Binary Bitmask Verification vs LLM Self-Review at Service Scale: The Robustness Verification Framework

At service scale ≥ 200 components, interpretation-based verification fails. LLM self-review depends on the writer-agent reviewing its own output — structurally insufficient when the agent cannot perceive the defect it introduced. We present the Robustness Verification Framework (RVF), a four-layer binary-first verification stack that subordinates LLM judgment to deterministic bitmask predicates. Empirical: deployed across 261 services in the ANKR substrate over four sessions in May 2026; surfaced 660 silent bugs across 175 source files (the "botched latency bulk-edit" pattern), 11 codex overclaims, 22 R-008 port-authority violations, 2 invalid capability_captured claims, and one fleet-wide gitleaks-quirk; 19 services formally proof-eligible per the 6-bit Layer C threshold; 5/5 negative-gate scenarios pass; 0 false-positive production blocks. We argue capability hallucination at scale is structural to LLM-author/LLM-reviewer systems and that binary verification is the only enforceable response. Companion to ArchMask (10.5281/zenodo.19417112), BitMaskOS (10.5281/zenodo.20047909), PRAMANA Protocol (10.5281/zenodo.19273330), and Vivechana SAR-006 (10.5281/zenodo.19456053).