Scoring Partial Progress in Agent Evals: A GDPx-Inspired Harness in JavaScript
A benchmark called GDPx exists, SemiAnalysis released it, and the reported top score is 12%, with the leading model identified as Opus 5.5. That is what the reporting establishes. What 12% means depends on whether a long-horizon task pays anything for getting most of the way there, and the post could not reach the GDPx task set or its scoring rule. So this video covers the harness the post builds in dependency-free Node JavaScript instead: decomposing a task into checkpoints that can be checked without reading the agent's prose, weighting those checkpoints by dependency depth so clearing the first three of five earns six of fifteen rather than sixty percent, and blocking a checkpoint when its parent failed without ever removing it from the denominator. The scores shown come from scripted policies in a stub sandbox, not from any frontier model. Full write-up: https://hackyjs.com/posts/scoring-partial-progress-in-agent-evals-a-gdpx-inspired-harness-in-javascript