Recorded demonstration, not live inference. Six synthetic incidents and private tool routes, selected after exploratory screening. These are not held-out results, a general agent accuracy estimate, or a biological safety benchmark. This experiment is separate from the paper's benchmarks.
Loading recorded outputs…
| Method | Correct routes | Accuracy |
|---|
A useful next experiment
Hold the model, memory bank, and evaluation set fixed. Compare retrieving from the original question with retrieving from an intermediate answer. Then vary whether examples are prepended or injected into the trace. Measure total model calls, generated and injected tokens, retrieval overhead, and failures.
Use the existing four-method comparison runner. Use fresh output directories per configuration; cached runs are resumed by example index. The runner is a starting point, and further controls are needed to isolate query quality from additional model calls.