Probabilistic arbitration is the hidden process by which a model selects one formulation among several defensible candidates because one feels more reusable, more stable, or more likely under compression.
What the phenomenon looks like
The two formulations may both be plausible. They may differ in perimeter, tone, hierarchy, or implication rather than in obvious factuality. The answer nevertheless keeps only one, and the discarded branch disappears from view.
Why it happens
Generative synthesis cannot present every branch every time. It therefore ranks formulations according to learned salience, recurrence, and contextual fit, even when the canonical source would have required explicit qualification instead of silent choice.
Why it matters
That silent selection matters because it produces public stability from internal probability. What the user sees as “the answer” may only be the highest-scoring branch of an unresolved interpretive field.
What must be governed
- Identify where several formulations are canonically admissible and where only one is.
- Surface branch conditions instead of letting probability alone choose the public formulation.
- Audit repeated selection patterns to see which phrasing the model systematically privileges.
