llm max tokens truncation: the json repair hid the cut

A model stops writing when it hits the max tokens limit. The API reports this in the finish reason, but many pipelines never read it. I saw one parse the cut text straight into JSON. The cut left an open array and a missing brace. A lenient repair library closed both and returned valid JSON. The parser was happy, but the record lost its last line items. Nothing failed, so nothing alerted. The trade-off is real. Repair saves a call when the model adds a stray comma. It also turns a hard failure into quiet data loss. I check the finish reason before any parsing. A length stop is a failed call, not a partial answer. I retry with a higher limit or split the task into smaller pieces. I log every repair as its own event, with the original text. Then I measure three things per release. I track the share of calls that stop on length. I track how often repair changed the output. I compare field counts against the source document.