Journal
Crash-free is not a north star
Crash-free session rate is a comforting floor. It is a terrible ceiling. Teams in Thailand still present it as if a rising percentage meant the product felt well. It often means only that the process did not die loudly enough for the OS to file a report.
On dual-SIM devices — ordinary here, exotic in most imported playbooks — a first session can evaporate when the radio flips, when a watchdog never fires, when an upload fails with a swallowed exception. The session is “crash-free.” The user is gone. If that user was trying to submit a KYC photo or a parcel scan, your north star just congratulated itself on a silent failure.
In Signal Architecture Studio we force an unbundling. Crash-free stays on the board as a veto: if it collapses, nothing else gets to speak first. It does not get to be the sentence you lead with when the rest of the ladder is on fire in quieter ways. Beside it we want a job-completion rate that actually names the job, and a silent-failure counter for the paths that historically die without a crash.
The objection I hear in Bangkok rooms is that silent failure is “hard to instrument.” It is. That is not an argument for leaving it off the slide. It is an argument for writing it into the event contract with an owner. If no one owns “KYC photo reached the server,” you do not have a stability metric. You have a crash dashboard and a hope.
Imported templates make this worse. A composite “stability index” that weights crash-free with ANR and a jank percentile will happily rise while photo uploads die. Executives like a single number. Composite indices exist to soothe that preference. They also hide vetoes. The Mobile KPI Atlas is a veto stack for this reason: lower layers may forbid you from celebrating the layer above.
A practical Monday shape, if you must take something from this note without enrolling: three numbers, not one. Crash-free sessions. First-attempt job completion for the job that makes you money or keeps you licensed. p75 time-to-interactive on a 4G trace you actually captured outside the office. Speak one limitation out loud — “we still do not see radio-flip drops as crashes.” Then sit down.
If your board still leads with crash-free alone, you are not lying exactly. You are answering a narrower question than the room thinks it asked. That gap is where Thai consumer apps lose trust, and where faculty red pens spend most of week two.