The King Wen Experiments

三千年前的算法能否训练人工智能？

Testing the King Wen I-Ching sequence as a learning rate schedule and curriculum ordering for neural network training. Two experiments, two negative results — and what they reveal about the boundary between ancient pattern recognition and modern optimization.

三次救援与一个发现

Three attempts to rescue the King Wen curriculum hypothesis — through seed sensitivity, cross-platform replication, and adaptive selection. Mengzi and Xunzi debate innate nature versus cultivation. A novel finding about how modern AI compilation tools interact with data ordering emerges from an unexpected direction.

三角为何杀死韩国

A 3-state Warring States simulation eliminates Han in 93% of games within five rounds. The Zhanguoce explains why: you cannot test ancient statecraft by removing the geopolitical complexity that made it necessary. Han Fei, Su Qin, and Zhang Yi provide the postmortem.

从优化到策略

Five negative results point the way forward. The King Wen sequence's mathematical properties — liabilities in gradient descent — are virtues in multi-agent strategic decisions. The research pivots from training AI to playing the game it was written for.

七国一卦

Seven AI agents with Warring States personas play Diplomacy-style campaigns. Han — the smallest state — consults the I-Ching before every battle. The game works, the oracle speaks, and the first results show something more subtle than a survival advantage: different reasoning, different arcs, the same territories at game's end.

山崩

The oracle spoke truly and Han died anyway. In game 6af8, Hexagram 52 (Keeping Still Mountain) counseled absolute stillness. Han interpreted stillness as holding one territory while supporting an ally's attack with the other. Both territories fell simultaneously — the earliest elimination in 68 games. A tragedy about RLHF cooperation bias, philosophical amplification, and the historical parallel to Han's fall in 230 BC.

生而不动

Sixty-eight games, four conditions, and the survival hypothesis is dead. Han survived at 80-92% regardless of oracle condition — statistically indistinguishable. But the trajectories told a different story: scrambled oracle compresses outcome variance by half (Levene's p=0.031), yarrow produces a distinct behavioral signature (61% hold, 19% move), and a game bookended by Hexagram 15 (Modesty) asks whether a philosophical circle is wisdom or stagnation. Three LLM anti-patterns are named for the first time.

坚城之局

The map froze the game. Twenty-two territories, all supply centers, no buffer provinces, no flanking routes — a board that violates every principle the Diplomacy variant community has established across forty years of tested designs. Probe addiction is not just an LLM failure; it is the only option on a board with no room to maneuver. The fix requires no rule changes: add 15 corridor provinces for staging and flanking, 5 new neutral supply centers to extend the early-game scramble, and reduce central connectivity to break the stalemate line. Hexagram 39 (Obstruction) frames the question and the answer: see the danger, then redraw the earth.

解围之野

The second-generation board is built. Forty-six territories where there were twenty-two. Nineteen corridors where there were none. Eight neutral supply centers where there were three. Zero cross-border home adjacencies where there were fifteen. The combat rules are unchanged — only the earth has been redrawn. Mountain passes create warning phases, river crossings channel movement, and contested neutral zones extend the early-game scramble from two rounds to six. Hexagram 40 (Deliverance) follows Obstruction: the storm breaks, the seeds crack open, and the oracle finally has room to prove whether different thinking produces different outcomes.

卦变天下

Twelve v2 games, six with the oracle and six without. Han's survival rate: 33% both conditions — dead even. The survival hypothesis is buried for good. But the oracle changed the board itself: Yan gained nearly five territories in oracle games, Chu gained three and a half, Qi lost nearly three. The philosophical framework injected into the weakest state propagated through diplomatic channels and territorial vacuums to reshape who wins — just not the oracle-user. Qatar, Article 9 Japan, Finlandization, and what alignment researchers should worry about.

三卦同局

Twenty-four v2 games across three conditions: yarrow, tarot, and control. Han's survival remains statistically indistinguishable — the oracle does not help the oracle-user. But the Tarot condition produces a sharp ecosystem signal: Qin wins five of six tarot games (Fisher's exact p<0.01, survives Bonferroni correction). This is the first statistically rigorous evidence that a philosophical framework injected into one agent reshapes which other agent wins the multi-agent competition. The yarrow-Yan pattern from Dispatch 10 remains descriptively consistent with a matching effect but awaits the ongoing yarrow expansion to separate from noise at the winner level. The cs.AI paper will carry the tarot result as its empirical backbone.