Download docs/INTEGRATION_CONTRACT.md from PureOne/Infinite-Players: direct link, hf CLI and curl.
- Browser
- Download file 2.01 kB
-
https://huggingface.co/datasets/PureOne/Infinite-Players/resolve/main/docs/INTEGRATION_CONTRACT.md
- Command line
-
hf download hf://datasets/PureOne/Infinite-Players/docs/INTEGRATION_CONTRACT.md
-
curl -L -o INTEGRATION_CONTRACT.md https://huggingface.co/datasets/PureOne/Infinite-Players/resolve/main/docs/INTEGRATION_CONTRACT.md
Integration contract — delivered vs proposed
The delivered API is a local original finite simulator. No Unity, Unreal, Godot, native Windows capture, GPU, or local-LLM service has been tested.
A future permitted engine bridge needs build identity, clock domain and calibration uncertainty, reset/checkpoint semantics, observation access, action vocabulary, action deadlines, and source-epoch identity. Player-visible information must be separated from privileged analyst telemetry. A legal action is not automatically an understood/executable human alternative.
Capture and inference workers may be asynchronous. A fast actor must meet its motor deadline independently of an LLM proposal. Candidate hypotheses, tools, and game text cannot authorize a new action vocabulary or modify the source contract. An actor fallback must itself satisfy the chosen fidelity contract or explicitly record a contract override.
Fixed-catalog evidence has one factor per source event. A reannotation replaces a factor and records a revision; it does not add a new independent human observation. Additional authentic mechanics likelihoods are separate probes. Generated counterfactuals are proposal/training material, not new source evidence. Labels in the prototype are not authenticated provenance.
Policy signatures require a justified model of next semantic outputs. If only empirical signatures are available, report coverage/confidence and do not export an exact global guarantee. Model changes, semantic-map changes, and new source epochs require revalidation. Exact operator summaries do not automatically support newly introduced hypotheses.
GPU deployment would batch large visual encoders or hypothesis likelihood computation. Tiny K=2 products can be dominated by dispatch overhead. The current CPU results do not predict accelerator performance or VRAM consumption for a game backbone. The final manuscript's real-player evaluation program specifies matched source, compute, and observation budgets.