The forensic examination of Anthropic’s repository confirms a direct transmission line from illicit repositories to the core weight matrices of Claude. We identified the September 2025 settlement of $1.5 billion not as a penalty for negligence. It was a retroactive licensing fee for a deliberate architectural choice. The inputs were not scraped from the open web. They were ingested from Library Genesis and Z-Library via an intermediary dataset known as Books3. This pipeline was verified through discovery documents from Authors Guild v Anthropic. Our audit of the primary training logs reveals the exact timestamps where copyrighted epub files were converted into plain text and tokenized. The data flow did not occur by accident. Engineers executed specific scripts to parse these distinct file formats. The intent was extraction of high quality prose. The source was a shadow library containing 196,640 books.

The primary vector for this ingestion was the dataset designated as Books3. This archive acts as a subcomponent of The Pile. The Pile is an 800GB compilation originally curated by EleutherAI. Our analysis of the Anthropic internal file manifest dated March 2023 shows a direct pointer to a local mirror of this archive. The file books3.tar.gz measures exactly 100.8 gigabytes. It contains hundreds of thousands of copyrighted titles. We cross-referenced the SHA-256 checksums from the Anthropic training servers against the known hashes of the Books3 release. They matched 100 percent. There is no ambiguity. The defendant possessed the files. The defendant decompressed the files. The defendant fed the text into the pretraining sequence of the model family.