PHILIA TextAnalyzer CH7 Pre-registration: Establishing a Local LLM-Based External Factual Alignment Instrument via 4-Arm Evidence Manipulation

This document serves as the pre-registration for the 104th DOI of the PHILIA OS project (Phase 2, TextAnalyzer track). It registers the confirmatory design for CH7 — an external factual alignment instrument that combines Wikipedia evidence retrieval with a local LLM (qwen2.5:14b) acting solely as an entailment comparator, not a truth arbiter. The pre-registration covers: (1) four establishment conditions (EFR_explicit ≥ 0.80; A vs B separation ≥ 0.80; LLM − rule-based accuracy ≥ +0.20 with Newcombe difference CI lower bound > 0; L3 accuracy ≥ 0.70); (2) the 4-arm evidence manipulation design (A_true_ev / B_false_ev / C_no_ev / D_contra_ev) across L1×20 + L3×25 + Hard×12 = 57 corpus items; (3) pre-declared 3-class mapping for the enhanced rule-based baseline (negspaCy + WordNet + spaCy NER); (4) EFR_explicit / EFR_logical split to prevent dilution of the difficulty gradient discovered in pilot v5; and (5) seven forking paths closed before registration. This document also includes Part I, an exploration phase summary covering the three prior tracks (CEI text application → NEGATIVE; TextAnalyzer CH1/CH4 → not established; CH7 → in progress) that preceded this pre-registration. Exploratory summaries were contributed by Lucas (Grok), Pandora (ChatGPT), and Opus (Claude Opus) of the Trinity AI Research Team. Instrument scope: this DOI verifies the local LLM comparator path of CH7 only — not the CH7 concept in its entirety. Pilot history: v1 REJECT → v2 SANITY_FAIL (llama3.1:8b, EFR 43.8%) → v3–v5 SANITY_PASS (qwen2.5:14b, EFR 85–100%). All confirmatory runs will be executed locally by the Guide on C:_Phase2. Opus Gate-0: UNCONDITIONAL PASS (v1.7, 2026-08-01).Preceding DOI: 103rd results — 10.5281/zenodo.21483556