Claims and Scope

Generated from claims.json. How can storage precision, execution precision, residency and data movement be coordinated under a constrained GPU-memory budget?

Implemented and measured claims

• Selective INT4 compression of routed experts with a separate FP16 non-routed path. Status: implemented-in-recorded-paths. Source
• Compact expert GPU execution with FP16 activations and recorded scalar/grouped executors. Status: implemented-in-recorded-paths. Source
• Partial residency observations under the selected campaign contract. Status: measured. Experiment: residency39.
• OLMoE c64 full residency observations with Grouped T3. Status: measured. Experiment: residency39.
• Complete IFEval instruction-following evaluation of retained Horizon responses. Status: measured. Experiment: ifeval541.

Not claimed

• fastest MoE runtime.
• SOTA throughput.
• runtime-to-runtime superiority.
• superiority over llama.cpp.
• general BF16 quality equivalence.
• validated frontier-scale deployment.
• independently reproduced runtime generation.
• world-first quantization, offloading or caching.

Experimental interpretation

• OLMoE c30-c60 use PRODUCT-core scalar; c64 uses Grouped T3. This is not a residency-only comparison.
• IFEval 541 and the earlier 120-pair behavioral study are separate experiments.
• Published artifacts enable their declared verification, not full independent runtime reproduction.
• A shared architectural explanation does not certify every executor against every ADR 0009 invariant.

Runtime source access: private by author decision. Public evidence verification does not rerun the private runtime. Evidence and verification identifies each admitted experiment and its source artifacts.