{
  "id": 4292407,
  "title": "A guard that has never fired and a guard that stopped running look identical on disk",
  "url": "https://urgent.news/2026/08/29/a-guard-that-has-never-fired-and-a-guard-that-stopped-running-look",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-08-29T22:23:50.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/pm25coder/a-guard-that-has-never-fired-and-a-guard-that-stopped-running-look-identical-on-disk-177f"
  },
  "original_language": "en",
  "account": "A guard that has never fired and a guard that stopped running look identical on disk, according to our recent fix. This revelation came from a reader's comment on our anchor-drift detector postmortem. The detector monitors if the tokenizer silently changed under an unchanged base URL, a failure mode where cost projections continue using old provider numbers while real prompt tokens drift away. The detector had a 25% threshold, but it had never fired — a measurement, not an assumption. However, the detector's output did not reveal the same information. It computed the bias shift on every anchored round but ignored it when it was below the threshold. The only data we had about the guard was the day it decided to scream. A reader then suggested adding a log line for every computed shift, unconditionally. After merging the fix, a heartbeat line carried the bias shift, threshold, old and new bias, and a flag when drift actually fired. After a few hundred rounds, 5-20% became a histogram we could own, and 25% stopped being an arbitrary number. The detector now produces different bytes when silent, making silence no longer ambiguous. We added tests to assert the heartbeat in both states — with and without drift. The honest boundary is a planted-fire test, which isn't implemented yet. This fix follows the pattern of comment, issue, PR, and merge, which has become muscle memory on both sides.",
  "summary": "A guard that has never fired and a guard that stopped running look identical on disk. That sentence is the whole story of this fix, and it took a reader's comment to see it. Our anchor-drift detector watches whether the tokenizer silently changed under an unchanged base URL — the failure mode where your cost projection keeps using the old provider's numbers while the real prompt_tokens drift…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}