A local research assistant worked better with a separate checker
A few weeks of running a research assistant showed that the goal was to keep work on local hardware, avoid sending data out, and answer questions that need several steps without making up sources. The first setup used one with a search tool in a . It was cheap and quick to set up, but it was removed first. Once a question went beyond about two steps, the model filled the context with its own earlier reasoning and then started agreeing with itself.
By the third step, it could confidently invent sources. A larger context did not fix the problem; it only delayed it. The setup that stayed used two . One model handled research, while a smaller verifier checked each claim against fresh sources.
The important rule was to keep the verifier from seeing the researcher model’s reasoning. The verifier only received the claim and had to ground it independently. The setup was modeled after Apodex, which separates the reasoner from the verifier. Its open mini model is described as 35B A3B, with only about 3B active per token, making it a good shape for a verifier that does not need heavy work on every step.
Key points
- The goal was a private research assistant running on personal hardware.
- One plus search was easy to set up but failed on deeper questions.
- A bigger context did not stop the model from inventing sources.
- The working setup used one research model and one smaller verifier.
- The verifier worked best when it saw only the claim, not the research model’s reasoning.