A balance change that does not move the outcome means the dominant term is elsewhere

Fact

When a tuning parameter is changed and the measured outcome does not move, that is a fact about the model, not about the parameter.

Halving a boss's projectile damage in a scripted fight left the win rate, the time of death and the boss's remaining health identical across every seed. The reason was that projectiles were never the problem: most of the health loss was self-inflicted blast damage from the player's own bombs. Logging each health change with what caused it found it in a single run.

The same signature shows up whenever one term dominates a sum — contact damage swamping ranged damage, one enemy type swamping the rest, a fixed cost swamping a scaling one.

Why it matters

The natural reading of "that change did nothing" is "the change was too small", so the parameter gets pushed further and further from a sensible value while the term that actually decides the outcome is never touched. The design ends up mis-scaled in two places at once, and every later conclusion drawn from the same harness inherits the blind spot while the numbers keep agreeing with each other.

How to apply

• Attribute before tuning. Log every change to the resource being balanced alongside its cause, then total by cause. The dominant term is usually a surprise.
• Treat an inert knob as a stop signal, not as an invitation to a bigger change.
• Include self-inflicted and environmental sources in the attribution. They are the ones nobody thinks to model, and the ones a scripted agent walks into.
• Check that the attributed totals sum to the observed change; a shortfall is a source you have not accounted for.

Related: [[simulated-error-is-per-decision]].