What happened
In August 2026, with 142 tests passing, OuranoOS was installed on a tablet and walked through the way a user would. Seven problems:
- Claims in the UI about features the backend did not provide.
- A 5 TB quota displayed as the user's — invented.
- A storage-cost simulator 17× too cheap.
NaNon screen, from dividing usage by a quota of zero.- Layout overflows at the device's real dimensions.
- (Found earlier, same app) a login path that never issued a session token.
- (Same) "multi-provider replication" writing every copy to the same device.
Why the tests could not see them
- The claims were strings. No test asserts that a sentence is true.
- The quota was a default in a model; tests used the same default.
- The simulator's constant was wrong in the same place the tests read it from.
- The
NaNneeded a zero quota, which no fixture produced. - Overflows need real device dimensions; unit tests render at a default size.
None of these is a failure of the tests. They are the class of problem tests do not address: is what the screen says true, and does it hold on the hardware?
What changed
- Finishing a batch now includes launching the app and walking through it. Written as a rule, applied to every ecosystem app since.
- Overflows became a test (Engineering Notes #9), because that one can be tested.
- The simulator's constant is derived from the storage market's price, not typed.
- How to drive an app over
adbfrom the command line was documented so the walk-through is reproducible — including the discovery that the phone refuses installs and screenshots while locked.
The pattern
The Ailoos mobile verification found random weights, a month of failed uploads and a tokenizer mismatch on a phone in one session. The Aephoron Android port found four traps only on the device. Eoonia's deep links only failed cold on hardware. The device is where the joins are tested, and the ecosystem's recurring failure is at the joins.
From the OuranoOS device review, August 2026.