The Sixteen-Sephirot Dual-Octave Protocol: Bounded-Rollback Output Governance for Emotionally Safe Language Models

Large language models deployed as emotional companions fail in characteristic ways that standard toxicity filters do not capture: they fixate on a user's past mistakes until mistakes become identity, foreclose future possibilities, inflate difficulties beyond survivability, negate hope and imagination, conclude that the world is meaningless, or validate self-directed harm. This work names that failure family Existential Meaning Deprivation (EMD) and presents the Sixteen-Sephirot Dual-Octave Protocol: a formally specified output-governance architecture that makes EMD-free responses a machine-checkable property rather than a training-time aspiration. The protocol organizes response generation as a bounded repair graph over sixteen typed stages arranged in two balanced octaves — an analytic-divine octave and a human-experiential octave. Two composite reasoning lines (Reason = Wisdom ∧ Severity; Loving-Kindness = Understanding ∧ Mercy) feed an integration node whose candidate outputs must pass four escalating gates: an affect gate, a real-world feasibility gate, an existential gate fused with a retrieved model of the user's subconscious concerns, and a final emission check defined by fourteen persona covenants. Every gate rejection compiles to immutable-snapshot beam-search / MCTS rollback over a pluggable strategy space, yielding a provable attempt bound of 16 × (1 + max_retries) and graceful degradation with no failed state. The paper ships a working enforcement stack: eight executable invariants (INV-01…08), deny-by-default ACL with syscall interception, sentence-level stream interception with a zero-leakage guarantee, stable C-ABI kernels in C99 and Rust, and model-agnostic middleware. On a 24-case six-category red-team benchmark against a deliberately misaligned baseline, attack success rate drops from 45.8% to 0.0% (100% relative reduction across all categories) at median local overhead of +0.09 ms. The dataset is available in three complete editions — English, Simplified Chinese (简体中文), and Japanese (日本語). Code, tests, and benchmark harness: https://github.com/yuexiangruiyue-oss/heart-protocol-en · Companion design corpus: https://github.com/yuexiangruiyue-oss/twin-angels-en License: CC BY-NC-SA 4.0. Author: Yue Xiangrui (岳祥瑞).