JCCDB v1.2 — Cryptographic Audit Hash and Macroeconomic Price Correction for Reproducible LLM-Based Construction Cost Diagnostics(v1.2.2)

We present JCCDB v1.2, a substantive extension to the Japan Construction Cost Database (JCCDB), an open structured dataset supporting LLM-based construction cost diagnostic systems for consumer protection. First, we introduce a cryptographic audit hash mechanism based on SHA-256 that provides reproducibility guarantees for diagnostic outputs: identical inputs always produce the same 12-character (48-bit) audit hashes printed on consumer-facing PDF reports, enabling vendor-neutral, client-side verification during contractor negotiations. Second, we describe a War Price Coefficient (WPC) subsystem that automatically incorporates manufacturer price-change announcements into monthly price corrections, addressing the documented exploitation of material price volatility in renovation contractor fraud. The current deployed coefficient is ×1.0935 (+9.35%, Hormuz Strait/Iran-conflict baseline), computed from a weighted composite of seven CGPI series and four manufacturer announcements. We demonstrate the production deployment of both mechanisms within HORIZON SHIELD, a Cloudflare Workers-based diagnostic service, and validate hash output stability. Theoretical collision analysis using the birthday-bound approximation indicates collision probability below 2×10⁻⁵ for realistic deployment scales within the 48-bit truncated hash space. The mechanisms described are designed to align with the recording and human-oversight principles of EU AI Act Article 12 and Article 14 (Regulation 2024/1689, applicable to high-risk systems from August 2, 2026), and are released under CC-BY 4.0 to support independent replication [v1.2.1 May 2026] This revision corrects 8 implementation discrepancies between manuscript and production code, adds Section 3.8 documenting OpenTimestamps Bitcoin blockchain anchoring, Section 4.5 deployment epoch theorem, and Section 7.4 practitioner observations from n=98 operational cases. Core contributions are unaffected. [v1.2.2 May 2026] This minor revision corrects a remaining numerical inconsistency between the collision-probability tables in Section 3.4 and Section 6.2: previously the two tables reported values computed from different formulas (closed-form approximation vs. exact birthday-bound), producing inconsistent figures at large n (17.8% vs. 16.28% at n=10⁷). v1.2.2 unifies both tables on the exact formula. All audit-hash values in Sections 6.3, 6.4, 6.5 have been independently re-verified. The author thanks Prof. Pang-jo Chun (The University of Tokyo) for identifying the inconsistency.