# Horizon RunMap

> Built independently by Lucas Ribeiro with limited hardware and no outside
> investment, Horizon coordinates model storage, movement, computation and
> evaluation so demanding AI can work within less graphics-card memory.

Current research entry point, reviewed 2026-10-01.
Horizon RunMap is an implemented MoE inference architecture with a public
evaluation app. The RTX 5070 campaign records execution, memory and speed;
the published IFEval edition evaluates 541 prompts; the app completed the
full IFEval process on another computer with an RTX 4060.

The architecture coordinates selective expert compression, compact W4A16
execution, full or partial GPU residency and resource admission. The immutable
cache is the documented path for bounded movement when residency is partial.
The runtime source remains private by author decision; the executable,
technical documentation and evidence artifacts are publicly available.

## Start here

- [Technical brief](https://horizonrunmap.com/technical-brief.md): compact canonical evaluation context and reviewer notes.
- [Architecture](https://horizonrunmap.com/architecture.md): full execution path, residency regimes and cache contract.
- [Evidence](https://horizonrunmap.com/evidence.md): RTX 5070 campaign, published IFEval, external app execution and original source editions.
- [Public app](https://horizonrunmap.com/en/download/): download and evaluation workflow.
- [Related work](https://horizonrunmap.com/related-work.md): sourced architectural comparison, including unresolved properties.
- [Reviewer brief](https://horizonrunmap.com/en/review): architecture, achievements, evidence and interpretation.
- [AI context](https://horizonrunmap.com/ai-context.txt): portable review context.
- [Review prompt EN](https://horizonrunmap.com/review-prompt.en.txt) · [Review prompt PT-BR](https://horizonrunmap.com/review-prompt.pt-BR.txt).

## Structured records

- [Claims](https://horizonrunmap.com/claims.json): versioned achievements, source-bound observations and scope.
- [Architecture record](https://horizonrunmap.com/architecture.json): system overview and profile-specific invariants.
- [Build inventory](https://horizonrunmap.com/agent-interface-manifest.json): SHA-256 inventory of this generated interface.

## Public evidence editions

- [Performance manifest](https://horizonrunmap.com/evidence/releases/technical-preview-residency39-v3/manifest.json): 13 configurations and 468 measured responses, E3.
- [Performance method](https://horizonrunmap.com/evidence/releases/technical-preview-residency39-v3/audit-method-v1.md): curated aggregates and hash verification.
- [Technical source map](https://horizonrunmap.com/evidence/releases/technical-preview-residency39-v3/horizon-technical-notes.md): architecture contracts and historical scope.
- [Behavioral manifest](https://horizonrunmap.com/evidence/behavioral/olmoe-0924-fidelity-v2/manifest.json): 120 BF16/Horizon pairs, E3.
- [Behavioral method](https://horizonrunmap.com/evidence/behavioral/olmoe-0924-fidelity-v2/audit-method.md): retained-response structural rescoring.
- [IFEval manifest](https://horizonrunmap.com/evidence/ifeval/olmoe-0924-ifeval541-v1/manifest.json): 541 prompts, E1 standalone_benchmark.
- [IFEval method](https://horizonrunmap.com/evidence/ifeval/olmoe-0924-ifeval541-v1/audit-method.md): frozen scorer and complete retained population.

## Report

- [Current report v2](https://horizonrunmap.com/technical-report-v2.md): principal technical report.
- [Report HTML](https://horizonrunmap.com/technical-report-v2.html): readable and printable report.
- [Report PDF](https://horizonrunmap.com/technical-report-v2.pdf): portable version of the same report.
- [Historical report v1](https://horizonrunmap.com/technical-report-v1.html): preserved 2026-09-16 edition.
- [Historical report v1 Markdown](https://horizonrunmap.com/technical-report-v1.md) · [Historical report v1 PDF](https://horizonrunmap.com/technical-report-v1.pdf).

## Optional

- [English presentation](https://horizonrunmap.com/en.md): full Markdown alternative to /en.
- [Apresentação em português](https://horizonrunmap.com/pt.md): alternativa Markdown completa de /pt.
- [Performance release index](https://horizonrunmap.com/evidence/publication-index.json): current and historical editions.
- [Behavioral release index](https://horizonrunmap.com/evidence/behavioral-index.json): separate paired-study discovery.
- [IFEval release index](https://horizonrunmap.com/evidence/ifeval-index.json): separate standalone-benchmark discovery.

## Technical review context

Horizon RunMap combines selective INT4 compression of routed experts, custom W4A16 GPU execution, full or partial expert residency and a public evaluation app. Full residency is a complete execution path; the documented immutable cache serves the partial-residency path.

The RTX 5070 campaign includes an OLMoE 0924 full-residency Grouped T3 result of 61.26 post-first-token decode tokens/s at 5.99 GiB whole-worker peak VRAM ([source](https://horizonrunmap.com/evidence/releases/technical-preview-residency39-v3/olmoe-0924-instruct-c64-residency39.json)). The separately published 541-prompt IFEval edition reports 45.66% prompt-level loose accuracy and retains all responses ([source](https://horizonrunmap.com/evidence/ifeval/olmoe-0924-ifeval541-v1/observations/summary.json)). The public app completed the full IFEval process on another computer with an RTX 4060 ([project account](https://horizonrunmap.com/en#external-4060)).

Use the linked technical brief, architecture, evidence and reviewer documents for definitions, contracts and related work. Keep each numeric result with its own recorded execution.