Frontier benchmark for pdf-strikethrough-detect: pre-registration

Pre-registration for a benchmark that tests whether frontier models can find struck-through text in documents as well as pdf-strikethrough-detect, an open-source package that detects struck-through text in PDFs and scanned pages, and at what cost. Ten models from Anthropic, Google, OpenAI and xAI run beside the package and two text-extraction baselines, on synthetic redlines, public redlined documents, US Congress bills and handwritten scans. These are the files tagged prereg-1, fixed before the pilot and before any model sees a test page: PREREG.md: the rules, from the arms, inputs, prompts and scoring to the statistics, the four primary endpoints and the budget. PLAN.md: the reasons behind them. corpus_manifest.json: the source documents, each with its URL and sha256. prereg1.json: what the tag fixes, including the model IDs, page lists, file and prompt hashes, and the cap. PREREG.md sha256: af555b03619ec659f369fcdae53e3ecdad31f141fddb3b66fea1606773107b68 The same files are in the package's GitHub release prereg-1. A second pre-registration, prereg-2, will be added as a new version of this record before the test run.