Chapter 3 · Proof

Does it work?

AILANG started 65 points behind Python on Claude Sonnet 4.5. Watch the gap close.

−

success-rate gap Claude Sonnet 4.5 · core tier

data: ailang.sunholo.com/benchmarks/latest.json · snapshot 2026-05-11

From −65 to parity on Claude Sonnet 4.5. The dips near the top aren't regressions — that's where saturated benchmarks get retired and harder ones come in. The bar keeps rising even after we catch up.

AILANG in production

100% AI‑coded apps. Running today.

Real WASM modules in your browser — no Python, no server — every extraction validated by a contract the compiler refuses to ship without.

Document Intelligence

AILANG Parse

DOCX, PPTX, XLSX, PDF & images parsed directly in your browser. Track changes, merged cells, comments — preserved, not flattened.

69 files 15 formats 3 / 3 track changes (vs 0 / 5 others)

Streaming & Voice

Voice DocParse + Website Builder

Talk to a document via Gemini Live. Or describe a site and watch it build — AILANG WASM modules driving the whole pipeline.

WASM runtime Gemini Live bidi audio no server

Game engine

Stapledon's Voyage

A hard-sci-fi relativity simulator. AILANG drives the game logic — special & general relativity, flybys, time dilation. Not a documents demo.

AILANG game core SR + GR physics open source

Everything you see is 100% AI-coded in AILANG — compiler, type checker, effects, Z3 verification, WASM runtime, and stdlib included. sunholo.com/ailang-demos

The Phoenix idea

A coding harness that grows itself — from a verifiable core.

built by @arniwesth

λ_ motoko

Phoenix architecture

Start with a small, verifiable AILANG core. The agent uses the core to write more AILANG — tools, extensions, even parts of itself. Every step goes through the compiler. If it can't be verified, it doesn't ship.

Verifiable coreAILANG types · contracts · effects all checkable

Self-verifying loopthe agent runs its own compiler · fixes its own errors

Self-extendingAI builds new tools in AILANG · verified before they're trusted

Self-correcting loop

When the AI gets it wrong… the language tells it.

Live session · openrouter/z-ai/glm-5 · task: wire an MCP subprocess into AILANG · 2026-05-07

~/motoko_agent · session_2026-05-07.jsonl step 61

no human in the loop · 8 steps of self-repair after the parse error

Don't ask AI to learn your language.
Build a language AI can speak. Cheaper to debug. Replayable. Verifiable.

Static types · the AI safety net Pre-runtime, parseable errors mean one-turn fixes — not production crashes.

Neurosymbolic verification AI writes the code, Z3 proves the contracts. Counterexamples beat tests.

Works in any stack Tight loops, explicit context, structured errors — push for these where you are.

ailang.sunholo.com language · playground · demos

mcp.ailang.sunholo.com live MCP for Claude / Cursor / Cline

github.com/sunholo-data/ailang source · compiler · stdlib

github.com/arniwesth/motoko_agent the motoko coding harness