Veritas Capsule 2 – Regime B Skeptic's Edition: Semantic Replay Under Stress (MiniLM Embeddings Dataset)

This dataset contains the full contents of Veritas Capsule 2 – Regime B Skeptic’s Edition, an escalated proof-of-concept for deterministic state alignment from a frozen schema (𝒞) alone. It uses real MiniLM sentence embeddings with verified weights checksum (SHA-256), runs both clean “bench mode” and hostile noise/drift simulations, and enforces automatic fallback to Regime A when invariance thresholds are violated. The capsule includes: Manifest & Schema Hashes binding model weights to metrics. Bench Mode Results: ψ-CQV = 0.95, Mirror Index = 0.93, Rδ = 0.03 (all thresholds passed). Hostile Mode Results: Controlled noise (σ = 0.1) and semantic drift; threshold violations trigger safe fallback. Verification Script (verify_capsule.py) for full local replication. MiniLM Embeddings Dataset (minilm_embeddings.npy), metrics JSON, and LaTeX/PDF proof report. This capsule is designed to withstand skeptical review by demonstrating: Semantic grounding via real model data. Stress test resilience with transparent fallback. Independent verifiability with no hidden dependencies.