First principles is a debugging tool, not a philosophy
Everyone quotes it. Almost nobody uses it at 2am when the retrieval pipeline returns garbage.
People talk about first principles like it is a personality. Something you have, like being tall. I think it is closer to a wrench. You pick it up when the normal tools stop working, and you put it down when the thing is fixed.
Here is when I actually reach for it. DocMind was returning confident, well-written answers that were wrong. Not slightly wrong. Wrong about which document the fact came from. My first instinct was pattern matching: search for 'RAG hallucination fix', try a bigger model, lower the temperature, add a system prompt that says please be accurate. That is not first principles. That is copying the last person who had a problem that looked like mine.
Break it until the pieces are boring
First principles means asking what has to be true for the answer to be correct, and then checking each of those things separately. For a RAG answer to be right, four things have to hold. The right chunk has to exist in the store. The retriever has to find it. The reranker has to keep it near the top. The model has to use it instead of its own memory.
Once I wrote that down, the problem stopped being 'RAG is hallucinating' and became four small tests. It turned out the chunk existed and the retriever found it, but the query was so vague that six other chunks scored higher and pushed the right one out of the context window. The model never saw it. It was not lying. It was answering the question with what it had.
The fix was not a better model. It was a state machine that rewrites the query when retrieval confidence is low, then tries again.
That single loop cut the hallucination rate by about 40% across 200 test queries. It came from decomposition, not from cleverness.
Why it feels slow and is actually fast
First principles feels slow because you are not doing anything visible for the first twenty minutes. You are writing down what must be true. Pattern matching feels fast because you are typing. But pattern matching has a hidden cost: when it fails, you learn nothing, and you have to start again with a new pattern. Decomposition fails informatively. Every test you run either confirms a piece or points at the break.
So I do not think of first principles as a way of seeing the world. I think of it as the thing I do when the world stops making sense and I need to find out which of my assumptions is the broken one. It is a wrench. Use it when you need it, and get good at knowing when that is.