Pelican: text-only painting results and replay

The fixed target is exactly:

Paint a pelican riding a bicycle.

No reference image is supplied. The model still receives the same two fixed 960 × 720 workspace cameras and flange feedback. All 11 models use AgentArk SimpleAgent, with seed 107, three connected attempts of at most 96 environment actions each, fixed brush/paper, infinite pigment, 120 Hz control, 480 Hz physics and time scale 1. A fresh canvas starts each attempt while the existing managed context, summaries and archives carry experience forward. Model weights do not change.

The completed independent collection contains 11 models, 33 attempts and 603 actions. Pelican's gallery and settings default to the benchmark author's personal ranking, supplied on 2026-10-20: Astra → Sol → Argon → Opus → Wonder → MiMo → Qwen → Kimi → DeepSeek → Doubao → Grok. This task-specific subjective recommendation is retained with attribution in the ranking metadata. It is not a numerical score or a human-consensus result. Submission reward checks the saving/submission workflow, not whether the pelican or bicycle is recognizable.

All 33 selected round-end PNGs are versioned under site/assets/media/pelican, with exact hashes and episode associations. Together with Golden Gate's 36 selected images, a fresh clone contains all 69 website assets. Original records, HTTP logs, unselected media and complete videos remain outside Git.

Actual historical settings

The portable results retain exact model IDs, source AgentArk commits, canvases, episode GUIDs, artifact hashes and limitations. These are historical settings, not an equal-compute comparison:

• Opus: retained successful execution used high, with tool_choice=auto, after max-mode failures and context-preserving recovery. Golden Gate's mixed max/high action history must not be substituted for this run.
• MiMo: thinking enabled, without an effort selector. The normal 20,480 output budget was temporarily raised to 65,536 for two provider requests after truncation. This exception is retained in the evidence; it does not silently change the standard template.
• Doubao: direct Volcengine Ark, doubao-seed-2-1-pro-260915, thinking enabled / high; no OpenRouter substitution.
• Astra: SimpleAgent's Codex subscription SDK backend, gpt-6-astra / max, without an API-equivalent hard output-token cap. Golden Gate used ultra, a different historical setting. The local SDK compatibility adapters are absent from the locked upstream package, so this repository does not advertise an unverified portable --model astra evaluation profile. Record replay does not require that transport, subscription login or a model API key.

Public evidence and coverage

The 11 complete records and 11 trajectories are pinned to dataset P90-RushB/AgentArk revision b16fb52e866ce6a33c52e287f04bfe4a77964453. Each file has a locked size and SHA-256 in the result manifest. No video, HTTP audit log, repeated request history, credentials or machine paths are embedded in this source release.

Use the results record for replay. All 603 executed actions are present there. The first-attempt trajectory exports for Claude Opus 5.5, Grok 4.7, Meta Muse Wonder, DeepSeek V4.1 Flash 1017, Qwen3.8-Max-Prime, Doubao Seed 2.1 Pro 260915 retain only their final 24 steps. Later attempts are retained; the result manifest reports each export's actual coverage. A trajectory window is not a complete request history or proof of how much context the agent retained. Public sanitized evidence preserves readable thoughts, actions and execution outcomes; it does not expose or guess encrypted reasoning.

The first nine models were re-recorded through the official Linux GPU replay entry point: all 27 round-end canvases matched their source pixels. Doubao and Astra retain their six native Linux GPU evaluation recordings; no additional physical replay is claimed for those two. All 33 selected native episode videos passed complete MP4 decoding, with frame counts checked against frames.jsonl and recorder manifests showing complete=true, truncated=false, encoder_exit_code=0. Episode GUIDs associate records, canvases and selected recordings; folder counters alone do not identify attempts.

These are historical local checks. A new machine or hosted Colab replay is a separate verification. Pixel equality across GPUs, operating systems or differing harness revisions is not promised.

Videos

The 11 published videos each contain one model's three attempts, in order, including calibration, failures, recovery and submission actions. They use native overhead, fixed side and pen-follow views. The follow camera is for the audience and was not an extra model observation.

播放速度：4× / Speed: 4× means four simulated seconds per playback second, independent of model/network waiting. Captions summarize and translate readable source statements in the first person. Plans are not presented as successful outcomes. Captions remain within their source step; brief explicitly labeled “字幕暂停 / Reading pause” segments support exceptionally short steps. The film is not an end-to-end latency measurement or the model's exact image stream. Golden Gate's published films use 6× and remain a separate collection.

Replay without inference

Open the Pelican Colab, select a T4 GPU, select a source model in step 5 and Run all. Keep MODE="replay". Step 7 prints AgentArk's existing viewer URL. The notebook embeds the same reviewed no-inference source bundle as the Golden Gate notebook and selects Pelican records and the text-goal configuration. Its extension was checked offline; it has not been claimed to pass a fresh hosted Pelican GPU execution. The existing Golden Gate hosted report retains only its stated scope.

For an installed local Linux GPU runtime, the same helper prepares a new isolated output directory without starting Unity or calling any model:

# Use the existing compatible environment and paths from install_runtime.py.
python scripts/colab_record.py download-record \
  --case pelican --model sol --destination '<new-input-directory>/sol-results.jsonl'

python scripts/colab_record.py prepare --case pelican --model sol --mode replay \
  --record-path '<new-input-directory>/sol-results.jsonl' \
  --run-dir '<new-pelican-replay-directory>' \
  --env-path "$AGENTARK_ENV_PATH" --mod-path "$AGENTARK_MOD_PATH" \
  --task-store "$AGENTARK_TASK_STORE_PATH" --render-mode existing-display \
  --graphics-api glcore33 --step-delay 0 --no-keep-viewer

python scripts/colab_record.py run --config '<new-pelican-replay-directory>/eval.yaml'

Replace sol with argon, opus, grok, wonder, deepseek, qwen, kimi, mimo, doubao or astra in both download and preparation. The helper verifies the selected record's immutable hash again before launching AgentArk's official agent_ark.ark_eval.run_replay; there is no second replay or evaluation harness. Fresh destinations preserve all original files. GPU mode keeps no_graphics: false; verify the actual Unity player.log renderer rather than inferring GPU use from hardware availability.

For optional native motion recording, pass --record-video --ffmpeg '<existing-absolute-ffmpeg-path>' to the helper's run command. Close the environment normally and perform the recorder integrity checks. inspect reports observed evidence but does not claim that MP4s were decoded or a runtime passed.

To prepare a new model evaluation, use python scripts/prepare_run.py --case pelican --model sol followed by the separately printed scripts/run_eval.py command only when inference is intended. Preparation alone is offline. Portable launch profiles and historical actual settings are distinct; recovery must retain the complete original context and physics boundary and must not be simulated by replaying a short trajectory window.