Retirement Ritual

Module 6, Lesson 6.5 · 90-day cadence · produces the capstone freeze · concept decision + recipe disable-revoke-archive

What this is. The once-a-quarter review every live scheduled task passes. Per-entry: cost reconciliation, five-signal check, idempotency audit, prompt integrity, decision. If the decision is Retire, the three-step ritual — disable, revoke, archive — runs immediately.

How to use. Copy the per-entry block into a new file for the day’s ritual: /capstone/automation-artifacts/retirement-ritual-<date>.md. Repeat the block once per register entry. Then run the aggregate checks once. Then answer the three closing questions.

If nothing ends up retired, the ritual was not honest.

Expect at least one retirement or redesign per quarter. Drift is the default; this ritual is the correction.

For each register entry, complete the following

Copy this block once per entry; fill in every row. Attach the completed worksheet to /capstone/automation-artifacts/retirement-ritual-<date>.md.

### Entry <N>: <Task name>

1. Cost reconciliation
   - Ceiling (per-run / per-month): $<N.NN> / $<N.NN>
   - Actual last 30 days (per-run avg / per-month total): $<N.NN> / $<N.NN>
   - Headroom: <percent under ceiling, or “BREACH if monthly > ceiling”>
   - Drift risk: <low / medium / high — what could cause the monthly number to rise in the next 90 days?>

2. Five-signal check
   - [ ] Named event — has the retirement trigger fired?
      Evidence: <specific observation>
   - [ ] Stale last-success — compared to my threshold (<X> days for this task):
      Evidence: <last-success date; gap to today>
   - [ ] Silent drift — are the artifacts still answering the original question?
      Evidence: <concrete observation from the last 2-3 artifacts — what’s being produced vs. what should be>
   - [ ] Ceiling breach — was the monthly ceiling breached in the last 30 or 90 days?
      Evidence: <Yes / No, with number>
   - [ ] Human-attention bankruptcy — is the aggregate review demand from all my tasks above 30 min/week?
      Evidence: <estimated weekly review time across entire register>

3. Idempotency audit
   - Current idempotency key: <as written in the register>
   - Can I name the key in one sentence, right now, without reading the register? <yes / no>
   - Does the artifact path actually use the key? Verify by listing files in the artifact folder; do they follow the pattern?
   - Any observed double-output in the last 30 days? <yes / no; if yes, describe>

4. Prompt integrity
   - Number of entries in the prompt-change log (last 30 days): <N>
   - Were all changes logged? <yes / no — honest>
   - Is the prompt meaningfully different from the version you first saved? <yes / no — and if yes, is the current version still doing the job you built the task for?>

5. Decision
   - [ ] Keep as-is.
   - [ ] Redesign (tighten prompt / adjust schedule / narrow scope / change cost profile). Note the specific change: <...>
   - [ ] Retire. Retirement reason (one of the five signals): <...>

6. If Redesign
   - Change to apply: <specific prompt/schedule/scope edit>
   - New prompt-change-log entry: <one-line>
   - New next-review date: <YYYY-MM-DD, within 90 days from today>
   - Expected effect: <what I expect to see different in the next three runs>

7. If Retire — the three-step ritual
   - Step 1: Disable the scheduler entry.
     - Scheduler: <Cowork tab (primary) / cron / launchd / Windows Task Scheduler>
     - Action taken: <specific — “Disabled and then deleted the Cowork-tab task ‘Morning brief’ at 2026-04-22 14:30”>
   - Step 2: Revoke credentials the task used.
     - List every credential the task held: <OAuth scope on Gmail label read; Gmail drafts-write scope; API key X>
     - For each, revoke or reduce scope: <action taken per credential>
     - Specific zombie token to watch for: <the one token that’s easiest to leave live>
   - Step 3: Archive the folder and update the register.
     - Moved `~/ai-architect-academy/automation/<task>/` to `~/ai-architect-academy/automation/retired/<task>-<YYYY-MM-DD>/`
     - Register row updated to: Status = “Retired <YYYY-MM-DD>”, post-mortem = <one sentence>

Aggregate checks (do these once after iterating through all entries)

Total monthly cost across all live tasks

Compared to total-suite monthly ceiling: $<N.NN>. Headroom: <percent>.

Total estimated weekly review time across all live tasks

Above 30 minutes? <yes / no>. If yes: which task will you retire to bring it back under?

Number of tasks retired this ritual

If zero: re-read Lesson 6.5 Block 4 and run the five-signal check again more honestly.

Number of credential revocations performed this ritual

These should match the number of retirements (plus any reductions-in-scope you did for still-live tasks).

Number of zombie tokens found (credentials still live for already-retired tasks)

Any non-zero number is worth noting in the register — this is the credential-hygiene signal Module 9 will build on.

Three honest questions to close the ritual

Answer in one or two sentences each. These go into the reflection paragraph at the top of the frozen register.

Which task am I most proud of?

Not “most sophisticated” — most useful, most genuinely leveraging my time.

Which task was I most wrong about when I built it?

What did I expect, what did I actually get, what did that teach me?

Which task do I suspect I’ll quietly let lapse in the next 90 days without going through this ritual?

Being honest about this now makes it harder to actually let it happen.

Every 90 days. Every entry. Disable, revoke, archive — in that order, no shortcuts.