aayambansall/delphi-benchmark-traces

Delphi executable-pilot agent trajectories

All 620 mini-SWE-agent trajectories from the preregistered executable pilot in Where Do Context-Engine Gains Come From? A Component-Level Decomposition of Repository Retrieval Under Matched Baselines and Audited Exposure (Bansal and Gangwani, 2026), together with the official SWE-bench harness verdicts and logs for every trajectory. The paper is at https://trydelphi.ai/research; code, per-case retrieval results, analyses, and the paper source are at https://github.com/aayambansal/delphi-benchmark.

What was run

• Tasks: 62 SWE-bench Verified instances (the paper's 62-instance C0 set).
• Agent: mini-SWE-agent 2.4.6, text-based action format, gpt-5.4-mini, step budget 50, 2.0 USD cost cap, official x86_64 instance images.
• Conditions (condition column): none (no seed), random (five non-gold files, seed 1042), delphi (top five files of the frozen Delphi ranking), hybrid_rerank_expand (top five of the fully matched conventional stack); each run twice (repeat 1 and 2). delphi_paths (Delphi's top five paths, no content) and delphi_chunks (paths plus the best-matching chunk instead of the file head) are the exploratory seed-interface arms, one repeat each.
• Seed block: identical wording across conditions, up to five paths with the first 40 lines of each file at the base commit, capped at 3,000 tokens.
• Outcome: verified repair under the official SWE-bench harness. 261 of 620 trajectories are resolved.

Layout

Path	Contents
trajectories.jsonl	one row per trajectory: run, condition, repeat, instance, model, exit status, API calls, cost, harness verdict, path to the trajectory file
pilot/runs/<run>/<instance_id>/<instance_id>.traj.json	full mini-SWE-agent trajectory (every model call and action, final submission)
pilot/runs/<run>/preds.json	submitted patches per run
pilot/eval/<run>/*.json	official SWE-bench harness report per run
pilot/eval/<run>/logs/run_evaluation/...	per-instance harness logs: patch.diff, eval.sh, test_output.txt, report.json, run_instance.log
pilot/datasets/<condition>/	the seeded SWE-bench inputs given to the agent under each condition
pilot/analysis-*.json, pilot/seed-content-analysis.json	pooled, per-repeat, seed-interface, and seed-content analyses
pilot/instances-round3-62.txt	the instance list
pilot/smoke-native/	a single smoke-test trajectory (not part of the pilot)

Runs are named <condition>-s50 (repeat 1) and <condition>-s50-r2 (repeat 2); s50 is the 50-step budget. Conditions in this release: delphi, delphi_chunks, delphi_paths, hybrid_rerank_expand, none, random.

Loading

from datasets import load_dataset
index = load_dataset("aayambansall/delphi-benchmark-traces", split="train")          # the trajectories.jsonl index

from huggingface_hub import hf_hub_download
import json
row = index[0]
path = hf_hub_download("aayambansall/delphi-benchmark-traces", row["trajectory_path"], repo_type="dataset")
traj = json.load(open(path))                                # {"info", "messages", ...}

Notes

Trajectories embed problem statements, file contents, and patches from the public repositories in SWE-bench Verified; that content keeps its upstream licenses. The trajectories, verdicts, and analyses themselves are released under CC BY 4.0. Hosted model calls were made with gpt-5.4-mini; API credentials are not included.

Citation

@article{bansal2026contextengine,
  title   = {Where Do Context-Engine Gains Come From? A Component-Level Decomposition of
             Repository Retrieval Under Matched Baselines and Audited Exposure},
  author  = {Bansal, Aayam and Gangwani, Ishaan},
  year    = {2026},
  url     = {https://trydelphi.ai/research},
  note    = {Preprint}
}