Reference

Glossary

---

agent

A language model operated in a control loop where it may invoke tools, observe their output, and decide subsequent actions, in contrast to a single chat completion.

agentic loop

The control structure propose action → execute tool → append observation → re-invoke model, repeated until a stopping condition. It enables self-verification against ground truth.

AGENTS.md

A plain-Markdown file giving an assistant always-on project context (environment, data model, conventions, definition of done). Read automatically by many clients; equivalent to a copilot-instructions.md, .cursorrules, or “custom instructions” elsewhere.

branching fraction

The probability that a particle decays through a particular channel. For Λ⁰ → p π⁻ it is ≈ 63.9%.

combinatorial background

The smooth, non-peaking distribution from pairing tracks that did not originate from a common parent. It is separated from signal by a fit, not removed by selection alone.

context (context window)

The bounded set of tokens a model attends to in a step: system instructions, files, prior turns, and tool outputs. Managing its contents is “context engineering.”

EDM4eic

The ePIC event data model, an EIC extension of EDM4hep generated with PODIO. Reconstructed output is stored in an events tree whose branches are per-event collections (e.g. ReconstructedChargedParticles, MCParticles).

harness

A model together with the machinery that makes it operational: model, context, tools, and the control loop. The method of this lesson is independent of which model fills the first slot.

invariant mass

The Lorentz-invariant combination \(m = \sqrt{(\Sigma E)^2 - |\Sigma \vec{p}|^2}\) of a set of four-momenta. For the daughters of a real decay it equals the parent mass.

JSON-RPC

A lightweight remote-procedure-call protocol using JSON messages. MCP is built on JSON-RPC 2.0.

LLM (large language model)

The generative model performing reasoning and code synthesis at the center of a harness.

MCP (Model Context Protocol)

An open client–server standard for exposing tools, resources, and prompts to language-model clients, so one tool implementation serves many assistants. See https://modelcontextprotocol.io.

MCP server

A program that advertises MCP tools/resources to clients over stdio or HTTP. Here, the uproot-mcp-server provides ROOT/EDM4eic analysis tools.

PODIO

The event-data-model toolkit (AIDASoft) that generates the I/O code behind EDM4hep/EDM4eic. Its files are readable by PODIO, ROOT (RDataFrame, TTreeReader), and uproot.

reconstruction

The inference of particle four-momenta and identities from detector signals. Outputs such as ReconstructedChargedParticles.PDG are estimates, not truth.

SKILL.md

The specification file of a skill: YAML frontmatter (name, description) plus a body giving applicability, inputs, steps, and success criteria. Loaded on demand when a request matches the description.

skill

A versioned procedure: a directory containing a SKILL.md that orchestrates the MCP tools toward a defined result. Where a tool is a capability, a skill composes capabilities.

stdio transport

The MCP transport in which the client launches the server as a subprocess and exchanges JSON-RPC on standard input/output. This lesson instead uses the streamable HTTP transport: the servers run inside eic-shell (eic-mcp up) and the client connects over http://127.0.0.1:910x/mcp.

tool

An operation the model may invoke through a typed interface (name, arguments, return schema). The only channel by which an agent affects the outside world.

uproot

A pure-Python ROOT I/O library, used by the tool server to read EDM4eic files and return compact JSON without a ROOT installation.

V0

A neutral particle reconstructed from two oppositely charged tracks at a displaced vertex (e.g. Λ⁰ → p π⁻, K⁰ₛ → π⁺π⁻); the canonical invariant-mass measurement.

yield

The number of signal candidates, obtained by integrating the fitted signal component, not by a raw bin count.