z0evals

Reproducible evaluations and public research reports for the Zer0 ecosystem.

z0evals is the frozen evidence/publication layer. It does not run production routing, own runtime policy, or replace the experiment engines that produce candidate systems.

Ownership

frontier-kb      external evidence / literature
       │
Evolution Lab    experiment design + execution + training
       │
Tokenomics       neutral measurements / receipts / verified outcomes
       │
z0evals          frozen datasets + analysis + figures + public reports
       │
z0intelligence   promotion / shadow / canary decisions
       │
Kerdoios         resource placement among allowed candidates

A study enters z0evals only after its input revisions are pinned and its results can be reproduced or independently inspected.

First study

studies/slm-router-v0/ is reserved for the local SLM routing study produced from the September 2026 z0intelligence / Evolution Lab work.

The initial report will investigate:

• compiler-first legal-action filtering
• FunctionGemma, Hammer 2.1, Nemotron-Orchestrator and Qwen routing behavior
• latency / quality / GPU tradeoffs on an RTX 3080 Ti
• bounded tool choice versus authentic orchestration
• the basis for a continuously distilled Q-Route policy

Layout

studies/    frozen study manifests, analysis notes, public data and figures
posts/      publication-ready narratives derived from studies
schemas/    versioned contracts
scripts/    import, validation and site-generation helpers
site/       generated GitHub Pages output

Importing a local DSH result

Do not push a whole local worktree or private trace directory.

Use an import spec that pins the source repository and commit, and explicitly allowlists each artifact to publish:

python -m pip install -r requirements.txt
python scripts/import_local.py imports/slm-router-v0.local.yaml
python scripts/validate.py

See imports/slm-router-v0.example.yaml.

Rules

1. Every claim in a post should resolve to a frozen study artifact.
2. Source-reported benchmark numbers and our reproduced measurements are different fields.
3. Raw secrets, credentials, personal traces and private prompts never enter this public repository.
4. Training/evaluation splits must preserve work-item lineage.
5. The system that generated a label cannot silently become its own verifier.
6. A study may report an incomplete or negative result. Do not hide failed hypotheses.

Related repos

• https://github.com/kvnloo/z0
• https://github.com/kvnloo/z0intelligence
• https://github.com/kvnloo/evolution-lab
• https://github.com/kvnloo/tokenomics
• https://github.com/kvnloo/kerdoios
• https://github.com/kvnloo/frontier-kb