Horizon RunMap — Technical Documentation and Experimental Evidence

Browse the architecture, original experiment reports and public evidence. Start with a document below, then follow its source and verification links.

Architecture
Compact expert representation, immutable host authority, GPU residency, miss transport and executor boundaries in Horizon RunMap.
Technical Notes
The original publication edition v3 technical source map: backend contracts, immutable cache lifecycle, measurement boundaries and release verification.
Experimental Campaign / Residency Report
The original consolidated residency campaign report: thirteen configurations, executor identities, timing, memory and transfer observations.
Evidence and Verification
Source-bound observations from three separate Horizon RunMap experiments, with public artifacts, verification methods and reproduction scope.
Quality Evaluation
Separate reading paths for the OLMoE 0924 IFEval 541 benchmark and the earlier 120-pair BF16/Horizon structural comparison.
Related Work
Architecture comparison with MoE offloading systems and a dated audit of HOBBIT, llama.cpp expert-cache proposals and vLLM offloading work.
Claims and Scope
Implemented and measured Horizon RunMap claims, explicit non-claims and experimental interpretation constraints, generated from claims.json.
Raw Evidence Releases
Manifest-addressed public evidence editions, original Markdown, JSON records, retained response bundles and standalone verification tools.

Experimental interpretation notes

OLMoE c30–c60 use PRODUCT-core scalar; c64 uses Grouped T3. The executor changes, so this is not a residency-only ablation.
Grouped T3 and scalar results identify different executor paths; keep that identity attached to each observation.
IFEval 541 and the earlier 120-pair BF16 × Horizon study are separate experiments.
Frontier-scale models provide motivation and context, not the experimental population of the published campaign.
Released artifacts enable their declared verification; this is distinct from independent end-to-end runtime reproduction.
Horizon does not claim SOTA throughput or universal runtime superiority.
Expert offloading, expert caching and expert quantization individually are established prior art.