FRACTAL introduces a self-similar multi-agent architecture for recursive goal decomposition. Every level of the decomposition tree repeats the same five-component pattern: Orchestrator, Ensemble, Knowledge Ledger, Decision Maker, Meta‑Observer. The design provides formal bounds on communication, orchestration criteria, convergence, and depth.
For communication, a two-phase blind‑proposal/stress‑test protocol yields intra‑node cost linear in ensemble size and round count, O(kr), whereas free‑form debate incurs quadratic cost. This result assumes a fixed number of evaluators.
The orchestration criterion separates generative and evaluative phases via a measurable diversity‑saturation threshold.
Regarding convergence, a supermajority rule guarantees explicit probability bounds, and a depth bound is derived from an operationalized goal‑complexity measure.
Inter‑layer escalation uses a Bayesian posterior update that unifies hysteresis and contract‑breach signaling into a single mechanism.
The paper illustrates execution with a worked example and a multi‑layer case study but does not present quantitative performance. The authors explicitly discuss formal and inductive‑bias limitations and propose a compatible bounded self‑improvement direction for future work. "Fractal" is used as a metaphor for self‑similarity rather than a strict mathematical property. Review