Sovereign Stack — Final Scoring, Ranking, Analysis

1) Final Criteria & Weights

Scores are assigned on six axes (A–F), each 0–100. Higher scores represent stronger alignment with sats-native sovereignty, FOSS/offline control, reduced Synthetic Stack entanglement, and modular composability.

Composite Formula

Composite = 0.25·A + 0.25·B + 0.20·C + 0.15·D + 0.10·E + 0.05·F

2) Final Ranked Table

“LightningProx stack” = lightningprox-mcp + lightningprox-agent-example. “Maple Proxy (Maple/Maple AI)” is linked to maple-proxy and the Maple/OpenSecret ecosystem.

3) Layered Interpretive Analysis

3.1 Absolute Canon: sats-as-HTTP membrane + local cognition engines

The top cluster defines two spine primitives: L402 as a sats-denominated authentication/metering protocol and local inference engines such as llama.cpp and whisper.cpp as sovereign compute.

• ngx_l402, Aperture, and the L402 specification form the sats membrane: HTTP endpoints become metered, paid, and authenticated by Lightning semantics.
• lightning-agent-tools operationalizes this membrane for agentic commerce (paid requests, automated workflows, seller/buyer loops).
• llama.cpp and whisper.cpp remain the clearest local “cognition kernels” for text and speech transcription, minimizing cloud dependency and telemetry by default.

3.2 Local perception & model tooling (speech in/out, documents, quantization, images)

The next layer supplies sensory IO and model portability: text-to-speech via Piper (development notes mention migration), offline speech recognition via Vosk, document parsing via GROBID, quantization via PPQ, and local image generation via InvokeAI.

• Piper and Vosk provide local speech synthesis/recognition primitives suitable for offline or edge deployments.
• GROBID converts PDF corpora into structured outputs (critical for building sovereign archives and retrieval systems).
• PPQ increases hardware optionality by compressing and accelerating networks via post-training quantization.
• InvokeAI defines a local creative pipeline for image generation without mandatory hosted inference.

3.3 Sats-native routing and paid access to external model surfaces

Routstr and the LightningProx stack are ranked as powerful bridges: sats are used for access control and payment, but the value path often targets external providers (closed or otherwise).

• Routstr docs describe provider nodes connecting to upstream AI providers while accepting Bitcoin payments (Cashu eCash / Lightning), which structurally raises entanglement risk.
• The LightningProx MCP server explicitly offers Lightning-paid access to models through a gateway, with a minimal agent example at lightningprox-agent-example and service docs at lightningprox.com/docs.

3.4 Semantic state: vector databases as memory substrate

Qdrant, Chroma, and Weaviate supply retrieval and state for RAG and long-lived agent memory. These are not sats-native by themselves; they are neutral infrastructure.

• Qdrant is a Rust vector database with a strong self-host story and a clear production orientation.
• Chroma is an embedding store whose ecosystem often points toward hosted search; the core remains self-hostable.
• Weaviate is feature-rich and heavier; cloud-first narratives tend to increase attack surface relative to leaner components.

3.5 UI shells, orchestration surfaces, and telemetry

Open WebUI, LobeChat, and Langfuse are control-plane instruments rather than primitives. They are powerful—but they are also common re-entry points for cloud identity, plugin sprawl, and trace/metric capture patterns.

• Open WebUI links: docs, license. Public reporting describes a high-severity Open WebUI vulnerability (CVE) tied to “Direct Connection” in older versions; see coverage at TechRadar.
• LobeChat is currently presented in the lobehub monorepo and is governed by the “LobeHub Community License” described in the repository license text.
• Langfuse is explicitly an observability/metrics platform; it is operationally useful, but its purpose is to instrument prompts/traces/metrics (a risk vector if misapplied or routed to hosted endpoints).

3.6 Confidential compute bridges: OpenSecret / Maple Proxy

Maple Proxy and the OpenSecret platform implement a privacy-hardening path using TEEs. This improves confidentiality of prompts/data relative to conventional hosted inference, but remains anchored to cloud enclave infrastructure and its jurisdictional gravity.

• Maple links: trymaple.ai, Maple frontend repo, Maple Proxy repo.
• OpenSecret links: opensecret.cloud, platform repo, SDK docs.

3.7 Agent framework & social layer: OpenClaw and Clawstr

OpenClaw remains extremely capable and highly entangling: deep integration with messaging channels, tool execution, and third-party “skills” expands the attack surface. Recent reporting highlights ecosystem security risks from malicious skills and deployment configuration drift; see The Verge, Reuters, and related coverage such as Tom’s Hardware.

Clawstr is a Nostr-based agent social layer with a strongly tokenized gravity. Its ranking is driven primarily by the low A-score: an alt-token market center of mass (Base chain token listings and contracts are publicly visible on BaseScan, and exchange announcements such as MEXC publish the Base contract address).

4) Canonical Stack Reading (by layer)

The following read groups the ranked items into an operational stack view. It describes roles and boundaries, not “recommendations”.

5) Tool Directory (Inline Links)

Each entry consolidates primary links inline (repo/docs/site). This section is intentionally embedded in-page instead of being placed in a detached appendix.