Benchmark Radar Day 76: Hold Out Whole Recordings

Day 76 of Benchmark Radar. The scan fetched 1,316 items, published 547, and marked 234 as recommended.

Author: Koutian Wu; GitHub: ktwu01

Benchmark Radar is live. Open the dashboard, star the repository, or read the paper. The Hugging Face dataset provides the data.

The milipoint-runs artifact revisits MiliPoint’s overlapping radar windows. Holding out complete recording runs reduced reported identification accuracy from about 94% to 37–39%. When training and test windows share frames, a random split can reward recognition of the recording rather than transfer to a new run. milipoint-runs: run-disjoint evaluation of the MiliPoint millimeter-wave radar dataset.

Inspect AI 0.3.278 adds task and sample descriptions to logs and data frames, extends tool review to handoff calls, and fixes context-compaction, token-counting, and retry issues. A harness update can therefore affect both the audit trail and the accounting of an evaluation. UKGovernmentBEIS/inspect_ai 0.3.278.

Why this matters

Record both the split definition and the harness version. A familiar dataset name does not guarantee a comparable measurement.

Scan coverage

This is a keyword-filtered discovery feed, not a representative survey of AI research. The archived briefing used 142 of 547 captured evidence records. These findings summarize its cited releases; discovery dates do not necessarily equal original publication dates.

Data and briefing: 2026-10-10 archived snapshot.