Reproducibility

What exact replay means

A deterministic replay re-executes the recorded action sequence using the same emulator, game version and relevant settings. Captured screenshots are compared at corresponding frames. The reported E005 replay matched every captured screenshot; this checks those samples, not a claim that every unrecorded frame was inspected.

The run record

Preserve the run ID, experiment version, policy/configuration, emulator version, timestamps, seed where applicable, score, timing, frame-based inputs, model calls, run class and replay comparison result. Hash all distributed files with SHA-256 and record size and provenance in a public manifest.

Latest verified record

Latest published verified Search record: 5,750 points. Run pleiads-20261005-131958-e6ceb4, zero model calls, confirmed at real-time speed. Exact Lua-input replay matched 2,239 / 2,239 captured samples.

Original manifest · Full replay comparison · Screenshot hashes · Record article and video

Published research evidence

The downloads now provide the full E002 numerical report and 4,800 inference outputs, coordinate annotations, E003/E004 sealed reports, and original E005/E006 manifests, ordered action sequences, metrics and replay screenshot hashes. The source-only laboratory release includes original Python orchestration, CV/strategy controllers, Search and the Lua bridge.

In the initial published checkpoint, the 5,050 and 5,510 runs underwent a new read-only comparison that checked every recorded screenshot against the existing replay. Sampling is every six emulated frames: 1,898 / 1,898 and 2,091 / 2,091 respectively. Hash CSVs are distributed; screenshot bytes are withheld. Exact recorded-input replay is distinct from MAME’s native INP playback, which is not claimed as verified by these records.

The source release is an immutable current revision; historical run manifests retain their different source-tree hashes. Image datasets, visual reference templates, model weights, emulator/game software and game-containing save states are excluded. The release therefore supports inspection of code and numerical analysis, rather than a self-contained replay of all experiments.

Reproduce legally

Obtain the required emulator and game software independently in accordance with applicable law. Arcade.it supplies no ROM, BIOS, extracted asset or game-containing save state. ROM Policy.

Publication boundary

Incoming bundles remain in quarantine. Validation checks hashes, paths, file structure and release classification. A separate provenance and license review is required before import; gameplay media requires human approval under the publication policy. Builds are staged and deployed atomically only after QA.