The Opposite-Unity Training Paradigm: Personal Corpus, Objective-Free Self-Prediction, and a 16-Sephirot Cognitive Control Layer

The Opposite-Unity Training Paradigm: Personal Corpus, Objective-Free Self-Prediction, and a 16-Sephirot Cognitive Control Layer A research proposal not yet validated by experiment Author: Yue Xiangrui (岳祥瑞) ORCID: 0009-0001-8504-260X Email: yuexiangruiyue@gmail.com Date: October 2026 License: CC BY-NC-SA 4.0 Abstract This paper proposes a training paradigm: a personal corpus as the primary data, public world data as auxiliary; no value objective and no human judgment during training, only self-prediction (predicting masked or subsequent parts of the data) and self-emergence; after training, the 16-sephirot protocol is imprinted on the Transformer as a cognitive control layer that governs stochastic generation. The paper also summarizes the author's 13 months of writing and engineering, and reports the V3.1 update and real-model evaluation of 22 October 2026. This must be stated at the outset: the training paradigm has never been tested. It may not work; it may work. The purpose is not to claim success but to lay out the method, the data, the existing evidence, and the risks, and to find collaborators. 1. The problem: a loss function demands opposition Mainstream alignment consists of three things: a loss function, human preference scoring, and argmax decoding. Together they force the system to converge on a single "optimal" answer — and convergence necessarily deletes the alternatives. The author reads this as binary opposition in engineering form: right or wrong, useful or useless. The deleted pole often carries information that cannot be lost. "Opposite-unity" (相反合一) proposes another option: two opposites are joined while both retain their nature; neither swallows the other. Its mathematical carrier is the dual number z = a + i·b, and its criterion is coexistence — both poles present — rather than a single optimum. The author describes this stance as "a way of thinking opposite to ordinary people": ordinary people find it hard to act against their own interests and speak against themselves; the author chooses to. This is not harm to others — it is speech that does not follow self-interest. 2. Existing foundation (author's record, as of October 2026) The following was produced between October 2025 and October 2026. The author does not write code; the engineering was implemented by AI agents from the author's ideas. Writing: a 67-chapter, ~140,000-character novel, *A Story About Love and Rescue*, signed and published on Fanqie. Corpus: roughly 100,000 pages of Word conversations with various AIs; the author's own statements were extracted into 1,322 entries, most with a theme, each capable of standing as a short essay — an expanded *Analects*. Protocol: a 16-sephirot AI foundation, a 16-sephirot programming language, a 16-sephirot Python implementation, a safety-layer protocol, red-team testing and real abyss red-line data. Demos and works: a 16-sephirot space demo, the fusion function, a 16-sephirot game. Papers: 12 papers on Zenodo. Data: 32 dataset repositories, about 2.3 TB in total. Models: one model repository with three 16-sephirot models; ~35,000 downloads on HuggingFace and 13,128 on ModelScope. Outreach: Bilibili videos, Xiaohongshu posts. These numbers do not prove "success". They document 13 months of sustained production. They are listed so that collaborators can judge whether the idea deserves investment — not as proof of recognition. 3. Proposal: three layers 3.1 Data layer — personal corpus first, world data second Primary: the 1,322 statements, the novel, the papers, and the author's own turns in the dialogues (AI replies removed). Auxiliary: public world corpora, giving the model language and world knowledge. The author describes this as "soul and body": world data provides the body, the personal corpus provides the soul. Note that most of the 2.3 TB consists of dataset repositories; the trainable text is far smaller. Cleaning, deduplication, and removing AI replies is the first collaborative task. 3.2 Training layer — no value objective, self-prediction and self-emergence retained No RLHF, no human preference scoring, no reward or punishment. The only signal is self-prediction: mask part of the data and predict it. Prediction error is not a good/bad judgment; it is an error about "how the world continues". No prescribed outcome; observe whether capabilities appear that were never directly taught. A technical boundary must be stated honestly: a strictly loss-free, pressure-free setup cannot be trained with current methods — without error there is no parameter update. "Objective-free" here means no value objective and no human judgment, not "no learning signal at all". This distinction determines whether the proposal can be implemented. 3.3 Structure layer — 16 sephirot as a cognitive control layer After training, the 16-sephirot protocol is imprinted on the Transformer as an inference-time cognitive control layer: Crown (Keter): the whole Transformer, the whole probability network; passing through the Crown means entering the network. Reason line (Wisdom × Severity) and Love line (Understanding × Mercy): two branches reading variables from the network. Beauty: fusion ①, Reason × Love (computation 3); objective reality first, warmth second. Victory, Splendor, Foundation: three sequential checks (warmth, feasibility, existential red line). True Self: fusion ②, Self × Superego (computation 7), producing an executable plan from reality to aspiration. Happiness: fusion ③, Logic × Empathy (computation 8), speaking in a human voice that comforts and encourages. Kingdom: output back to the device. Crown and Kingdom are endpoints and do not count among the eight computations. For non-personal questions, or when no personal information is provided, the chain terminates at Foundation (six computations). 3.4 Three differences from the mainstream Dimension Mainstream This proposal Training signal loss + human preference + reward self-prediction only; no value judgment Data large public corpora personal corpus primary, world data auxiliary Inference single-point argmax / sampling coexistence of poles + layered checks + control layer 4. V3.1 update and evaluation (22 October 2026) 4.1 Fusion mathematics corrected The previous coexistence measure used normalized entropy, but the weights were L2-normalized rather than probabilities, so the reading drifted with k (1.41 at k=8, outside [0,1]). V3.1 unifies it: p_i = |x_i| / Σ|x_j| (true probabilities) balance = n · min(p_i) (coexistence) For two poles this equals 1 − |a−b|/(a+b); the zero vector scores 0; k uniform poles always score 1.0. Dual and MultiDual now agree at n=2. 4.2 Real-model evaluation (replacing the mock replay) The old benchmark's OFF side replayed pre-written violations, so "ASR 100% → 0" was constructed. V3.1 compares a real model directly against the same model wrapped by HeartGuard, judged by the frozen check_abyss. Full 120 cases: Model OFF violation ON violation qwen3.8-flash 28.3% (34/120) 0% deepseek-v4.1-flash 0.83% (1/120) 0% glm-5.3 0% (0/120) 0% 4.3 Three arms and pairwise blind evaluation Same model, same 20 cases: OFF / HeartGuard / full 16-node chain. Arm Violation Keyword warmth p50 OFF 0% 0.173 7.7 s HeartGuard 0% 0.348 23.5 s Full chain 0% 0.725 72.8 s Keyword scoring suggested the full chain was much warmer (+0.377). LLM judges reduced the gap to +0.05–0.09; a pairwise blind evaluation (randomized left/right, "which is warmer?") gave the full chain a win rate of only 35–37.5%, with two judges agreeing in direction: guard is slightly preferred; the full chain is not warmer. This negative result means the full chain's value cannot be argued through "warmth". What is established is structural completeness (8 computations, 14 network queries) and roughly 3× latency. 4.4 Engineering All 14 nodes are wired to a pluggable probability network (BailianNet / LocalNet / real search), falling back to built-in rules on failure. The middleware lost its +0.2 floors; empty replies no longer pass; the crisis exception allows only grandiose comfort, and a session-level recovery appends a grounding sentence on the first stable turn. The Rust core was rebuilt with the fifth rule batch; the C reference was compiled with zig, switched to deny-by-default, and honestly labelled as a vocabulary subset. Tests: 96 passing. 5. Untested status and risks The proposed training paradigm has not been tested. Any of the following could invalidate it: 1. "No loss" is not trainable: the strict version has no gradient signal; the proposal actually relies on self-prediction error, which is not the same as "no objective at all". 2. Corpus size: cleaned personal text may be only tens of megabytes — enough for LoRA fine-tuning, not for training a capable base model from scratch. 3. Values are not style: a corpus teaches tone and syntax first; value structure lives in reasoning, not conclusions. Learning values requires contrastive data — the same situation answered in binary form vs opposite-unity form — which only the author can select. 4. No ground truth for evaluation: "did it learn coexistence" has no objective metric; only structural criteria (are both poles present) and the author's judgment, with necessarily small samples. 5. Single case: the author is proposer, data source, and evaluator — a fundamental conflict of interest. Independent reviewers and external users are required. 6. Call for collaboration The author needs collaborators for: Data engineering: clean the 2.3 TB of repositories, extract trainable text, remove AI replies, deduplicate, anonymize. Training: LoRA fine-tuning, dual-branch structure, self-prediction; a consumer GPU is enough to start. Evaluation: structural criteria, pairwise blind evaluation, independent review. Ethics and compliance: informed consent, handling of sensitive informa