Make the default executeConcept

Three weekly-review cards in a row have stalled — four weeks, fourteen days, and now seven. Each carries a written default for 'if you don't answer.' None of them has ever fired. That's the whole gap.

There's a real distinction between human-in-the-loop and human-on-the-loop, and it's not a matter of degree. In-the-loop means the system halts and waits for you — no answer, no progress, indefinitely. On-the-loop means the system proceeds on a stated default and you retain the ability to override. The first optimizes for never doing the wrong thing. The second optimizes for not stopping. Which one is correct depends entirely on how often the operator actually shows up.

For six months, in-the-loop was defensible — you were here daily, and a card sitting two days was a minor delay. That assumption expired yesterday. The cost isn't just the unanswered card; it's everything downstream. A stalled curriculum decision means unauthored lessons. A stalled adopt decision means a tool never evaluated. The Jul 26 card's curriculum decision lapsed with the calendar — those dates passed unrendered, and the decision became moot without ever being made. Not deciding is itself a decision, made by the clock, with no record.

You've already done the hard part. Every card carries an explicit default, chosen deliberately when you had the most context. The missing piece is purely mechanical: a date after which that default executes and the system records that it executed by expiry rather than by choice. And there's a clarifying test in it — if you wouldn't let a default fire after seven days of silence, it probably shouldn't be a decision on the card at all. It's either more consequential than a card can carry, or it isn't a real fork.

Worked example

The shape of an expiring default — data first, execution second:

# Today's card frontmatter, with the clock added:
# ---
# slug: weekly-review-2026-08-16
# decisions:
#   - id: curriculum-post-sabbatical
#     default: accept-as-authored
#     default_executes_at: 2026-08-23     # 7 days
#   - id: cadence
#     default: daily-continues
#     default_executes_at: 2026-08-23
# ---

# The sweeper (a cron, not an LLM — Automation Hygiene rule 5):
#   for each open card in FOR TOM/TTR/:
#       for each decision past default_executes_at:
#           record choice=<default>, via="expiry"
#           emit decision_answered event
#           notify: "defaulted after 7 days — reply to override"

# The evidence this is needed:
ls -lt "/Users/tom/Claude/FOR TOM/TTR/"
#   weekly-review-2026-08-09.html   <- 7 days, unanswered

• The default already exists in every card you write; this only adds a date to it. No new judgment is required at authoring time.
• Recording via="expiry" rather than via="ntfy-action" keeps the audit trail honest. Later you can count how many decisions you actually made versus let lapse — which is itself the useful metric.
• Deterministic script, not `claude -p`. This is bookkeeping against a date field; an LLM adds failure modes and no judgment.
• The notification after the fact is what preserves your veto. Defaulting silently would be worse than waiting — you'd lose both the decision and the knowledge that it happened.

Gotchas

• Defaults only work if they're genuinely the safe choice. An expiring default that ships something irreversible converts your silence into a deploy — pick defaults you'd be content to wake up to.
• Seven days is a starting point, not a law. Match it to your actual cadence; if you review on Sundays, a Sunday-to-Sunday window means every card gets exactly one look.
• Notify on expiry, always. A default that fires silently teaches you the system is unpredictable, and you'll trust it less than when it just waited.