Fugu
Verified facts Overview

Public-path ANE substrate · source private

Emit. Compile. Audit.

Verified Neural Engine substrate outcomes — not a how-to.

See proven facts What stays private
Mako, the Codex project pet

Toolchain

Public path

Proven · Apple public compiler / Core ML path

Vertical slice

Emit→Compile→Audit

Proven · compile-clean Phase 11a slice

Entitlement bypass

Not required

Proven · for the useful public workflow

Laboratory source

Private

Proven · no public reconstruction surface

What Fugu is

A Neural Engine substrate that can emit a package, compile it through Apple’s public toolchain, and audit the result — measured on real Apple Silicon.

This page publishes verified capability claims and proof labels. It is not a converter tutorial, not an operator catalog, and not a reconstruction manual.

Verified facts

Each row is a distinct claim. Numbers from sibling projects are not merged into Fugu headlines.

Claim Status Notes
Emit → compile → audit vertical slice runs on-device Proven Phase 11a compile-clean slice on Apple’s public path
Useful workflow needs no entitlement bypass Proven Same public-compiler posture as the ANE-RE / Qwen programs
Native Swift substrate extracted and packaged as its own artifact Proven Laboratory source remains private; this page is outcomes only
Placement / program inspection hooks exist in the lab stack Proven Capability claim only — interfaces and probes unpublished
Universal PyTorch→ANE conversion as a finished product Unmeasured / incomplete Ambition, not a shipped universal converter claim on this page

Scope (high level)

Emit

Author a Neural Engine–bound package from a controlled specification surface. Details unpublished.

Compile

Hand the package to Apple’s public compiler path and obtain a device-loadable result.

Audit

Check that the compiled artifact matches the lab’s integrity / placement expectations.

Evidence discipline

Public claims stay labeled Proven, Estimated, Unmeasured, Falsified, or Closed.

What stays private

Source repositories, graph topologies, operator sequences, tensor layouts, seed packages, CLI contracts, thresholds, weight locations, unpublished compiler workarounds, internal ledgers, and integration blueprints remain private. Hints about what is possible are intentional. Recipes for how to copy it are not.

If a detail would let a skilled engineer rebuild the substrate from this page alone, it does not belong here.