AI Policy

PatchDiff uses large language models (LLMs) to read raw patch notes from public sources and extract structured data: which subject was changed, what stat, the from/to values, and whether it's a buff, nerf, bug fix, or other. AI is used only for this text extraction — not for images, user profiling, or anything else.

One developer writes the extraction prompts, picks the models, monitors failures, and reviews bug reports. Individual change lines are not manually verified one-by-one — the volume makes that impractical for a solo project. Each analysis is annotated with the date it was processed so you can check its freshness.

• Hallucinations. The model may invent changes or attribute them to the wrong subject.
• Misclassification. Buffs may be labeled as nerfs and vice versa, especially for reworks or conditional changes.
• Omissions. Vague or non-numeric changes may not be extracted.
• No game context. The model parses numbers and keywords — it can't judge whether a change is truly a buff or nerf in the current meta.

PatchDiff does not use generative AI for any images, video, or audio. All visual assets are hot-linked from official game CDNs or are standard UI elements.

If you spot a wrong classification or missing change, email with the patch URL. We'll re-run the analysis with the current model and fix it if the output improves. Old patches may also be re-analyzed automatically when the pipeline is upgraded.contact@patchdiff.com with the patch URL. We'll re-run the analysis with the current model and fix it if the output improves. Old patches may also be re-analyzed automatically when the pipeline is upgraded.