Public paper scaffold. Release status: scaffolded. License posture requires human review.
This repository is Francisco Abner Rivera's public paper scaffold for exploring expert pruning and sparse model efficiency as an Energy Per Intelligence research direction.
The work belongs to the Franzabner public technical brand. It is not a released pruning workflow, not a deployed agent, not a private harness disclosure, not a validated benchmark, not a model release, not a dataset release, not a Hugging Face artifact, not a deployment, and not a client result.
| Item | Status |
|---|---|
| Public posture | Paper scaffold |
| Release status | Scaffolded |
| Method status | No released pruning workflow |
| Benchmark status | Not validated |
| Result status | No evaluated results released |
| Model status | No model weights released |
| Dataset status | No dataset released |
| Agent or harness status | No deployed agent and no private harness published |
| Hugging Face status | No model, dataset, or Space created by this repo |
| License posture | Existing license files are unchanged; human review required before any license change or external reliance |
Mixture-of-Experts and sparse model architectures raise a practical engineering question: can removing, disabling, or routing around selected experts reduce energy per useful output without making accuracy-only claims?
The public research question is:
Can expert pruning improve Energy Per Intelligence while preserving a clear boundary between public paper scaffolding and private model or agent workflows?
This repository may describe hypotheses, public-safe terminology, and paper structure. It does not claim that an efficient operating point has been measured, that any pruning workflow is released, or that any agent/harness workflow is public.
The intended review path is:
- Define public-safe expert-pruning questions and terminology.
- Keep model, dataset, agent, harness, and measurement choices under human review.
- Document planned measurement fields without publishing fake benchmark or eval results.
- Record limitations before any report, card, or external reference is published.
- Route any future Hugging Face-facing card through the release discipline in
hf-card-templates.
| Repo | Role |
|---|---|
| franzabner-proof-stack | Public proof routing and status discipline |
| energy-per-intelligence | EPI metric framing and research surface |
| epi-bench | EPI tooling scaffold; no validated benchmark claim here |
| epi-meter | Public hardware measurement scaffold; no released measurement claim here |
| hf-card-templates | Hugging Face release-readiness templates and boundary gates |
- Paper scaffold for expert pruning, sparse model efficiency, and Energy Per Intelligence.
- Public-safe research questions and status language.
- Skeleton code and analysis placeholders.
- Boundary notes for future measurement, report, or card publication.
This repository does not claim:
- a released expert-pruning workflow;
- a deployed agent;
- a published private harness;
- a validated benchmark;
- evaluated results;
- model weights;
- a dataset;
- a Hugging Face model, dataset, or Space;
- a deployment;
- client or customer use;
- revenue outcomes;
- production readiness;
- a private corpus, training corpus, endpoint, private harness, private agent workflow, or company-private infrastructure.
Human review is required before:
- publishing measured energy or accuracy results;
- publishing benchmark outputs;
- publishing raw traces or datasets;
- publishing model artifacts or weights;
- publishing agent or harness details;
- linking to any external Hugging Face artifact;
- changing license posture;
- claiming release, deployment, client usage, production benchmark status, or validated method status.
Public examples must be synthetic, scaffolded, or explicitly approved. Private corpora, private model weights, private endpoints, private agent harnesses, private training workflows, private infrastructure, and sealed implementation details stay out of this repository.
This repo keeps expert-pruning EPI work public as a paper scaffold while holding method, model, agent, harness, data, and result claims behind review.