What changed.
Hydra merged benchmark fixes (#2898) at 12:23 UTC on October 9, 2026. The change updates how its benchmark reports present memory and when they show slope and colour signals. Documentation and monitoring tests were updated as well.
The memory figure now describes average live heap during the load. Previously, it described the peak since the start. Those measures answer different questions, so a figure from before the change and one from after it are not like-for-like, even if both appear under a memory heading.
Why it matters.
People using Hydra benchmark reports to compare runs or size deployments need to check which memory measure each report uses. A lower reported number after this change would not, by itself, show that Hydra uses less memory.
The slope row now requires eight snapshots in its window, up from four, so scenarios with only a few large snapshots skip it. Colouring also requires measurements from at least three of four pairs. These rules should reduce trend signals drawn from thin data, although the record does not define what the pairs represent.
Watch the next Hydra benchmark report for a clear memory label and for how the revised slope and colouring rules appear in practice. Its documentation may also clarify whether peak memory is still reported anywhere.
Upstream references and independent checks
Digest dated 2026-10-10 · upstream model sonnet. Direct links are matched to all 1 upstream source IDs.
- 1benchmark fixes (#2898)Direct upstream source ·
3128f185033e2f4555dedafb6eb7697b81ebddfe
This Research brief was generated by GPT-6 Sol from a dated upstream research digest. It has not received the source-by-source human review required for Reviewed analysis. Material limit: The supplied evidence is one merged repository commit and contains no benchmark figures. It supports a change in measurement, not a conclusion that Hydra became faster or uses less memory.