{
"source": "Captured 2026-07-31 on GitHub Actions ubuntu-latest (workflow run 30604713594), the same runner class the gate executes on. The previous baseline was taken on a developer machine on 2026-02-08 and was not comparable: it reported +104% on xas_group_benchmark_single purely from host differences, while autobk_stage_linear_direct had in fact become ~12x faster than it recorded.",
"benchmarks": {
"xas_group_benchmark_single": {
"median_ns": 3080937000.0
},
"xas_group_benchmark_parallel": {
"median_ns": 1492776000.0
},
"autobk_stage_legacy_lm": {
"median_ns": 4162000.0
},
"autobk_stage_linear_direct": {
"median_ns": 221000.0
}
},
"thresholds": {
"regression_percent": 25.0
},
"notes": [
"Threshold widened from 10% to 25%. These are absolute timings on shared CI runners, so part of the run-to-run spread is host noise rather than signal: across two consecutive runs of identical code, xas_group_benchmark_parallel moved 4.8% (1424.873 -> 1492.776 ms) while xas_group_benchmark_single moved 0.29%. 25% leaves roughly 5x headroom over the observed spread while still catching a genuine 2x regression.",
"Re-capture this file from a CI run, never from a developer machine, or the comparison measures the host instead of the commit.",
"The durable fix is to benchmark the merge base and the head within a single job and compare relatively, which removes the dependency on a committed absolute baseline entirely."
]
}