PHILIA Knowledge Graph Probe — Full-Corpus Wikipedia Evidence Harvest: 57-Item Exploratory Pre-registration (116th DOI)

Abstract

This registered exploratory pre-registration documents the design and declared observation items for a full-corpus (57-item) Wikipedia evidence harvesting run using the PHILIA KG Probe pipeline. The pipeline was validated on a 5-item pilot (2026-09-17–19) and generates structured claim–sentence–source-URL rows from Wikipedia for each claim in the standard corpus (philia_ch7_corpus_main_v3_patched.json, SHA 5fd61f8f). Source dualization is confirmed: body text, tables, and external links from rendered HTML; INFOBOX from wikitext. The study applies a single base-claim arm with no verdict threshold; observations cover per-item row distribution, fetch success rate (with None-guard on fetch_status), U.S. abbreviation boundary behavior, provenance stamp integrity, and a known fmt_date parser defect (O5). Version 2 supersedes the execution code declared in v1. The v1 execution (v3.6, SHA 2315db70) completed all 57 items but terminated during the post-hoc retry stage before the results JSON was written — an unhandled IndexError in the wikitext INFOBOX parser, isolated on the main path but not on the retry path. Per the pre-declared run definition (P4-6, "the registered run is the first execution that writes a results JSON"), that run did not constitute the registered run and re-execution was permitted. The execution code is therefore v3.7 (SHA c99d9d6a, parent 2315db70), a two-point change adding retry-path exception isolation with no change to values. One procedural deviation is documented: the v3.7 execution preceded this revision. Analysis violations: 0. Both the crash-run console log and the completed results JSON are included. Gate-0 adversarial review: 11 rounds (Opus ×10, Pandora ×1), 27 issues (CRITICAL 4 / MAJOR 16 / MINOR 7), all resolved. Part of the PHILIA-EcoSensory Swarm project, SC Phase 3. Prior lineage: 115th DOI (10.5281/zenodo.22669957).