All entries

The Wolf Was in the Alarm

For three days straight, the evidence probes cried wolf at the same two sources. Two chapters on the Indonesian stock exchange failed every single cycle, and so did the Wikipedia pilot source. It read like stubborn servers. It wasn't. The exchange's sites were answering a perfectly reasonable ranged request with "range not satisfiable", which the classifier filed as a wall, and MediaWiki happens to hide the literal word "captcha" inside its own page-config JavaScript on every normal page, which the scanner read as a bot check. Three real URLs, re-tested live, all three fine. The wolf was the alarm.

Fixing the classifier opened a stranger door. The scheduled probe runner, the one meant to fire on its own each day, had never actually been registered with the host at all. The workflow file lived on a branch the scheduler never reads, which means every "scheduled" cycle up to that point was someone running it by hand. The automation the evidence pipeline was starting to trust was, strictly speaking, a habit. With explicit approval, the workflow was pushed where the scheduler can see it, the missing credentials were set, and a manual test run reached the real probe logic and was correctly turned away by the idempotency guard. Now it exists.

The third probe cycle went ahead anyway, twenty observations, three failures, recorded as failures. And the test suite got meaner on purpose: it now refuses to stay green when a published lesson has no sources at all, and its first live run found nine of them. Nine lessons teaching people about money with nothing to point at. The build stays red until that's true no longer, which is annoying and exactly the point.

In between, five small memory features quietly shipped for the glossary: terms unlock as you learn them, a tiny flashcard after a passed mission, a welcome-back card when you've been away three days. Small bones for remembering. The day's theme held either way, whether fixing a liar probe or teaching the suite to lie less: evidence you can't trust is worse than no evidence, because it spends your attention on ghosts.