Projectmouse-brain-ccf-joinOctober 16, 2026

Joining mouse brain datasets for AI agents

The design for a library, an MCP server and Skills that let AI agents answer questions spanning Allen Institute and Janelia mouse brain datasets. Each join checks that both sides use compatible atlas and taxonomy versions, and each answer carries a record that a verifier reruns. A pre-registered experiment measures what each part adds, in Claude Science and in other agent harnesses.

In short · design review

Questions that span mouse brain datasets need joins across file formats, coordinate conventions, atlas versions and cell-type taxonomies. Each crossing has a trap that gives a plausible wrong answer and raises no error. This design puts each source behind an adapter that translates it into one shared model, checks every join for compatible versions before it runs, and records what was read so a verifier can rerun the answer. Agents reach it through an MCP server and Skills. A controlled experiment, with gold answers taken from published sources, measures what each part adds.

Question families
2
probe to cell types · projection-anchored
Data sources
3
ABC Atlas · DANDI · MouseLight
Requirements
19
each with a check
Tool conditions
4
Agent SDK; subsets in two more harnesses

Summary from the design review, v1

  • Problem: Many questions about the mouse brain need several datasets joined in one shared 3D atlas. Each crossing between datasets has a trap that gives a wrong answer without an error, and nobody can check what an agent joined.
  • Design: A Python library translates each source, through its own adapter, into one model of atlas versions, cell-type taxonomies and observations. A join runs only between compatible versions. Every answer records what it read and every fix it applied.
  • Agent surface: An MCP server and two Skills give agents tools that find data, plan joins, refuse invalid ones, and run code in a sandbox next to the data. They work in Claude Science and in Claude Code.
  • Infrastructure: Batch work runs as Slurm jobs in Apptainer containers, which is how Claude Science computes on a cluster. The same job scripts run on the homelab's Slurm, on a Slurm cluster in AWS at the end, and on a partner's cluster. The MCP server runs as a container service.
  • Evaluation: A pre-registered experiment runs four tool conditions in the Agent SDK, and subsets of them in Managed Agents and Claude Science, against gold answers from published sources. A component stays only if it measurably helps.

Phases P0 to P5

  1. P0frameCurrent phase

    What happens
    This review
    Exit
    the review is accepted
    DeliverableEngineering design review: mouse-brain-ccf-joinv1

  2. P1spikeNot started

    What happens
    One reference script per v1 join that reproduces its trap and its fix. Claude Science loads a stub connector and an imported Skill, connects to a Slurm login node on the homelab, and submits a job. An Apptainer job on the agent partition reads a MERFISH slice in place
    Exit
    each script matches a hand-checked value; Claude Science's job runs through sbatch; the Apptainer job's isolation flags and read rate are recorded

  3. P2core and adaptersNot started

    What happens
    The repository mouse-brain-ccf-join is created. The model, the crosswalks, the three adapters and their contract tests
    Exit
    the P1 scripts' answers reproduce through the library

  4. P3agent surfaceNot started

    What happens
    The MCP server, the Skills and the Claude Science package. The server deployed on k3s, and the agent partition on the homelab's Slurm
    Exit
    one v1 question answered end to end in Claude Code and in Claude Science, with a record the verifier reruns

  5. P4experimentNot started

    What happens
    The pilot, the freeze, then the full sweep as Slurm job arrays on the homelab
    Exit
    results published, failures first

  6. P5AWS build and releaseNot started

    What happens
    A Slurm cluster on AWS from infrastructure code, building on an earlier Terraform-built Slurm cluster that Claude Science has already driven, with the MCP server beside it. All of it in a public repository. The released package, the public endpoint, and the Managed Agents harness
    Exit
    someone new stands up the AWS build and reruns one condition from the documentation alone

Pages 8

  1. Design reviewv1Engineering design review: mouse-brain-ccf-joinThe design for a library, an MCP server and Skills that let AI agents answer questions spanning Allen Institute and Janelia mouse brain datasets. Each join checks that both sides use compatible atlas and taxonomy versions, and each answer carries a record that a verifier reruns. A pre-registered experiment measures what each part adds, in Claude Science and in other agent harnesses.
  2. Addendumv1Addendum: threat modelA STRIDE threat model of mouse-brain-ccf-join: a level-1 data flow diagram with five trust boundaries, and 56 threats, each with its likelihood, impact, mitigations, status and the check that would show the mitigation holds.
  3. Background · 1What an answer about the brain depends onBrain-atlas data is versioned in five independent places. An answer that leaves any of them out cannot be checked, and AI agents leave them out by default.
  4. Background · 2What already exists: MCP servers and assistants over Allen dataA community MCP server wrapped the whole Atlas API. The institute's own MCP work is sustained and in use. Neither pins versions or carries provenance. The institute's AI assistant does not either, and it says so.
  5. Background · 3Same input, different answerThese are documented cases where a changed default, reference release, identifier or platform altered scientific results on identical inputs. Usually the release notes said nothing. Two are computed for this project from Allen's own files.
  6. Background · 4Fluent, confident, unverifiableThese AI deployments in science and medicine failed on what could not be checked: invented citations, numbers with no lineage, accuracy that did not survive new conditions, and novelty measured against nothing in particular.
  7. EvidenceEvery case behind the requirementsEvery verified case on one timeline, traced to the requirements and entities it motivates, and in one table.
  8. Data modelThe data model, full screenThe entity explorer: every entity with its fields, relations and the requirements it carries.