The 9-23 System Audit memo is on manual hold: the poller never pulls it, done does nothing, and Claude reads your ink only when you say so. Today: the difference between a routing fence and a work queue, and the one sentence that releases the audit.
Ten documents on the tablet are held this way. If you inked the audit's seven batch boxes expecting the usual overnight application, nothing happened, by design. The paper says so on page one, but the rule is easy to forget when every other FOR REVIEW doc applies itself.
Two routes exist. Legacy: a document tagged done is pulled at the next 15-minute poll, composited with your ink, and handed to a headless Claude that edits files, writes decision records and can start downstream work (a GO on a briefing starts a 40-minute pursuit). That is an application workflow, not an inbox.
Manual hold: the exact visible name is registered in PDB/config/remarkable-routing.json with an owner and mode manual_hold. rm_pull.py, rm_process.py and the retention sweep all check that file first and skip the name. The ink stays on the tablet. When you ask in chat, Claude fetches only that document's archive, composites it into the project's own returns/ folder, transcribes the marks, and proposes changes. Applying them needs your explicit direction in that conversation.
Why the audit uses it: the seven batches authorise real code changes (alarm rewiring, shelving skills, retiring jobs). The memo was written so that a tick means 'approved to run in a dedicated session', and the hold guarantees a human sentence sits between the tick and the run. A fence is not a queue: nothing waits behind it to execute.
List the held names, then check whether the audit has been read.
python3 -c "import json
d=json.load(open('/Users/tom/Claude/PDB/config/remarkable-routing.json'))['documents']
[print(k,'-',v['owner'],v['mode']) for k,v in d.items()]"
ls '/Users/tom/Claude/FOR TOM/PROJECTS/SYSTEM-AUDIT-2026-09/'
A held document waits for your sentence, not your tag.