Step 16: tool results

{
  "note": "The audit documents the measurement problems directly (sparse compute and horizon overlap, a non-unit slope across METR horizon versions, 73 percent of substantive quantitative events from one programme). It does not survey published forecasts to count how many intervals leave these out, so the step from the problems to the frequency of the omission rests on the author's characterisation of common practice rather than a tally.",
  "support": "overstates",
  "deployment": "Closing normative point of a single-author audit of the public frontier-AI measurement record, arguing that forecasts should report instrument, protocol and provenance choices.",
  "instance_id": "82181a16-ae92-4ffe-8e60-37d1d2df7875",
  "source_kind": "preprint",
  "source_read": true,
  "worth_reading": false
}