48-Hour Test Reveals How Token Budget Limits Cause AI Log Summarizers to Mislead

A developer running automated overnight log summaries on a free AI model noticed the tool confidently reported a root cause that did not match the actual incident logs. Over 48 hours of testing, it emerged that the model was only ever processing the most recent portion of logs due to a fixed token budget, causing older, critical context to be silently dropped. As the pipeline ran longer, compressed digests of previous summaries replaced raw log data, meaning errors in early summaries became unquestioned premises in later ones. A control test confirmed the issue: feeding the same logs in three different orderings produced three entirely different, equally confident incident explanations. The author proposes a two-pass digest approach that first summarizes log chunks individually and then tracks how many early digests get evicted before the final summary is generated, making context loss visible rather than hidden.