_u64, 2_bf16. The suffix is a range-checked type annotation that context never retypes, so 1_i64...

Taoのタオイズム・Gowersの数学わず・Jevとヨブ

Frontier v4: 115 Claude Code runs of Claude Opus 5.5 at five effort levels, Code quality judged by Muse Spark 1.3 and Grok 4.6. Best combined score 91.11 at high, second only to Fable 5.1 at max; medi...

GPT-2 (124M) compiled from its ONNX graph by Loam, running in the browser as a Wado WebAssembly component.

Claude Code sessions in this repository can no longer delegate work to subagents or to other sessions. Delegated work errs often, and only its final report comes back, so it can be neither watched ...

wado test --coverage reports which lines, branches and functions the tests ran. It prints a per-file summary and writes LCOV, JSON and a baseline to build/coverage/. The standard library is held to...

Loam turns an ONNX model into ordinary Wado source at build time, with tensor shapes checked like TypeScript types. GPT-2 now runs that way in the browser. An introduction for programmers who have nev...

A wado test test can now pass generated source to core:eval::eval and get back what the program printed and how it ended. That is the missing piece for testing a code generator as a function: gener...

A Loam module loads its weights from a safetensors checkpoint at run time, and GPT-2 (124M) continues a prompt with the tokens onnxruntime picks. The checkpoint is checked against the graph twice: ...

This fixes about 45 miscompiles, ICEs and compile-time blowups in lowering and the NIR optimizer. Each fix comes with an e2e fixture, 55 new in all. Heap-effect summaries are now solved once and ke...

Claude Code reads AGENTS.md directly, so each directory keeps a single agent guide and no CLAUDE.md link beside it. Every reference in code comments and docs names AGENTS.md, and each one points at...