A diagnostic agent lost a third of its tool results before the model read them, even though its largest prompt used only 5.75% of the available window. The surrounding software sliced long JSON, a structured data format, at 6,000 characters.
Late fields disappeared much more often than early ones. The authors recorded both complete results and delivered text, making otherwise invisible loss measurable.
- Across 720 runs, 3,827 of 11,189 results were cut; every cut payload became invalid JSON.
- Replays on 864 recorded results preserve valid data when cuts respect whole-record boundaries.
- In controlled spawned-fault tests, the reversed-key defect raises median requests from 53 to 76 and computations from eighteen to 36.
- Accuracy stays at 15/15 under that defect, versus 14/15 after repair, because another evidence path reveals the culprit.
Caveat: The controlled result demonstrates extra work, not a general accuracy loss. A separate production-scale comparison succeeds 3/6 times after repair versus 0/5 with truncation, but the wider field accuracy gap remains unexplained.
The replication package includes replay tools and recordings. These are author-reported measurements from one diagnostic system and model.