openttd_transport_ai: cannot reproduce DS-V4-Pro's score
Hi, while reproducing the EdgeBench leaderboard we found that every model we ran scores 0 on openttd_transport_ai (all submissions crash), while DS-V4-Pro is reported at 15.2. Our setup is verified bit-for-bit official (work image digest matches Docker Hub; internet: false), and control AIs (pure-sleep) survive the full 20 years in your judge image, so the engine is fine.
The blocking issue seems to be the task setup: the docs say OpenTTDLab is "already installed" for local testing, but the work image ships no OpenTTD binaries and internet: false prevents downloading them — agents must write a full Squirrel AI completely blind, with only pass/fail feedback. (This even hides the fact that the bundled baseline AI itself crashes at tick 1 on an API call that doesn't exist in NoAI API 13.)
Could you confirm whether the published DS-V4-Pro score was produced under exactly these conditions, or with a testing-capable variant (e.g., binaries pre-cached in the work image)? Thanks!
Hi,
Thank you for the detailed and careful review. After inspecting the images and the run logs on our side, we can confirm your finding: the work image ships no OpenTTD binaries, so the agents were effectively writing their Squirrel AI blind, with only pass/fail feedback and no way to test locally. The previously published DS-V4-Pro score (15.2) was indeed produced under exactly these conditions — it was not run with a testing-capable or binary-pre-cached variant.
This was not our intended setup. We plan to fix the work image (bundling the OpenTTD binaries so OpenTTDLab local testing works as the docs describe) and re-run the affected models to correct the scores in the near term.
Given this, you have two options for your reproduction:
- Keep the blind setup and compare against the original (blind-run) scores — these are directly comparable to what is currently on the leaderboard.
- Wait for our fix, then compare against the corrected scores produced with the binaries available.
We'll follow up here once the updated image and scores are ready. Thanks again for flagging this.
Hi,
Thanks for the quick confirmation and the transparency — much appreciated. We'll keep the blind setup for now and compare against the original scores. Looking forward to the fixed image and updated scores; hope the re-run goes smoothly. Thanks again!
Best regards