Project Summary

LLM agents increasingly choose which tool to call from a large registry, and which-tool-did-it-pick is a precondition for safe deployment — incident analysis, capability auditing, and red-teaming all need a routing signal humans can actually inspect. Today that signal is usually a cosine similarity score inside a 1,536-dimensional embedding space.

I’ve built and shipped Meridian, an open-source MCP that replaces opaque embedding routing with a deterministic orbital classifier: every candidate receives a physics signature, a celestial class, and a one-line decision rule explaining why it ranked where it did.

Live now:

• MCP endpoint: mcp.ask-meridian.uk

• GitHub: Meridian MCP Repository

Meridian v2.1.0 already ships with:

• OAuth 2.1 + PKCE

• stdio + Streamable-HTTP transports

• deterministic orbital routing

• 47 unit tests

• npm package + GHCR image

• Cloudflare Worker deployment

• GitHub Actions CD pipeline

This $5,000 grant funds a public benchmark for tool-routing failures:

• a labelled routing dataset

• a two-judge evaluation matrix

• an open-source eval harness

• and a reproducible write-up

The goal is to make tool-routing quality measurable in the same way perplexity standardized language-model evaluation.

Goals & Execution Plan

Goal

Make tool-routing failure rates a measurable, comparable, and citable metric.

Deliverables (4 Months)

1. Labelled task→skill dataset (~500 pairs)
Tasks spanning coding, research, operations, and creative domains. Each task includes:

• one correct skill

• four distractors

• paid human labelling via Prolific or Surge AI

Published publicly on HuggingFace.

2. Two-judge evaluation matrix
Routing judged by:

• Anthropic Sonnet 4.6

• xAI Grok-4

Plus:

• self-hosted BGE-large-en embedding baseline running on Modal

This creates a reproducible comparison between:

• frontier-model routing

• embedding-based routing

• deterministic routing

3. Open-source evaluation harness (CLI)
One-command evaluation against:

• Meridian

• LangChain routers

• LlamaIndex routers

• vanilla embedding systems

• MCP-compatible routing systems

4. Public write-up
A LessWrong / Alignment Forum post with:

• reproducible code

• benchmark methodology

• routing analysis

• published dataset

Success Criteria

• Dataset downloadable on HuggingFace by month 3

• Harness runnable externally with one command by month 4

• At least one external framework adopts or cites the eval within 6 months

• Cross-classifier results produce publishable insight regardless of outcome:

  • If deterministic routing wins → interpretable routing is viable

  • If it loses → interpretable routing carries measurable capability cost

Funding Breakdown

• Anthropic + xAI API credits (judges + baselines) — $1,500

• Human labelling (~500 pairs) — $1,400

• Starlink hardware + 4 months service — $950

• Cloudflare Workers Paid + GitHub Models inference — $400

• GPU compute (Modal / Lambda) — $300

• Buffer (~10%) — $450

Total: $5,000

No salary or stipend is included. Development work is performed independently alongside contracting income.

Minimum Viable Funding: $2,500

Reduced scope:

• single judge (Sonnet 4.6 only)

• ~250 labelled pairs

• no Starlink reliability layer

Still useful, but less reproducible and less citable.

Team & Track Record

Independent solo engineer.

Recent shipped work

• Meridian MCP
  GitHub Repository

• Lens — WebXR Vision Lab pairing SmolVLM + Meridian routing
  Lens Repository
  lens.ask-meridian.uk

• Photon — photonic retrieval router using the Meridian backend
  Photon Repository

• Writing & architecture notes
  ask-meridian.uk/blog

Published work includes:

• classifier walkthroughs

• deterministic routing analysis

• OAuth operator-pays architecture

• Cloudflare Workers vs GitHub Pages deployment trade-offs

Risks & Failure Modes

1. Deterministic routing loses the benchmark

Embedding or LLM-based routing may outperform Meridian.

Outcome:

• still produces a useful public benchmark

• still yields a publishable result

• clarifies whether interpretability costs capability

2. Dataset quality issues

Routing datasets are difficult to label reliably.

Mitigation:

• second-pass review

• manual validation sampling

• reduced dataset size if noise exceeds threshold

3. No external adopters

Possible outcome:

• benchmark remains useful as a public reference

• still functions as an internal regression metric for Meridian

4. Funding shortfall

Lower funding reduces:

• dataset size

• judge diversity

• reproducibility

The project still ships regardless of funding outcome.

Prior Funding

$0 external funding raised.

All work to date has been self-funded through contracting income, with approximately:

• £1,500 (~$1,900 USD)
  spent across 2025–2026 on:

• Cloudflare Workers Paid

• GitHub Models inference overage

• domains/DNS

• monitoring

• infrastructure