Pleiads Search milestone: 5,540 verified real-time points
Watch the complete 5,540-point run on YouTube, or use the embedded player in the evidence and media section below. The original MP4 and replay evidence remain available.
The research question was whether legal-input continuations of an existing policy could improve the Project Pleiads laboratory milestone. E006 screens schedules in accelerated play and separately confirms promising schedules at native real-time speed from a clean reset. The controller is native5 Automated Policy Search, with zero model calls. This is separate from the Perception/Qwen branch.
The closed real-time run pleiads-20261005-120227-03ddb1 scored 5,540 points, with 208.601088 seconds of gameplay survival and 211.943424 seconds of emulated duration. The website checkpoint read when this candidate was prepared contained 5,510 points; the difference is +30 points. The maximum is a laboratory milestone, not a world-record certification or evidence of Qwen performance.
The policy is a continuation of pleiads-20261005-093429-14a115, generated by post_landing_combat_restart_grid. Its first changed input frame is 12186. The original manifest retains the complete schedule, parent identifiers, source revision, timing command and legal-control flags. The outcome ended in observed game over at frame 12936.
The existing exact-frame Lua applied-input replay exited successfully and matched the terminal PNG. The publisher independently compared 2,156 of 2,156 captured frames, at capture interval 6, with no missing or mismatched captures. This is exact applied-input replay, not verified MAME INP playback. Ordered CSV rows preserve all applied inputs, including same-frame entries, without reordering.
Evidence: original manifest, metrics, ordered applied inputs, full trace comparison and all screenshot hash comparisons. Original benchmark PNG bytes are excluded from the data artifacts. One separately classified editorial terminal image and the complete video are submitted for human rights review.
The H.264 editorial video covers 211.90 seconds at 1920 by 1080, 10 fps. It includes laboratory overlays identifying the run, experiment and Search with zero model calls. Captures were sampled every 6 emulator frames; the replay comparison covers every scheduled capture. No intentional cuts or speed-up were applied to this complete record video. Audio is absent. The selected gameplay media depicts Pleiads (Tehkan, 1981); game software and extracted assets are not distributed.
This is a selected result from adaptive search with related policy prefixes, not an independent holdout or a causal comparison. Input fractions do not locate the player or identify collision boundaries. A single reproduced trajectory cannot establish generalisation, a trained phase detector, VLM benefit or completion of the 279,100-point research target. The next step is to test additional continuations while preserving the same real-time and replay verification requirements. Publication and gameplay-media rights approval remain human decisions.
- Run ID
pleiads-20261005-120227-03ddb1- Experiment / controller
- E006 / native5
- Score
- 5,540 points
- Survival
- 208.601 s (emulated game time)
- Real-time / replay
- Confirmed / exact recorded-input replay
- Screenshot matches
- 2,156 / 2,156 captured samples
- Original manifest SHA-256
0f6f7552c1f050b592be7d220438c17b0959fd2167fb69e107e1b7e56afc3f9d- Evidence files
- Original manifest · Ordered actions CSV · Replay comparison · Screenshot hashes
All captured samples were checked, at one capture every six emulated frames. Screenshot bytes and game software are excluded. CSV preserves the original TSV action order. This is a laboratory milestone, not a world-record claim.
This is a laboratory milestone, not a world-record claim.
E006 · Editorial revision 2 · Published from reviewed checkpoint bundle-20261005T123958Z-df0bf332.