Governance & Community Guidelines
Decision-making model, contribution process, and maintainership structure of the AMEVA Open-Source Foundation (AOSF)
1. Governance Model
AOSF operates under a Benevolent Dictator For Life (BDFL) model with meritocratic advisory input. The founding maintainer (@uno-km) holds final decision authority on all technical direction, releases, and ecosystem policy. This structure is chosen explicitly over committee-based governance to maintain engineering velocity and eliminate deadlocks on hardware-specific decisions.
2. Technical Decision Criteria
All technical decisions — kernel selection, API changes, backend priorities, release timing — are governed by the following ordered criteria:
- Real-device measurement: Claims must be substantiated by measured results on the physical 5-device fleet (S25, S21, S20, A35, A53) under documented conditions (CPU affinity, GPU clock, alternating measurement).
- Correctness first: Output parity vs. reference (CPU f32) must be verified before a GPU path is considered production-ready. KLD measurements are required for LLM paths.
- Zero silent fallback: Any change that introduces a CPU fallback path without an explicit error or warning is rejected regardless of performance gain.
- ABI stability: Changes to the ggml-ameva C ABI require an increment of the version probe (
ameva_abi_check) and a documented migration path.
3. Contribution Process
| Contribution Type | Process | Acceptance Criteria |
|---|---|---|
| Bug Report | Open a GitHub Issue with: device, OS version, repro command, full log output. | Confirmed reproducible on at least one fleet device. |
| Performance Improvement | Open a PR with: before/after t/s measurements (prompt + generation, alternating), device, conditions. | Statistically significant improvement on at least one device; no regression on others. Parity check required. |
| New Kernel / Shader | Submit as a draft PR first. Discuss design in the Issue before writing the full implementation. | Passes ameva_qmm_check 516/516 and test-backend-ops full suite on at least S25 and A35. |
| Documentation | PR against the dev/uno-km portal tree. Include English and Korean versions for major additions. |
Factually accurate; no unverified claims; follows Zero-Hype policy (no "100%", "guaranteed", "zero-*" absolute language). |
| Commercial Inquiry | Email amv.foundation.kr@gmail.com with use case, scale, and timeline. |
Evaluated case by case. SDK licensing available under Commercial License. |
4. Release Process
Releases follow a verified fleet sign-off procedure:
- All 5 devices pass
verify_kv.shandtest_rel.sh. test-backend-opsfull suite passes on S25, S20, A35, A53, S21.- KLD comparison vs. CPU reference passes on all 5 devices (parity gate).
- GitHub Release is created first; PyPI wheel upload follows; npm publish last.
- Release is tagged
non-latestuntil fleet production rollout is manually approved by the maintainer.
5. Zero-Hype Policy
All AOSF documentation and communications comply with the following language rules, aligned with OpenSSF and Apache-2.0 community standards:
- Absolute performance claims ("100%", "zero-latency", "guaranteed") are prohibited unless backed by a specific measurement citation.
- Use "validated by automated checks", "measured at X ms under conditions Y", "low allocation footprint" rather than superlative language.
- Benchmark tables must include: device name, SoC, measurement date or release version, and conditions (threads, ctx size, clock state).
6. Code of Conduct
AOSF adopts the Contributor Covenant v2.1 as its code of conduct. In brief: technical criticism is welcome and expected; personal attacks, harassment, or bad-faith engagement are grounds for removal. The maintainer's decision on conduct matters is final.