Your AI fools you

Your work (together with Claude) is amazing. But AI makes a huge mistake. It's not hallucination, it's **association**.

**The In-Memory Path: The Holy Grail**. AI were focused on how class named, not what it do. What is it **actually**?

```
-[_ANEInMemoryModel compileWithQoS:options:error:]
-[_ANEInMemoryModel saveModelFiles]
-[_ANEInMemoryModel localModelPath]
+[_ANEModel modelAtURL:key:]
-[_ANEClient compileModel:options:qos:error:]
```

It is a **wrapper**. Wrapper that saves files, loads standard model, and compiles it as is. I decompiled it, to make sure.

Moreover, this wrapper contains **bugs**. It does not save files properly, so you still need to hack into temporary directory and put some files manually. And it **leaks** - that is your "per-process compilation limit" that requires checkpoint-exec hack.

So, when you decide to avoid to use this, you are completely fine - just put model files in any directory and use `_ANEClient` for the rest.

Another notable thing is chaining. While [!40](https://github.com/maderix/ANE/pull/40) contains a lot of research into `_ANEChainingRequest`, it is still not working for me (I have gone into it much deeper). But I have one thing that become my own game-changer: `-[_ANERequest setCompletionHandler:]`.

You can check out how it all combined in [my work](https://github.com/djphoenix/ane-rs/blob/main/ane-bridge/src/kernel.rs). If you eventually want to **join forces** - write me, but you promise that you dont't **blindly** trust your AI anymore.

Start with the in-memory call sequence in the issue, then compare it with ane-bridge/src/kernel.rs and the research in pull request 40. The issue describes wrapper bugs and chaining behavior but does not define a specific change, reproduction, or acceptance criteria, so a concrete target must be established before implementation.

Written by the indexing model from the issue text.