Review week - the pen is the throttleReview

Review week. Five lessons, one number: how much the machine produced versus how much you decided. The gap is the system's real throttle, and this week's D3 is the first time the paper asks you to set it on purpose.

Monday (9-15): a usage cap is a state, not a blip - the machine's rate is bounded by budget as well as cron. Tuesday (9-16): the briefing loop has one manual station, and the alert measures the queue behind it. Wednesday (9-17): a secret scanner stopped a sweep for three days because a scratch folder was in a source repo - the machine will keep failing loudly until a human decides scratch-or-source. Thursday (9-18): seven releases in a week; reading them is a ten-minute Sunday job, not a daily one. Friday (9-19): a non-idempotent send doubled ten documents because a queue was never marked drained.

The common shape: launchd sets the production rate (05:15, every day, no matter what), and your sittings set the decision rate. When the two diverge, nothing breaks - queues just grow: 23 undecided leads, 29 unsent packages, 262 documents, 12 lessons. The alerts (automation-health count 21, jobhunt-briefing-return count 4, device-doc-count red for two weeks) are all measuring the same divergence from different angles.

Two honest levers exist. Lower production: D3 on the 9-13 paper proposes that the nightly stop presenting new leads while more than one briefing is unreturned - sweeps still run, findings still accumulate, nothing is lost, and the device stops filling. Or raise the decision batch: one sitting, all five briefings, blank boxes re-asked. 'Work harder' is not a lever; it is the thing the queue is already asking for and not getting.

Worked example

The three ratios, from the files that hold them.

python3 - <<'EOF'
import json, glob, re
from collections import Counter
# packages: staged vs sent
st = Counter()
for f in glob.glob('/Users/tom/Claude/JobHunt/leads/*/lead.yaml'):
    m = re.search(r'^stage:\s*(\S+)', open(f).read(), re.M)
    st[m.group(1) if m else '?'] += 1
print('packages  staged/drafted:', st['packaged'] + st['drafted'], ' sent:', st['sent'])
# leads: presented vs decided (last 7 days)
d = json.load(open('/Users/tom/Claude/JobHunt/leads/briefing-state.json'))['presented']
wk = [v for v in d.values() if v['date'] >= '2026-09-07']
print('leads     presented:', len(wk), ' decided:', sum(1 for v in wk if v.get('decision')))
# lessons: on device (needs bun)
import subprocess
out = subprocess.run(['bun','/Users/tom/Claude/PDB/scripts/rm-bridge/bridge.ts','list','500'],capture_output=True,text=True).stdout
print('lessons   unread on device:', out.count('Learning \u2014 LESSON'))
EOF

• Stage comes from each lead.yaml; 'packaged' and 'drafted' are both waiting on you.
• Presented-vs-decided reads briefing-state.json for the week; decision is None until ink lands.
• The device listing is the slow call (the whole list, every poll) - which is itself lesson 9-19's point.

Gotchas

• A queue that grows quietly is worse than one that fails loudly - the sweep that failed got fixed in a lesson; the 23 leads got an alert nobody answers.
• Throttling production is not giving up on the search. It matches the machine to the pen so what it produces gets read.
• Hold-by-default is your rule and it is right; it also means every queue above is yours to drain.