Gameplay Data Kit (stub-v1): a verifiable tape format for world-model and agent teams
A downloadable tool
STUB NOTICE FIRST: every sample in this kit is a synthetic placeholder made in our own build with a seeded simulation. It is not engine output and it is not real gameplay. The kit exists to show the format and to let you check the method before real tapes exist. The public real-tape sample ships only after the engine2040 tape exporter (HS15) lands and the verifier passes on real tapes. Nothing on this page claims more than the replay hash proves.
Who it is for: a team that trains or evaluates world models or game agents and wants to see what an action-conditioned, rights-clean tape should carry before it buys any, or a data seller who wants a reference shape for their own tapes.
What is in the zip (176 files, counts read from the zip): a manifest schema with 21 required fields and a state dictionary of 6 fields (name, unit, frames, rate, version each), a datasheet, a dataset card, a stdlib verifier (VERIFY.py), a draft stub-sample licence and a paid-pack licence template, a provenance file, a real test log, and 10 stub samples in 3 scenarios (arena-duel x4, siege-gate x3, urban-entry x3). Each sample: a manifest with its seed and replay hash, 12 PNG frames (320 x 180, STUB burned in), 60 state ticks at 20 Hz, raw and normalized controls (60 rows each). Totals: 120 PNG frames, 600 state ticks, 1200 control rows, 10 of 10 samples flagged stub true, 0 real engine tapes.
Check it in 5 minutes, Python 3 only, nothing to install: unzip, open a terminal in the folder and run python VERIFY.py --manifest PACK-MANIFEST.json --data STUB-SAMPLE. Expected last line: ALL 10/10 VERIFIED. The verifier re-hashes the state ticks and both control files of every sample (the replay hash) and checks every frame PNG against its own sha256 and its size. Flip one byte in a frame and it names the frame file and the sample; flip one in the state or controls and it names the sample. TEST-LOG.md records dated runs, including the same seed run twice with the same hash.
Free sample: sample/gameplay-data-kit-free-sample.zip (20 files, one stub sample, the verifier and a one-page how-to-verify-a-tape sheet) verifies as ALL 1/1 VERIFIED.
What you do next: verify, read DATASHEET.md and DATASET-CARD.md, test your own loader against MANIFEST-SCHEMA.json. If you want real tapes, reply to this page and name the unit you would pay for (an hour, a scene or a licence); no price exists until a buyer names one.
Honest limits: placeholder data only; no engine code is included (engine2040 is named by revision, read-only); the stub licence is a draft that has not been signed off by a lawyer; no paid product exists yet.
Licence: royalty-free for evaluation and research, including commercial evaluation; do not resell or redistribute the raw files (LICENSE.txt, LICENSE/STUB-SAMPLE-LICENSE.txt, draft). Sources: none, every file is generated in our own build; provenance in THIRD_PARTY.md.
AI disclosure: assisted. The listing and the docs were drafted with AI assistance; the stub frames and state are generated by seeded scripts (Pillow); no generative model made any sample, and every sample is re-hashed by VERIFY.py.
| Published | 2 days ago |
| Status | Released |
| Category | Tool |
| Author | sabako |
| Tags | dataset, game-agents, reinforcement-learning, replay-verification, synthetic-data, tool, world-models |
| AI Disclosure | AI Assisted, Text |
Download
Click download now to get access to the following files:




Leave a comment
Log in with itch.io to leave a comment.