https://koyoojin.pages.dev/log Why I decided what I did and what I gave up. More about what I reversed than what I made. en Fri, 25 Sep 2026 00:00:00 +0900 https://koyoojin.pages.dev/posts/it-was-the-rhythm https://koyoojin.pages.dev/posts/it-was-the-rhythm Fri, 25 Sep 2026 00:00:00 +0900 Told the writing didn't read as human, I suspected translationese; measured, the problem was rhythm and emphasis. A record of taking that rule out of one repository and reapplying it across the whole site. https://koyoojin.pages.dev/posts/whose-uncommitted-changes https://koyoojin.pages.dev/posts/whose-uncommitted-changes Fri, 25 Sep 2026 00:00:00 +0900 With several sessions, nobody can tell whose changes are left in the working tree. Work idle for 40 hours moved again seven seconds after my commit. https://koyoojin.pages.dev/posts/buildable-is-not-worth-building https://koyoojin.pages.dev/posts/buildable-is-not-worth-building Fri, 25 Sep 2026 00:00:00 +0900 I built a file feature twice and reverted it twice. Both worked, and no line in the test table asked who would use them. https://koyoojin.pages.dev/posts/a-number-without-its-settings https://koyoojin.pages.dev/posts/a-number-without-its-settings Thu, 24 Sep 2026 00:00:00 +0900 Rerunning a number published a month earlier, no single setting gave it back. The number had been recorded; the conditions behind it had not. https://koyoojin.pages.dev/posts/the-day-the-gauges-went-in https://koyoojin.pages.dev/posts/the-day-the-gauges-went-in Thu, 24 Sep 2026 00:00:00 +0900 A day of turning rule documents into checks the AI runs on its own. The ruler turned out wrong fourteen times, and what is still unproven is written down too. https://koyoojin.pages.dev/posts/thirteen-bytes-over https://koyoojin.pages.dev/posts/thirteen-bytes-over Tue, 22 Sep 2026 00:00:00 +0900 A write the device refused came back without an error. I found six places where failure looked like success, and made each of them say so. https://koyoojin.pages.dev/posts/the-example-would-have-wiped-the-card https://koyoojin.pages.dev/posts/the-example-would-have-wiped-the-card Mon, 21 Sep 2026 00:00:00 +0900 Every place I trusted the board vendor's example or its documentation, I had to go back. The cause of the sluggishness and the cause of the reboots were both somewhere other than where I first pointed. https://koyoojin.pages.dev/posts/picked-chopin-described-schubert https://koyoojin.pages.dev/posts/picked-chopin-described-schubert Wed, 09 Sep 2026 00:00:00 +0900 The catalogue was fine and so was the index. The small model chose one piece and wrote about another, and the first guard I added went blind in the same commit. https://koyoojin.pages.dev/posts/machines-idle-queue-growing https://koyoojin.pages.dev/posts/machines-idle-queue-growing Sat, 05 Sep 2026 00:00:00 +0900 The model's headline conclusion rested on one theorem, and that theorem came with a precondition. On a real line the precondition did not hold. What was wrong was not a parameter but the frame. https://koyoojin.pages.dev/posts/true-until-it-was-split https://koyoojin.pages.dev/posts/true-until-it-was-split Wed, 02 Sep 2026 00:00:00 +0900 I split one automation into daily, weekly and monthly. The code ran fine; six sentences describing that code became false without a sound. https://koyoojin.pages.dev/posts/one-line-of-model-name https://koyoojin.pages.dev/posts/one-line-of-model-name Sat, 29 Aug 2026 00:00:00 +0900 The model two apps ran on was retired. It looked like editing one constant; within a day both apps had broken, for different reasons. https://koyoojin.pages.dev/posts/a-ruler-still-growing https://koyoojin.pages.dev/posts/a-ruler-still-growing Wed, 26 Aug 2026 00:00:00 +0900 The model came out 3% off. What was wrong was not the model but the number I had taken as my baseline — and fixing the baseline closed the gap without changing a line of the model. https://koyoojin.pages.dev/posts/batches-are-not-always-full https://koyoojin.pages.dev/posts/batches-are-not-always-full Wed, 26 Aug 2026 00:00:00 +0900 There is a loss a static capacity model cannot see even in principle. A 3.5% throughput gain with no capital spend was sitting inside it — and that is what simulation is for. https://koyoojin.pages.dev/posts/open-not-clean https://koyoojin.pages.dev/posts/open-not-clean Wed, 26 Aug 2026 00:00:00 +0900 I built metrics out of filings anyone can download. Most of the pipeline's code is not arithmetic — it is there to avoid the traps. https://koyoojin.pages.dev/posts/a-false-date-looks-like-a-date https://koyoojin.pages.dev/posts/a-false-date-looks-like-a-date Wed, 26 Aug 2026 00:00:00 +0900 On a page whose rule is “only what I measured”, five of the dates were ones I invented. Nothing caught them, a past day cannot be measured again — and the next day the tool doing the checking made the same mistake with the sign reversed. https://koyoojin.pages.dev/posts/printing-the-blank https://koyoojin.pages.dev/posts/printing-the-blank Wed, 26 Aug 2026 00:00:00 +0900 I collected the numbers I had measured into a new tab. What took the time was not the numbers but how to write the six rows that have no write-up. https://koyoojin.pages.dev/posts/a-dead-socket-and-a-thinking-model https://koyoojin.pages.dev/posts/a-dead-socket-and-a-thinking-model Tue, 25 Aug 2026 00:00:00 +0900 The address we were using stopped answering — and with the response arriving all at once, that produced no signal at all. I never found the cause; I made the system speak instead. https://koyoojin.pages.dev/posts/the-stack-did-not-move https://koyoojin.pages.dev/posts/the-stack-did-not-move Mon, 24 Aug 2026 00:00:00 +0900 A crash from app review, diagnosed twice and fixed twice — and all three times it came back on exactly the same stack. The cause was not my code, and a fourth arrived after I thought every route was closed. https://koyoojin.pages.dev/posts/counted-what-i-kept https://koyoojin.pages.dev/posts/counted-what-i-kept Sun, 23 Aug 2026 00:00:00 +0900 The first question after adding a visitor count was: what about everyone who came before? There is no way to know — and that was a decision, not an accident. https://koyoojin.pages.dev/posts/perfect-score-wrong-piece https://koyoojin.pages.dev/posts/perfect-score-wrong-piece Fri, 21 Aug 2026 00:00:00 +0900 To stop recommendations inventing albums, I built a vetted catalogue — and found the verifying side wrong in both directions, one part of it citing itself as evidence. https://koyoojin.pages.dev/posts/graded-its-own-mistake-as-mine https://koyoojin.pages.dev/posts/graded-its-own-mistake-as-mine Fri, 21 Aug 2026 00:00:00 +0900 To honour a constraint — it must run without a key — I wired up the on-device path, and it ran end to end with zero keys. Then I measured it on the same recording, and the constraint had to go. https://koyoojin.pages.dev/posts/whether-the-grader-listens https://koyoojin.pages.dev/posts/whether-the-grader-listens Thu, 20 Aug 2026 00:00:00 +0900 I built two speaking-test apps, copying one from the other — but whether scoring reads a transcript or listens to the audio redrew the engine, the tabs, and what could be scored at all. https://koyoojin.pages.dev/posts/the-screen-has-no-authority https://koyoojin.pages.dev/posts/the-screen-has-no-authority Mon, 17 Aug 2026 00:00:00 +0900 I added a board to a static site. What can you trust the browser with, when anyone can open and edit it? Every feature returned to the same answer. https://koyoojin.pages.dev/posts/the-instrument-was-shaking https://koyoojin.pages.dev/posts/the-instrument-was-shaking Sun, 16 Aug 2026 00:00:00 +0900 The verdict kept flipping between samples. The model had not changed — the instrument was shaking. And what the fixed instrument then caught. https://koyoojin.pages.dev/posts/defaults-are-not-neutral https://koyoojin.pages.dev/posts/defaults-are-not-neutral Sun, 16 Aug 2026 00:00:00 +0900 Defaults for unknown values carry a direction. Reverting them one at a time, four of five pushed the same way — and together they multiplied the answer by 3.4. https://koyoojin.pages.dev/posts/empty-handed-beats-invented https://koyoojin.pages.dev/posts/empty-handed-beats-invented Sun, 16 Aug 2026 00:00:00 +0900 What an app does when there is nothing to recommend. Dropping the unconfirmed instead of flagging it — the same call, made in three different places. https://koyoojin.pages.dev/posts/pin-the-reader-first https://koyoojin.pages.dev/posts/pin-the-reader-first Sun, 16 Aug 2026 00:00:00 +0900 A scheduled job that had never once failed kept picking things nobody could act on. Not a bug — the criterion itself. https://koyoojin.pages.dev/posts/docs-dont-ride-along https://koyoojin.pages.dev/posts/docs-dont-ride-along Wed, 12 Aug 2026 00:00:00 +0900 Fixing one README turned up the same kind of lie in three more places. Code has several layers that catch a false claim; prose has none. https://koyoojin.pages.dev/posts/a-layer-not-a-screen https://koyoojin.pages.dev/posts/a-layer-not-a-screen Tue, 11 Aug 2026 00:00:00 +0900 Porting one splash screen to Android, I got it wrong three times — never the drawing, always the layer I put it on. https://koyoojin.pages.dev/posts/not-the-signature https://koyoojin.pages.dev/posts/not-the-signature Mon, 10 Aug 2026 00:00:00 +0900 Shipping one version as a test overturned the previous post's conclusion — and exposed a real failure in the release pipeline. https://koyoojin.pages.dev/posts/checks-passed-screen-wrong https://koyoojin.pages.dev/posts/checks-passed-screen-wrong Sun, 09 Aug 2026 00:00:00 +0900 I ran 88 automated checks while adding Korean/English switching — and both times, I only found the bug by looking at the screen. https://koyoojin.pages.dev/posts/one-number-bump https://koyoojin.pages.dev/posts/one-number-bump Fri, 07 Aug 2026 00:00:00 +0900 It started with a request to bump versionCode to 22, and ended with finding out the app was a debug build signed with a debug key. https://koyoojin.pages.dev/posts/who-could-know-what https://koyoojin.pages.dev/posts/who-could-know-what Fri, 07 Aug 2026 00:00:00 +0900 Four apps in three weeks: retracing how the work split between Claude and me, using comment density and the traces left in a permissions file. https://koyoojin.pages.dev/posts/warchronicle-battles-as-data https://koyoojin.pages.dev/posts/warchronicle-battles-as-data Tue, 04 Aug 2026 00:00:00 +0900 Getting to where one battle is one JSON file and everything else is calculated — and why frames had to be cut on the map. https://koyoojin.pages.dev/posts/classical-mood-cross-platform https://koyoojin.pages.dev/posts/classical-mood-cross-platform Tue, 04 Aug 2026 00:00:00 +0900 Rather than faking what the platform could not do, I chose to write down what I gave up. https://koyoojin.pages.dev/posts/fab-alert-beyond-sms https://koyoojin.pages.dev/posts/fab-alert-beyond-sms Fri, 31 Jul 2026 00:00:00 +0900 Adding RCS, MMS, forwarding, and self-distribution — how an alarm I built for myself became an app other people use. https://koyoojin.pages.dev/posts/skill-md-as-a-contract https://koyoojin.pages.dev/posts/skill-md-as-a-contract Fri, 31 Jul 2026 00:00:00 +0900 A job that starts from an empty session every time cannot remember context — it has to nail the context down in a document. https://koyoojin.pages.dev/posts/sms-alarm-always-fires https://koyoojin.pages.dev/posts/sms-alarm-always-fires Tue, 28 Jul 2026 00:00:00 +0900 Why I stopped leaning on a foreground service and moved the decision to the moment the message arrives.