state-vs-realitylisted
Install: claude install-skill RICK-ORPHEN/agent-discipline-skills
# state-vs-reality — reconcile the declaration with the behaviour
## 0. One line
**Look at what the system does, not at what the value says.
Then count every reader.**
## 1. Why this exists — eight defects measured in a single day
Every one of them had the same shape: *the source of truth was correct, and nobody was
reconciling it.* Not one was technically hard. **They were all failing silently.**
| # | Declaration | Reality | Damage |
|---|---|---|---|
| 1 | three vocabularies coexisted in an `industry` field | the signup form stored free text verbatim | **6 of 8 accounts had received no industry term at all** |
| 2 | `priceMonthlyJpy: null` meant "price undecided" in one module and "declared free" in the checker | every account used it free, ungated | **matched neither declaration** |
| 3 | middleware honoured `expires_at` | **the cron did not** | expired grants kept working |
| 4 | reads were normalised by `normalizeIndustry` | **the write path (a settings dropdown) was not** | 4 accounts lost their term on a single save |
| 5 | an integration reported `status: active`, `reconnect_required: false` | **failing every hour** (86 times over 26 hours) | nothing surfaced; nobody noticed |
| 6 | "one alert per incident" should be enough | **silence for a month** (open, unacknowledged, 1,401 occurrences) | before that, 22 alerts an hour for the same incident. It swung between both extremes |
| 7 | the delivery ledger records everything | reopening reset the stage to 0 and **c