# Research report - User-Delegated BYOC v0.1

**Research cut-off:** 30 September 2026  
**Purpose:** fact-check the prior-art and novelty boundary before any publication.

## Method

Reviewed primary documentation for the named commercial products and standards. Red-team searches covered IETF/RFC material, MCP documentation, official product documentation, arXiv, GitHub, decentralized inference, and compute-marketplace material. The paper cites primary sources wherever available; QVAC's explanatory article and Leash's repository are explicitly marked as product/project sources rather than standards.

## Strongest prior art

1. **LM Studio LM Link + Tailscale** is the closest mature user-device precedent. It combines linked-device discovery, remote local models, end-to-end encrypted connectivity, and integrations with Codex/Claude Code. Its scope is a LM Studio-controlled linked-device ecosystem, not an independent consent layer for arbitrary applications.
2. **BazaarLink BYOC** is the closest gateway-shaped product found. It brings Ollama/LM Studio/vLLM/llama.cpp hardware behind a single API. Its published architecture places BazaarLink's gateway in the request path, so it does not satisfy the paper's optional plaintext-independent-broker property.
3. **QVAC and Leash** show that encrypted peer-to-peer delegated inference and peer capability tracking are already being implemented. They defeat any broad claim that peer-delegated inference is novel. Their published material indicates SDK/ecosystem or paired-peer scope, not broad third-party application interoperability.
4. **Tailscale** is the clearest control-plane/data-plane precedent: coordination and identity/policy functions can be separated from encrypted device data transport, including relay fallback.
5. **OpenAI Sign in with ChatGPT** demonstrates user-carried entitlement for supported applications, but not user-selected compute.
6. **NVIDIA DGX Cloud Lepton BYOC** proves that customer-owned machines can join a managed AI platform; its enterprise node-management meaning is distinct from user-delegated application compute.

## Falsification result

No reviewed source demonstrated all eight requested characteristics together at broad, neutral interoperability scope:

1. user identity;
2. user-controlled or user-selected AI compute;
3. arbitrary third-party applications;
4. capability discovery;
5. scoped user consent/delegation;
6. model/provider independence;
7. revocation; and
8. an authorization broker outside plaintext inference data flow.

This is a **time-bounded research finding**, not proof of absence. The paper therefore says “appears to be missing” and names the closest counterexamples.

## Remaining uncertainties and conservative choices

- **Microsoft Foundry Local:** Official documentation verifies local OpenAI-compatible serving. The paper treats it as enterprise/local-serving prior art and deliberately does not make a detailed claim about Microsoft identity flows, because that distinction is not central to the user-centered scope of this concept.
- **Warp data path:** Warp documents custom endpoints but the reviewed material does not establish an end-to-end request path for each mode. The table says *Unclear* rather than claiming bypass.
- **LM Link:** LM Studio explicitly states that the Hub cannot see chats and that communications between linked devices are end-to-end encrypted. This supports the Hub-discovery/data-plane distinction, not a general claim about every piece of device metadata.
- **BazaarLink:** Its public architecture makes the gateway path clear, but a public page alone is not a complete security assessment. The paper only says it is in the declared request path.
- **“Arbitrary applications”:** This is a deliberately demanding criterion. Configurable OpenAI-compatible clients are not treated as proof of a neutral cross-application delegation standard.

## Publication gate

The requested local deliverables are present:

- formatted Markdown paper;
- plain-text paper;
- PDF paper;
- Hugging Face README with metadata;
- comparison table and references;
- this research report.
- a local pre-publication repository preview.

No Hugging Face repository was inspected, changed, pushed, or published. A remote repository diff cannot exist until a target repository is supplied or created with explicit approval. The local artifact preview is the file inventory below.

```text
byoc_concept_paper/
  README.md
  PREPUBLICATION_PREVIEW.md
  assets/switchboard.png
  paper/Bring_Your_Own_Compute.md
  paper/Bring_Your_Own_Compute.txt
  paper/Bring_Your_Own_Compute.pdf
  research/RESEARCH_REPORT.md
```