Trinary-19683: An Open-Source Falsifiable Deterministic Cognitive Layer

While large models are still "guessing" the next word probabilistically, an open-source project called Trinary-19683 is trying to give AI a "deterministic cognitive anchor." It uses 3^9 = 19,683 ternary states to build a symbolic reasoning layer that is recomputable, falsifiable, and evolvable, targeting the hallucination problem of large models head-on.

From Static Hexagrams to a Dynamic Cognitive System

The author previously proposed the idea of a "ternary nine-dimensional 19683 cognitive system": based on a 3^9 state space, overlaid with five-domain semantic bridging. The system has now been fully implemented and open-sourced, with deliverables including an offline model, an inference workbench, a service console (15+ APIs), and an Apache-2.0-licensed Python package.

What it aims to be is not a divination tool, but a deterministic symbolic cognitive layer for LLMs and agents. The author explicitly distinguishes the two: traditional I Ching tools are "cast a hexagram + look up text," whereas in Trinary-19683 every hexagram, every jump, and every transition carries a deterministic signature and an auditable trace.

Ternary and three-valued logic: Ternary uses the three digits 0, 1, and 2, with each digit position capable of carrying three states. Compared with binary's two-state encoding, it is naturally suited to expressing three-valued logic such as "yes/no/undetermined" or "thesis/antithesis/synthesis." In symbolic AI, three-valued logic (such as Kleene three-valued logic) is often used to handle situations with incomplete or contradictory information: beyond true and false, it preserves "unknown," avoiding the need for binary systems to make assertions when evidence is lacking. Trinary-19683 uses 3^9 to construct its state space, meaning 9 dimensions each take one of three values, yielding exactly 19,683 discrete states, so that each state can be uniquely indexed by an integer ID, thereby supporting deterministic recomputation and auditing.

The Technical Skeleton of the Deterministic Reasoning Layer

The core of the system is a fully recomputable reasoning mechanism. The complete 19,683-hexagram state space uses D/R/T ternary encoding, and hexagram IDs are aligned with line states through a three-generation correction engine, achieving zero missing entries.

On top of the state space, the system defines five-dimensional structural signatures and 301 rules for computing a deterministic adaptive rank. The author claims the process is 100% recomputable and achieves 100% through leave-one-out validation. In addition, the system supports dynamic potential-flow reasoning, three-state axiomatic semantics, and temporal evolution paths (suggest / trace interfaces).

This means the reasoning process is not black-box sampling, but symbolic operations in which every step can be independently reproduced and verified.

Leave-one-out validation: Leave-one-out is an extreme form of cross-validation in which one sample is held out from the dataset as the test set each time, and all the rest are used for training or fitting, repeating until every sample has been tested once. It is often used in small-sample scenarios and can make fuller use of the data, but computational cost grows linearly with the number of samples, and evaluation results are relatively sensitive to individual samples. In symbolic rule systems, leave-one-out can be used to test whether a rule set can still correctly derive any given state after that state is removed, but "100% pass" only shows that the system is self-consistent within its own state space, and does not directly equate to generalization ability over external semantics.

Cognition Board v2: Writing "Falsification" into the Architecture

What truly distinguishes this project from ordinary symbolic systems is the falsification dynamics introduced by Cognition Board v2.

The system supports graded rebuttal, reversible transitions, and self-contradiction detection. In the author's words, it teaches the model "when to abandon a belief, and where to go instead." This amounts to handling hallucination at the architectural level rather than the prompt level: when the model produces free associations, the system can absorb them, verify them, and solidify reliable associations into deterministic rules.

The metacognitive board L4 further provides M(X) meta-evaluation and M(M(X)) recursive self-reflection, and the entire process is falsifiable. The multi-agent protocol L5 localizes conflicts to specific transition edges and supports rebuttal collisions and recomputable joint chains.

Falsifiability and falsification dynamics: Falsifiability was proposed by Popper and holds that a proposition must be possible to overturn by observation or experiment in order to be scientifically meaningful. Introducing falsification into AI cognitive architecture means the system must not only generate beliefs but also maintain, for each belief, the conditions under which it can be rebutted and the transition paths involved. Graded rebuttal allows the system to lower belief confidence level by level according to the strength of evidence, reversible transitions ensure that error correction does not destroy the existing state space, and self-contradiction detection is used to find conflicts within a reasoning chain. This mechanism moves hallucination governance from "post-hoc prompt correction" forward to "belief lifecycle management," but whether the design of the falsification rules themselves is complete remains the key to its effectiveness.

Multi-Agent and the Open-Source Ecosystem

The L5 multi-agent protocol deserves separate attention. When multiple agents collaborate, disagreements are often hard to localize. This system pinpoints conflicts to a specific transition edge, making joint reasoning chains recomputable and auditable. This offers a new approach to building explainable multi-agent systems.

The project is open-sourced on GitHub (q1z2q3-debug/trinary-19683) under the Apache-2.0 license, and the author welcomes stars, forks, and issues.

A Rational View: Innovations and Points Still to Be Verified

A note here: the capabilities above all come from the author's self-description, and independent third-party verification is still insufficient. Metrics such as "100% leave-one-out validation" lack publicly available evaluation details, and actual generalization ability remains to be tested by the community.

Using the I Ching hexagram system as the encoding framework for symbolic reasoning also raises questions about whether its semantic bridging is rigorous enough and whether it can truly constrain the free associations of large models; more empirical evidence is needed. Still, the direction of using a falsifiable, recomputable deterministic layer to hedge against the hallucinations of probabilistic models is itself worth exploring.

For developers interested in AI cognitive architecture, symbolic reasoning, and hallucination governance, this project offers an open-source sample worth studying.

One-line Summary

Trinary-19683 open-sources a ternary symbolic cognitive layer that uses falsifiable reasoning to hedge against large-model hallucinations.

Paragraph Summary

Trinary-19683 is an open-source symbolic cognitive layer built on a 3^9=19,683 ternary state space, intended to provide LLMs and agents with a recomputable, falsifiable, and auditable deterministic reasoning anchor. The system includes an offline model, an inference workbench, a service console, and an Apache-2.0 Python package, with core mechanisms involving D/R/T ternary encoding, five-dimensional structural signatures, 301 rules, and deterministic adaptive rank. Cognition Board v2 introduces graded rebuttal, reversible transitions, and self-contradiction detection in an attempt to govern hallucination at the architectural level; the metacognitive board L4 supports recursive self-reflection, and the multi-agent protocol L5 can localize conflicts to specific transition edges. The author claims 100% leave-one-out validation, but independent third-party verification is still insufficient, and the rigor of the semantic bridging remains to be empirically tested.