LLMs: em-skills, NUANCE-MCP, NUScope and NUScope-TOMO¶

How a language model works with NUANCE data: what it may do, which code does it, and how every result stays traceable to the instrument, the file and the person.

The layers¶

flowchart TB
  subgraph Registry["em-skills registry (emskills.org)"]
    S1[sem-particle-size]:::skill
    S2[provenance-verify]:::skill
    S3[fib-phase-fractions]:::skill
  end
  subgraph Server["NUANCE-MCP server"]
    T[Instrument tools<br/>acquire_*, stage, optics]
    P1[NUScope plugin<br/>nuscope_* tools · sealed acquisitions]
    P2[NUScope-TOMO plugin<br/>tomo_* tools]
  end
  LLM[LLM agent<br/>Claude, Ollama ...] -- "which skill, when, how safely" --> Registry
  LLM -- "MCP tool calls" --> Server
  T --> M[(Microscope)]
  P1 --> F[(Sealed, signed .h5<br/>instrument DOI inside)]
  P2 --> V[(Tomography projects)]
  F -. "nuscope cite" .-> D[DataCite DOI<br/>IsCollectedBy instrument]
  ID[EMDataBridge identity<br/>ORCID + key] -. signs .-> F
  classDef skill fill:#ece4f6,stroke:#4E2A84

Layer	Role	Owner
em-skills	The knowledge: a `SKILL.md` per task tells an agent what a tool does, its scope, its safety class and its exact call, and ships a fixture with a reference check that CI runs	emskills-mcp/em-skills
NUANCE-MCP	The runtime: MCP server that exposes the instrument and the plugins' tools to the agent, with schema-validated inputs	NUANCE-IT/NUANCE-MCP
NUScope	Data tools (read any NUANCE file, analyze in physical units, verify) and the guarantee that every acquisition is saved as a sealed, signed file with the instrument DOI	this package
NUScope-TOMO	Tomography tools: project summaries, phase fractions with error bars, recipe runs	NUANCE-IT/NUScope-TOMO
EMDataBridge	Who: the researcher's ORCID iD and signing key	NUANCE-IT/EMDataBridge
nuance-datacite	Which instrument, and the DOI of the dataset	NUANCE-IT/NUANCE-Data-Cite

A skill is not code that runs inside the model. It is a contract: call this MCP tool with these arguments for this kind of data, expect this result, and here is the test that proves the tool does it. The tools themselves live in NUScope and NUScope-TOMO and are served by NUANCE-MCP.

Skills backed by NUScope¶

Skill	Tool	Package	Safety
`sem-particle-size`	`nuscope_analyze` (op `particles`)	NUScope	read-only
`provenance-verify`	`nuscope_verify`	NUScope	read-only
`fib-phase-fractions`	`tomo_phase_fractions`	NUScope-TOMO ≥ 0.3.2	read-only

Each one pins the released package in its requirements.txt, and its verify.py recovers a known ground truth from a synthetic fixture with that exact version. If a NUScope release changes a result, the registry's CI fails before an agent ever sees the change.

Existing skills that NUScope can serve¶

The registry's first skills describe tasks NUScope and NUScope-TOMO already do, or could do with small additions. Backing them with NUScope gives them file readers for every NUANCE instrument, calibration from metadata, and provenance.

Skill	NUScope / NUScope-TOMO today	Gap
`scalebar-calibrate`	`calibrate(pixel_size_nm, reason)` records the correction in the history; `fft` + `radial_profile` measure a grating period	fit the cross-grating pitch automatically
`lattice-dspacing`	`fft` (1/nm axes) + `radial_profile` give periodicities in nm	peak fitting with uncertainty
`eels-edge-id`	reads EELS spectra and spectrum images with calibrated energy axes	edge database and fine-structure fitting
`fib-mill-seg`	NUScope-TOMO segmentation (random forest, U-Net) and particle analysis in 3D	lift-out path planning
`tomo-align`	NUScope-TOMO aligns FIB-SEM slice stacks	tilt-series (electron tomography): planned
`sem-topography`, `stem-drift-lock`, `edax-phase-map`, cryo-EM skills	readers and provenance	the methods themselves

Rules for skills built on NUScope¶

Verify before trusting. An agent should call nuscope_verify (skill provenance-verify) on any file before analyzing or publishing it, and stop if it is tampered.
Calibration comes from the file. Tools read pixel size, dispersion and camera length from the instrument's metadata. A skill never asks the model to type a pixel size, and a manual correction goes through calibrate with a reason, which is kept in the history.
Results are new, sealed files. Image results are written next to the input with the full history; inputs are never modified.
People sign, agents don't. Files are signed with the identity of the person running the server. An agent cannot sign on anyone else's behalf, and nothing is countersigned without that person.
DOIs are drafts. An agent may prepare DataCite metadata (nuscope cite). Creating even a draft DOI uses the facility's credentials and should be asked for. Publishing is permanent and is never automatic.
Safety class in the manifest. Data skills are read-only. Skills that act on the microscope declare it in ## Safety and are served by NUANCE-MCP's instrument tools, with their schema limits.

Adding a NUScope-backed skill¶

Make sure the operation is reachable as an MCP tool (nuscope_analyze covers every operation in nuscope.analyze; add a dedicated tool for anything else).
In em-skills/skills/<slug>/: SKILL.md (front matter: kind: MCP server, tool_call: <tool>, org: NUANCE-IT, repo: NUANCE-IT/NUScope), examples/invoke.json, fixtures/generate.py and fixtures/sample/ (small, deterministic), verify.py that recovers the ground truth with the released package, and requirements.txt pinning it.
Run python3 scripts/validate_skills.py --run in em-skills.