This document serves as the foundational text for the precursor project, ARC-Epistemic, which laid the theoretical and empirical groundwork for the Epistemic Tribunal.

It is a fact of no small significance that the contemporary obsession with “artificial general intelligence” has produced machines that speak with absolute certainty while possessing almost no mechanism for doubt. They are, in the strictest sense, solipsists. They generate a single, highly probable string of tokens and defend it to the death, entirely blind to the alternative hypotheses they might have entertained.

This repository, the ARC-Epistemic project, is an attempt to introduce a necessary measure of intellectual humility—or at the very least, structured self-interrogation—into these engines.

What we have here is not merely another dataset for the machines to memorize, but a crucible. It is designed to test whether a reasoning system can distrust its own brittle internal consensus, adjudicate between competing stances, and arrive at a truth through the dialectic method rather than blind assertion.

The Poverty of the Single-Pass Solver

The prevailing methodology of our time—the greedy, single-best ranking system—is intellectually impoverished. It actively discards the coalition evidence that might save it from error. The machine guesses once, perhaps twice, and declares victory.

ARC-Epistemic is the first benchmark to make that exact loss measurable. We focus squarely on Diversity-Aware Consensus. We force the machine to generate multiple, independent hypotheses, score them against structural invariants, and aggregate their relative weights. We demand that it resolve ambiguities not by guessing louder, but by evaluating the distribution of evidence.

Empirical Vindication

We have tested this thesis, and the results, quite frankly, speak for themselves. In our rigorous evaluation over a 703-task cohort (utilizing a Qwen2.5-7B-Instruct architecture upon an H200 NVL stack), the core claim was entirely confirmed.

  • Relative Lift: The application of our diversity-aware consensus mechanism (M3) yielded a staggering 78.8% relative lift over the greedy baseline (M0).
  • The Gap Statistic: On the “selector divergence” family of tasks—where multiple rules satisfy the training examples but only one is robust—the baseline achieved a pitiful 1.9% accuracy. Under our regime, it reached 11.3%. A nearly sixfold improvement on an unseen model.
  • Recovery from Error: The machine successfully recovered from failure in 31 distinct instances, directly proving that error can be overcome when a system is forced to consider competing hypotheses.

It is a striking vindication of reasoned deliberation over probabilistic reflex.

Project Architecture & Roadmap

This project is currently scaling from its initial 303 synthetic tasks to a comprehensive 1000-task benchmark, meticulously stratified by difficulty and epistemic family.

  1. The Control Family: Tasks designed to be unambiguous, serving as the necessary baseline for structural integrity.
  2. Selector Divergence: Tasks where multiple distinct rules satisfy the training data, but produce divergent test outputs. These are the true test of the dialectic method.
  3. Hypothesis Revision & Causal Disentanglement: Tasks that force the machine to untangle correlated variables and revise early assumptions.

The final phase involves the generation of a supervision dataset—a complete ledger of reasoning traces, hypotheses, and coalition evidence—to train a new generation of models that natively understand metacognitive aggregation.

A Concluding Note

We are faced with systems that are vast, powerful, and profoundly gullible. If we are to build machines that actually reason, rather than merely approximate the appearance of reason, we must teach them how to argue with themselves. This project is the first step toward that necessary friction.