ADR-001: Physics-Centric, Simulator-in-the-Loop Architecture

Status: Accepted
Date: 2025-12-27

Context

Reservoir simulation failures commonly manifest as numerical effects (non-convergence, timestep collapse, unstable solver behavior) rather than as clean logical violations. Purely text-driven or static rule-based AI assistance risks producing plausible but unsafe recommendations.

Decision

CLARISSA treats simulators as first-class learning substrates. Simulator execution, solver diagnostics, and numerical outcomes provide primary feedback signals for learning and validation.

Consequences

Positive

• Learning is grounded in physical/numerical consequences.
• Architecture remains compatible with multiple simulators.

Negative

• Requires adapter infrastructure and reproducible execution environments.

Neutral / Open

• Specific reward shaping strategies may evolve by phase.