Execution Cluster
Depending on the router's execution plan, up to four distinct Python-based agents act on the BEV data simultaneously. These agents execute pure mathematical and deep-learning logic to prevent LLM hallucination.
Retrieval Agent
The Retrieval Agent identifies raid scenarios that are mathematically similar to the input clip. It operates on a database of preprocessed Bird's Eye View sequences using FAISS similarity matching.
It uses a dual-track architecture: Track-1 employs a Spatial Graph Transformer combined with a Temporal Transformer, and Track-2 uses a Dilated Temporal Convolutional Network. Both tracks project outputs into a shared embedding space that is indexed into a knowledge base. If a defensive formation matches a failed tackle from the 2022 season, it instantly pulls that precedent.
Analysis Agent
The Analysis Agent evaluates the defensive structure using a Spatial-Temporal Graph Convolutional Network (ST-GCN).
The ST-GCN captures key structural properties such as the compactness of the defensive formation, the spatial coverage of defenders, and the presence of gaps or weak zones. If the geometric area of the defensive chain suddenly expands before contact, the agent flags a structural break using an MLP prediction head that outputs seven distinct tactical scores.
Simulation Agent
The Simulation Agent explores alternative scenarios by generating possible future trajectories. It operates as a diffusion-based trajectory generator.
The BEV JSON input is ingested and combined with timestep embeddings. It passes through a Social Temporal Block comprising raider intent conditioning and social processing across an iterative forward loop (T=0 to T=50). The reverse denoising loop then applies a diffusion scheduler to output a simulated BEV JSON—allowing coaches to literally ask "What if the raider stepped left instead of right?"
Description Agent
The Description Agent converts structured Bird's Eye View data into a textual representation. This transformation is absolutely essential for enabling reasoning at a higher LLM level.
The agent processes frame-wise BEV coordinates and role information to generate structured natural language statements that describe the spatial state of the raid at each timestep. It encodes relative positioning, movement direction, and proximity between the raider and defenders across the raid.
Semantic Bridge
Without the Description Agent, the Synthesizer LLM would be forced to read raw [x, y] coordinates, leading to massive hallucinations. By translating geometry into semantic text (e.g., "The raider moved 2 meters towards the left corner"), the LLM receives context it can actually understand.