The Context Bomb
In this benchmark, 64% of questions had raw records larger than a 32k-token context window. This synthetic example shows how truncation can hide an early result and how a question-aware projection can preserve it. Nothing is uploaded.
What's real here (and what's synthetic)
The 1.28M-token median (92% >100k, 53% >1M) is measured from our secondary external comparison (MIMIC-IV-on-FHIR, a FHIR-AgentBench re-grade) — single model (GPT-5.5), single substrate, single seed. The patient on this page is synthetic (we can't redistribute MIMIC); its token count is computed live in your browser with the exact o200k tokenizer, so the overflow is real even though the patient isn't.
The COVERAGE COMPLETE badge checks this local demo: it turns green only when the projection keeps every in-scope HbA1c result, including the first and last. The naive recency cap turns it red because it drops the earliest value. In the benchmark, 40 of 107 fit-but-wrong blunt-projection answers involved first or earliest questions; that association does not prove one cause. See the benchmark →
Give the model the relevant evidence, not the whole record.
Bonfire is being designed to compile bounded, cited slices. This page is a local concept demo, not a Bonfire product result.