Architecture

ZhenLayer separates the problem into two cooperating models. Validators run a containerized, physics-heavy BOPTEST emulator that produces deterministic ground-truth measurements for a building under a defined weather and occupancy schedule. Miners receive only the training slice of those measurements and fit a simplified RC-network model of the building, submitting back the calibrated parameters.

Summary

• Validator model: BOPTEST FMU, Docker-containerized, deterministic. Training and held-out periods are split via SHA-256(round_id) — miners never see the held-out period.
• Miner model: ~5-parameter RC network (resistances + capacitance), sub-second per simulation, portable across optimizers.
• Synapse: CalibrationSynapse ships training measurements out; miners respond with calibrated_params within a per-round budget (1,000 evaluations).
• Verification: exactly one re-simulation per submission against the held-out period, scored with CVRMSE / NMBE / R² / convergence.
• Weights: 5% relative score floor, power-law amplification (p=2), normalized to sum=1.

Canonical source

The living specification is here:

• zhen-subnet/docs/ARCHITECTURE.md