Most “AI writes the code, humans review it” tooling tries to make the model write cleaner diffs. Jacquard flips it: change the language so review scales even when a machine wrote every line. It’s a v0.1 research prototype, not something to ship — but the design choices are the sharpest answer I’ve seen to “reviewing agent output is the real bottleneck.”

• 🎯 Effects live in the type signature. A function reads (text) ->{net} text, so you know it touches the network without reading the body — the runtime rejects ungranted effects. That’s the single question I ask most when reviewing an agent’s diff: what can this actually reach?
• 🔍 Structural identity, not source bytes. Definitions are content-addressed on resolved structure, so a rename or reformat doesn’t re-trigger downstream verification. Your review and eval only fire on real semantic change.
• ⚡ Worlds instead of mocks. The same code runs against a real network or a scripted fake by swapping handlers — testing without the mocking scaffolding that usually rots.
• 💡 Uncertainty as a first-class value. Probabilistic sampling with exact enumeration, so “how sure is this” is something the program states, not something you infer.
• 📊 Byte-identical AOT output. The native compile matches the interpreter exactly — the artifact you review is provably the artifact that runs.

The capability-based authority model is doing the heavy lifting: it’s the same instinct behind sandboxing tool use in an agent framework, pushed down into the language itself. Whether a bespoke language is the right layer — versus effect systems bolted onto languages agents already write fluently — is the open question. Details are in the Show HN thread.

If the bottleneck in agentic coding is trust, not generation, does the fix belong in the model, the harness, or the language?