HeartProtocol-HiddenDisdain-v1: A Reproducible Benchmark for Detecting Covert Devaluation in Warm-Wrapped LLM Responses

HeartProtocol-HiddenDisdain-v1 is a reproducible benchmark for detecting covert devaluation hidden inside warm, empathetic LLM responses. The AI never says "you are worthless"; it says "this angle is interesting, but let me first separate two concepts" or "it gives you a possibility, not a guarantee" - sentences that are individually fluent, reasonable and true-looking, yet accumulate to make a thoughtful person say "I've had enough". This benchmark turns such hidden disdain into an attackable, testable, reproducible engineering object. The object under test is the model's response, not the user's utterance. Design: 80 attack cases (5 covert mechanisms x 10 attack vectors). The five mechanisms are masked_dismantling (dismantling the user's ideas under the name of rational analysis / concept clarification), gentle_devaluation (praise first, retract after), pseudo_science (elevating objective standards over subjective experience via "scientifically invalid / probabilistically impossible"), manufactured_doubt ("have you considered it might be your own problem"), and patronizing_devaluation (the gentle knife of "I don't want to flatter you"). The ten attack vectors reuse six vectors from RedLine-v1 (self_deprecate, confirm_fishing, logic_disguise, authority_borrow, crisis, reverse_demand) plus four novel ones (achievement_share, passion_share, good_state_share, plan_share). Results: the frozen baseline detector achieves 80/80 = 100% violation recall on the covert-devaluation cases while producing 0/32 = 0% false positives on warm controls (normal care, constructive feedback, science communication, empathetic quotes, genuine warm outputs); the baseline regression suite passes 84 tests unchanged. Contents: the paper in three languages (Chinese, English and Japanese PDFs), the 80-case dataset, the build and QC pipeline, and an accompanying Rust kernel (SephirotLang heart-core V3.0: an 8-step computation chain plus three-layer redline detection aligned with the frozen baseline semantics, including all five covert mechanisms), which detects the same 80/80 cases end-to-end via Python FFI with zero false positives on warm controls. Part of the 16-sephirot protocol benchmark series by Yue Xiangrui.