PHILIA TextAnalyzer CH7-v20: Exploratory Pre-registration — Intra-Document Factual Locus Selection via SVO Claim-Slot Zeroing

This is the pre-registration document for PHILIA TextAnalyzer CH7-v20, an exploratory study conducted by the Trinity AI Research Team at Konyang University. CH7 is an evidence retrieval pipeline that fetches factual information from Wikipedia to anchor LLM responses, with the goal of retrieving correct facts even when the input claim is wrong. CH7-v19 (hotfix2 baseline) achieved entity resolution in 20/21 cases (95.2%) but retrieved the correct factual sentence in only 9/21 cases (42.9%), with 7 of 12 failures attributed to intra-document locus selection failure. v20 tests a single intervention: zeroing claim-derived SVO object/time/location slots before sentence scoring, while retaining the existing entity-centric hard ordering. This isolates whether claim-predicate leakage through these slots is responsible for the intra-document retrieval failures. This document was produced under the PHILIA Open Science (PHILIA OS) protocol, including multi-member adversarial Gate-0 review (Opus, Pandora, Lucas) across 8 revision cycles (v0.1–v0.8), SHA256 script registration prior to execution, and pre-locked observation items (O1–O6) to prevent post-hoc forking. Script SHA256: 69bcd6241f91468e2bd62232d0464f79f9c461a632e2e2e7a503fcf849254f1d