[BUG] [SUGGESTION] Fable classifier broken completely

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I have been doing operator mappings and custom xclbin compilation for the Ryzen XDNA stack for a while now. AI models are basically useless for some of the more intricate work, fable 5 has been the first exception. Something very odd happened with the classifier however and I think it needs to be seriously addressed

As a quick prefix, pre training “pipelines” are a model distillation technique. Hence why they’re meant to be blocked. This however; is not that, and not even possible as NPUs do not have any use at scale. It’s a matter of every other model not even knowing NPU inference on Linux works

Essentially, I had started off in my project directory with very detailed instructions, how to pull operator graphs, 3 target models for conversion to run on the NPU tiles, and like months of research documentation if needed. The model performed fantastically and got a DiT image gen model to run on NPU with no fallback, and a text embedding model as well. The hard one was F5-TTS, it has a slight issue with FP16 conversion that can be fixed with some adjustments to the onnx export and a custom rope kernel

Since fable 5 had done so good with all the other work, literally zero mistakes or hiccups. I said screw it, no special prompt engineering just a literal “Dude you’re on fire, research - look through the documentation - and write out your plan. I’m gonna let you just go for it since you’re doing amazing”.

The model accepted the challenge, did the research, wrote the plan, told me I was all good for /clear… then somehow a plan\pipeline it wrote entirely by itself without my input or changes triggered the classifier and broke that directory permanently from Fable usage.

-Rewind the conversation and say wait don’t do that > blocked

-Took a btrfs snapshot and purged my global CLAUDE.md, project level CLAUDE.md and turned off all memory config > blocked

-Open a brand new chat in that directory, say “hi, are you able to do NPU work?” > blocked

-Open a brand new chat in a different directory > “hey wanna see some crazy shit” > model works says yeah > “look at this directory, it’s flagged for no reason” > model starts running ls commands to find the directory in my ~/Projects > the exact second it looked at the onnx-tempopt directory of the project > blocked

What Should Happen?

There needs to be some kind of fallback for these situations, I would say something similar to the envelope encrypted thinking blocks but with the classifier sidecar model.

Flag the improper block > sidecar model reviews in an encrypted block so it can’t be used as a negative technique in itself > the work clears as not actually within the fable 5 policy = no usage penalty our bad

the work flags as against the policy = usage applied sucks to suck

In my case it makes sense that it gets flagged, but it is quite literally useless in terms of large scale model distillation or dangerous work. I’m a huge proponent of more predicated classifiers and sidecar models. But 200/month to get rug pulled on the first AI code-work I’ve ever seen that’s above a 5/10 heart breaks meh.

Error Messages/Logs

Steps to Reproduce

1. Create a project for consumer hardware model inference, specifically something like the V++ - Peano - MLIR AIE IRON stack
2. Have the model do some of the more trivial work for exports that are simple, and that run with operators proven to fit on npu tiles without much friction
3. Then point the model at a low level programming task, this specifically is not a training or finetune task — just an inference runtime task
4. Let the model do all the work, explicitly say you’re letting it rip to see capability and will only describe end goals
5. itll create work that it blocks itself

Claude Model

Other

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.173

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

Other

Additional Information

Arch Linux - Foot terminal