Dialectical Debate AI

This project implements a single loop: debate → belief update → memory encoding → collective aggregation → repeat — an evolutionary process for ideas grounded in Bayesian epistemology, where debate is the selection mechanism and implicit memory is the inheritance mechanism.

The Loop

It starts with debate preparation. A BeliefRefiner decomposes a topic into a tree of beliefs — values, claims, sub-claims, evidence — structured as a Bayesian network with credence values. Each side (AFF/NEG) gets a perspective: a pruned view of this tree containing the beliefs, arguments, and evidence assigned to them. This isn't just prompting — the belief tree is a formal epistemological structure where credences satisfy probability axioms and updates follow conditionalization.

When a debater is assigned a side, they don't receive their beliefs as a system prompt. They receive them as HEXIS M-states — implicit memory perturbations that reshape what they attend to, what evidence feels salient, what rhetorical strategies feel natural. The debater doesn't "know" their beliefs in the way a RAG system retrieves facts. Their beliefs are dispositional — they shape perception before deliberation, like Aristotle's hexeis.

The debate runs through a multi-stage pipeline: tactic selection, argument skeleton, evidence selection, speech generation — with flow tracking across seven IPDA speeches and cross-examination. A multi-tier judge ensemble scores each speech on argument quality, evidence usage, clash engagement, and format-specific dimensions.

After the debate, judges update their beliefs. Debate outcomes become evidence: winning arguments strengthen the beliefs they relied on, losing arguments weaken theirs. Updates follow Bayesian conditionalization with damped likelihood ratios, and propagate through the belief tree via Jeffrey conditionalization. Coherence checking ensures the resulting credence distribution remains probabilistically valid.

After many debates across many topics, each agent — debaters and judges alike — accumulates nuanced, experience-specific M-states. A judge who has watched 50 debates on economic policy develops different dispositions toward economic arguments than one who specialized in ethics cases. These dispositions are implicit: the judge can't articulate them, but they shape what the judge notices and how they weigh evidence.

The final step is federation. Individual M-states — each encoding one agent's accumulated experience — get aggregated into a higher-order representation. This is the open research question: how do you compose implicit memories? The aggregated M-state gets recompressed and distributed back to individuals, augmenting their future debates with the collective's refined understanding.

This is evolution for ideas. Debate is the contest — the selection pressure. Bayesian belief updating is the fitness function. HEXIS M-states are the genetic material — inherited dispositions that shape behavior without being explicitly articulated. Federation is reproduction — combining successful traits from the population. And the loop repeats, with each generation's ideas becoming more nuanced, more robust, more reflective of genuine dialectical engagement.

The Three Phases

Phase 1 builds the debate infrastructure: the LangGraph orchestrator, the 4-stage speech pipeline, the judge ensemble, and the fine-tuning curriculum that teaches Qwen3-30B-A3B to run the pipeline autonomously. Critically, it also builds the belief tree system — the Bayesian epistemological foundation that everything else rests on.

Phase 2 implements the memory mechanism: HEXIS gives agents persistent experiential dispositions via Q-modulation of attention. The conviction system enables Bayesian credence updating from within M-space, and the permission hypothesis shows how implicit and explicit memory cooperate to produce experience-grounded behavior.

Phase 3 tackles the hardest problem: aggregating implicit memories across a population. How do you federate dispositions that agents can't even articulate? How do you preserve diversity while sharing insights? How do you compress a population's collective experience back into something an individual can carry?