Educational-slide Retrieval Reproducibility Package v7.0.0

This versioned package accompanies the manuscript “When Does OCR Help Educational-slide Retrieval? Multi-positive Relevance, Lecture-held-out Splits, and Encoder Dependence.” It contains the author-written analysis code, frozen protocols, nonpixel entity and relation manifests, lecture-held-out split assignments, result tables, cryptographic hashes, and independent audit reports used for the v7 analyses. The main LPM surface comprises 1,064 figure entities, 644 caption entities, and 42 lecture groups; 222 captions have more than one relevant figure. A separate direction check covers 1,107 eligible M3AV talks and 24,223 admitted relations. The package does not contain lecture videos or audio, slide images, reconstructed frames, figure crops, contact sheets, model weights, or frozen feature matrices. These materials must be obtained from their official providers under the applicable terms. The MIT License applies only to author-owned code and original documentation. LPM- and M3AV-derived nonpixel materials remain subject to their upstream terms, as detailed in THIRD_PARTY_NOTICES.md and THIRD_PARTY_RIGHTS_AND_REDISTRIBUTION.md. The uploaded archive is byte-identical to the GitHub v7.0.0 release asset: 7,905,023 bytes; SHA-256 3a088b5fd6c137e79c0b58c6a84cef5ee67a89912818015b0aad003bcbe10ed0.