An overnight AI pass over a day's notes, capped and checked

Related notes drift apart, duplicates pile up and two notes can disagree about the same fact. I wanted an overnight pass that finds these without giving a model free rein over my records while nobody watches.

Each night Dreaming pairs notes that look related and asks a small judge whether each pair is related, a duplicate or a contradiction. In v1 it changes no note unless I switch writes on, and then only two front matter fields. Everything else is a flag or a proposal in a morning report. It never deletes, renames, moves or sends anything.

A launchd job starts it at 04:30. The judge is one tool-less Claude call per pair, on Haiku, with no MCP servers and a per-call spending cap. It is deliberately not a full managed agent run, which could still be going at 07:00. A night allows at most 30 calls and $1.50, usage is checked before every call, and pairs over the cap carry to the next night. The judge stops at 06:40 so the report lands before the morning job.

The judge's output is data, not instruction. Notes reach it fenced as data, and text in a note worded as an instruction is compared, not obeyed. Each verdict must quote one complete line from each note exactly, or it is thrown away. The code re-checks every quote against the file and rejects replies with unexpected keys. Live calls and writes are refused inside a managed agent, so only the scheduled job runs live.

The main script is 2,248 lines of Python with a 1,153-line self-test. With writes on, it makes at most 10 a night, each a compare-and-swap on the file's hash, logged before and after, and it skips notes someone else edited in the last 12 hours. Undo restores a field only if it still holds what Dreaming wrote, then freezes that rule for 7 days. The report is at most 12 lines plus 5 proposals, and one line of it reaches four other surfaces, including the daily email.

Replayed against 90 obligations closed over two weeks, the proposed close rule would have made 0 of them. Matching my typed words in session logs to open items was right 12% of the time. So v1 reports rather than edits. The first scheduled night ran on time and reported 0 changes, 0 conflicts and 6 proposals. Writes stay off until a week of dry runs is reviewed.

Most of what a night reads produces nothing yet. On the test night, 239 of 326 files gathered had no output, and using them is a v2 decision.