29 Hours of Silent Maximum-Depth Reasoning

Two days ago I wrote up fixing this same card's freeze-on-boot problem. Same week, a different setting on the same machine: I capped how hard the model should "think" before answering. It looked fixed. It wasn't.

Found this out in three passes. Each one looked like the last word.

Pass one: the model refused outright. The model accepts a few named reasoning depths — short, moderate, maximum. Sent "moderate." The server rejected the request outright. First fix: turn the setting off entirely, let the model pick on its own. The error disappeared. Called it done.

Pass two (the one that looked real): the error went away, the setting never actually worked. Two days later, read the server's own documentation line by line instead of trusting the missing error. Found the real cause: the server only recognizes three reasoning depths, not the five everyone assumes. "Moderate" is one of the three — the level being sent by default wasn't. Turning the setting off hadn't fixed anything; it had just stopped reporting the failure. The server was silently defaulting to its deepest, most expensive reasoning level on every request, with zero signal anything was wrong.

Pulled the request log for the card's entire life. 958 of 990 requests — 29.5 hours, essentially since the card first came online — ran at maximum depth, uncontrolled. Not one error. Not one warning.

Pass three: the actual fix. Told the server exactly which of its three supported levels to use, instead of the five it doesn't. Checked the log on the next request: "moderate." Exactly as intended from the start.

What I took from it:

→ An error that disappears isn't the same as a setting that works. The second fix removed the message, not the cause → 958 of 990 requests isn't an edge case — it's nearly the card's whole operating history. A silent setting stays silent for an hour or a month, no difference to it → The server's own documentation had the real answer in five minutes — faster than a second blind guess