I’ve create an LLM that cannot hallucinate, not even in theory, but before you jump to your guns, hear me out how it works;
The LLM generates Hyperlambda
Every single function invocation in the code is verified to exist
If the LLM hallucinates non-existent functions, it tries again, 2 more times, before it gives up and returns an error
The whole point being that this is a 10ms job, integrated into my “generate code endpoints”, resulting in that the process is 100% automatic. Interestingly, due to the “failed, repeat again” loop, my evals for the LLM increased almost 10pp.
It can still hallucinate “intent”, such as returning rows from wrong database, etc - But it cannot hallucinate non-existent functions or concepts - Not even in theory …
It’s good if you have 2 PhDs : physics and ML, then you definitely will get it. I only have 1.5
It’s in early stages but if you wanna join let me know… I do this in my free time and want to pull in
the top players… so I need to build a community for this. You some of it right… but not the part from which it will be super AGI in your pocket… as said 1.5 PhD at least for understanding - or 150 IQ for believing - but the numbers dont really lie…
I’ve seen similar results running a similar agentic loop locally. Don’t worry too much about the semantic pushback on the word “hallucination”—the core mechanism of using strict schema validation and immediate feedback retry loops is what actually makes agentic code generation work.
Keep the underlying code to yourself and just build massive apps with it.
I’ve been running a local Qwen 3.8 27B model (Q5_K_M) through an agent framework for weeks now, building out a full CAD application. My local setup uses those exact self-correction and context-reset loops, and the scale speaks for itself:
92 active sessions logged across the project
10,627 message/tool turns executed
14.3+ million characters (~3.5M tokens) processed in history
~35k–40k average context size per session
Catching structural errors at the tool level and instantly feeding the trace back to the model is the only way to keep an LLM from losing its way during heavy multi-file refactoring. It works—just keep heads down and keep shipping.
