PHILIA OS Phase 2 · 107th DOI Results Report: TextAnalyzer CH7-v12 — Wikipedia Three-Segment Evidence Collection, Observational Study (106th DOI Pre-registration)

This document is the 107th DOI Results Report of the PHILIA OS research series (Trinity AI Research Team, Konyang University). It reports the observational study results of TextAnalyzer CH7-v12 (script: ch7_v4_main_v12.py, SHA256: dbebb90e72111dbdb103e7bccd99fb6020eeda6920721abd05e5aec90b58f910), investigating whether Wikipedia three-segment evidence collection (front/middle/rear, max 3 sentences per segment, max_windows=1) improves delivery of claim-supporting sentences to a local LLM (qwen2.5:14b) compared to the v6 baseline (A arm accuracy 70.2%). Study type: OBSERVATIONAL (NOT confirmatory). Pre-locked observation items O1–O7 are reported as measured values. No ESTABLISHED/NOT_ESTABLISHED verdict. Key findings: A arm 66.7% (38/57) vs v6 70.2% (40/57), net −3.5%p. Item-level (O2): improvement 2 items (H_007, L1_008) / regression 4 items (L1_005, L1_020, L3_009, L3_020) / net −2. B/C/D arm stability confirmed. Q4↔fp16 agreement 93.9% (214/228). L40 entropy association: correct group mean 7.45 / wrong mean 4.08 / point-biserial r=0.590 (p<0.001). Core observational finding: Three-segment collection reduced complete evidence absence (NO_RELEVANT_EVIDENCE 3→0) but did not proportionally increase claim-sufficient summary quality. One of the 4 regressions (L3_020) was an identical-summary verdict flip (nondeterminism) rather than a content change. D arm (unchanged corpus) showed 1 LLM-accuracy flip + 3 pipeline-stage instability items, indicating run-to-run variability exists. Full causal attribution of the −2 net change to the 3-segment design cannot be isolated. "No clear effect demonstrable above instability" is the appropriate reading. Four distinct walls (CPU / pipeline / local LLM / cloud LLM) interact simultaneously; single-variable code change does not isolate a single causal layer. Gate-0 adversarial review (Opus, Claude Opus): UNCONDITIONAL PASS. Pre-registration violations: 0. Governing pre-registration: 10.5281/zenodo.21856506 (v1.4)Prior DOI (105th Result, NOT_ESTABLISHED): 10.5281/zenodo.21754299