Why would a model trained only on short decays fail differently than one given no training data at all?

This question seems to be asking about machine learning generalization rather than frontier physics or propulsion—it's not directly connected to the Antigravity archive's scope. That said, I can address it using the learning theory in these papers.

A model trained only on short decays would fail systematically in ways a blank slate wouldn't, because it has learned spurious patterns specific to that regime. Here's why:

The key insight comes from the two-phase learning model [AG-2025.04-1127]: networks first fit the training data rapidly, then enter a slower "compression" phase where they learn which patterns generalize and which are accidents of the training set. A model trained *only* on short decays completes phase one but may never fully compress away the short-decay-specific features—it has overfit to a narrow regime.

By contrast, an untrained model has no learned biases at all. When you then train it on a broader dataset (or ask it to generalize), it starts from scratch and can learn the true underlying physics without interference from short-decay artifacts.

More concretely: the error analysis in [AG-2026.01-858] suggests errors accumulate when a model's learned representations don't capture the true structure of a task. A short-decay-trained model has encoded misleading representations; it doesn't just "know less"—it has learned the *wrong* relationships, which can be harder to unlearn than starting blank.

This archive focuses on frontier physics (propulsion, relativity, quantum mechanics); for detailed guidance on training regimes in machine learning, you'd want pedagogical ML literature rather than these physics-adjacent papers.