Trajectory Guard: agent trajectory corpora for "Beating Noise Is Not Beating a Cap: Learned Early Abort for Self Healing Code Agents"
收藏资源简介:
Trajectory corpora and audit sidecars for a study of learned early-abort policies in self-healing code agents, evaluated against a randomized step-cap frontier in- and out-of-distribution. 43 files, 13,022 step-level records, covering five models across three serving stacks and 14B to 550B parameters, on HumanEval, MBPP, MBPP+, BigCodeBench, and SWE-bench Lite. Each record carries the nine trajectory features available to an abort predictor at decision time, the eventual outcome label, and recorded rather than asserted model provenance. Audit sidecars carry every raw model reply. No one can regenerate a lane. The DeepSeek V4 Pro endpoint reached end of life on 2026-08-07, one day after that lane was harvested, and now returns HTTP 410. This archive is the only remaining evidence that those trajectories occurred. Absolute filesystem paths identifying the harvesting machine were replaced with neutral tokens of the same length, so the stored feature values remain exactly consistent with the text they describe. See README.md and MANIFEST.json. Analysis code: https://github.com/Exile404/Trajectory-Guard



