DEJAN's Content Optimizer implements a closed-loop, multi-role Bayesian experimental framework designed to uncover what makes language models prefer specific content:

• Hypothesis Generation and Classification: A Seeder examines the target query, baseline content, and competitive results to generate initial hypotheses (claims) with prior probabilities. A Classifier then maps each claim into a structured, two-level Rank Factor Class taxonomy (spanning dimensions like alignment, substance, architecture, style, framing, and proof).
• Two-Stage Acquisition: Each round, a two-stage acquisition policy first selects a rank-factor class weighted by historical freshness (\(1 / (1 + n)^{\text{diversity\_weight}}\)) to enforce exploration, then applies policies like Thompson Sampling, Upper Confidence Bound (UCB), or \(\epsilon\)-greedy to sample a specific claim within that class.
• Targeted Variation: The Ideator generates an updated variant isolated to testing that hypothesis, either modifying a short extractive snippet or generating precise line-numbered patch edits (<replace>, <insert>) directly against full page markdown.
• Sampling and Consensus Evaluation: A Ranker judges the new candidate against competitor documents over multiple temperature-sampled passes (typically 5 to 11 samples) to compute a robust median rank, supplemented by optional cross-model observer scoring to observe non-target model reactions without polluting the primary optimization signal.
• Bayesian Belief Updates: If the candidate improves upon the current best rank, it earns a binary reward (\(1\) for a win, \(0\) otherwise), updating the claim's \(\text{Beta}(\alpha, \beta)\) posterior distribution (\( \alpha \leftarrow \alpha + \text{reward}, \beta \leftarrow \beta + (1 - \text{reward}) \)).

Runs iterate until the target reaches rank 1, converges via settled claims, or exhausts its round budget, with an autonomous custodian proposing new factors whenever unclassified winning hypotheses accumulate sufficient evidence.