A calculation is originally built by one team and validated against known inputs. Over time, a second team builds their own version because the original is too slow, too rigid, or inaccessible. A third team pulls results from both and averages them. Input data sources shift. A library is upgraded. A rounding convention changes. None of these changes are individually significant. Together, they produce divergence that surfaces as unexplained variance in reports.

Symptoms that indicate model instability is active.

When reported figures cannot be reproduced or explained, decision-makers lose confidence in the infrastructure. Capital allocation, hedging, and regulatory reporting all depend on numbers that are deterministic and auditable. Model instability erodes this foundation. In a stress event, it becomes impossible to distinguish between a real change in exposure and an artifact of calculation drift.

We establish a single, versioned source of truth for calculation logic. Every model has an explicit definition, a version identifier, and a reproducible build. Inputs are captured at execution time so that any result can be replayed and explained. Validation layers compare outputs across versions and flag divergence before it reaches downstream consumers. The goal is not to prevent models from changing, it is to make every change visible, intentional, and traceable.

We encountered model instability across all three proof-of-work engagements. At BatteryOS, parallel implementations of the same optimization logic produced divergent results. In our ETRM infrastructure work, P&L attribution models had been forked across desks without version control. At Greenflash, calculation logic had drifted between real-time and batch paths. Our Calculations mandate exists because of how consistently this signal appears.