What happens
You say what you noticed. A small tool — not the AI’s judgment — routes the capture: the target is a fact, and if it doesn’t match a registered ledger the command stops and says so rather than guessing. Your words land verbatim inside an immutable, timestamped machine block; the session’s structured reading of them is added after the block, never inside it. Fact layer and judgment layer stay separate by construction.
Your ratification is the write gate — the agent reads its understanding back, and your yes is what commits the capture to the buffer. Then one confirmation line, and the session returns to the interrupted work. Minimal disturbance is the point.
The buffer itself is the project’s deliberate RAM: a growing list meant to shrink. Entries dissolve into the project — a rule, a document update, an undertaking of their own — and a pressure gauge warns when the list runs full, without ever blocking a capture.
Why it’s shaped this way
Two failures taught the shape. An early capture lost its target because routing was left to interpretation — so routing moved to a tool, deterministically. And observations kept accumulating as permanent clutter — so the buffer was redefined as temporary memory, with dissolution as the success condition. A fix proposed from the field always aims at the generator, the kernel, the next release — never at patching the running instance from inside a capture.
The shipped file: fieldnote.md ↓ — verbatim, with the baseline’s own ledger roster inside. The format behind it is the Fieldnote Format; the buffer’s place among the user documents is its own story.