BNLGit
BNLGit doesn't fine-tune your model or swap it out. It changes the one thing that actually decides whether an agent gets it right: what the model sees. Here's the same question, the same model — without the layer, and with it.
The model reasons over structured, high-signal facts instead of pattern-matching across raw files — so it stops inventing behaviour that isn't there.
Every claim points to an exact line at an exact commit. You (or a CI check) can confirm it in one click — trust becomes checkable.
A fact block is a fraction of the tokens of the files it summarises, so more of the system fits in context — cheaper calls, wider awareness.
Summaries are tracked per blob SHA and re-derived on change, so the agent never quietly reasons over last week's code.
These illustrate the mechanism, not a lab result. The rigorous benchmark is in progress; we'll publish real numbers when they're earned.