Background · Chapter 1October 13, 20265 min read

What an answer about the brain depends on

The Allen Institute's brain atlases are among the most used reference data in neuroscience. Every answer drawn from them rests on five things that change on their own schedules.

In short · background

One question about cell types in a brain region rests on a data release, a taxonomy, an atlas annotation, a gene annotation and a set of mapping parameters. Each has changed in ways that moved results. An agent that does not say which ones it used gives answers nobody can check.

Atlas voxels relabelled, 2016 → 2017
27.8%
at 100 µm; computed for this project
Gene symbols out of date
2,400
of 32,285, against Ensembl 116
Cells in one spatial dataset
4.0M
from one mouse brain
Clusters in one taxonomy
5,322
in four levels

The Allen Institute publishes reference data that much of neuroscience builds on. That includes a three-dimensional atlas of the mouse brain, cell-type taxonomies built from millions of single cells, and spatial maps of where those cells sit. All of it is free to read from public cloud buckets. The Allen Brain Cell Atlas alone describes its whole-mouse-brain taxonomy as 5,322 clusters, nested in 34 classes, 338 subclasses and 1,201 supertypes (ABC Atlas docs).

Ask a question of that data, for example "which cell types are in primary visual cortex?", and the answer depends on more than the data. It depends on five things, and each of them is versioned separately.

Five things that change

Fig. 1One question, five versioned inputs. Every consequence shown comes from a primary source or from an analysis of Allen's published files.
  1. The data release. The atlas is distributed as dated manifests, for example releases/20241130 and releases/20250531 in the public bucket. Each manifest lists components, and each component has its own version date. In one spatial dataset, the expression matrices date from 2023-08-30 and the cell metadata from 2024-11-15 (ABC Atlas docs).
  2. The taxonomy. Cell types are nodes in a tree, at a level, in a named release. A newer consensus taxonomy now sits beside the one above (brain-map.org). "Microglia" means something different at each level and in each tree.
  3. The atlas annotation. The labels that say which voxel belongs to which brain structure have been released in 2015, 2016, 2017 and 2022 versions. An analysis this project ran on Allen's own 100 µm volumes finds that 27.8% of voxels changed label between 2016 and 2017. Allen's Python SDK switched its default to the 2017 labels without a changelog line (chapter 3).
  4. The gene annotation. Genes are joined by symbol or by stable id, against a named release. The atlas's gene table still uses pre-2020 names like Sept4. Checked against Ensembl release 116, 2,400 of its 32,285 symbols differ, and 15 old symbols now name a different gene.
  5. The mapping settings. Cells get their types from an algorithm with parameters. When Allen's MapMyCells changed one default, the same file got different confidence scores. The labels barely moved, but anything filtered by confidence changed (forum thread).

None of these changes was a mistake. Atlases get better, taxonomies are revised, genes get renamed for good reasons, and the new default is better calibrated than the old one. The problem is answers that do not say which version they used. You cannot check those answers, and you cannot compare them with each other.

Why AI agents make it harder

A scientist working by hand usually knows which atlas version their notebook loaded, because they loaded it. An AI agent answering in a chat does not show that work. Several things push it toward answers that cannot be checked:

  • It answers from memory. A model's training text mixes many releases of the same names. Nothing in a fluent answer says which release it came from.
  • The data does not fit in its context. One spatial dataset is 4.0 million cells, with 14.2 GB of expression matrices (ABC Atlas docs). An agent cannot read it. It has to call tools or write code, and each step can quietly pick up a default.
  • The traps are invisible to it. That spatial dataset comes from one mouse brain on a 500-gene panel, and a second, imputed version covers genes outside the panel. Millions of cells from one animal are one sample. A count of an imputed gene is not a measurement. An expert spots both at once. An agent states them only if something makes it.
  • Tools often return data without its version. A tool that answers "structure 385 is VISp" without naming the annotation hands the agent a fact with no date on it.

Anthropic's engineering guidance points the same way. It recommends that tools return compact, meaningful results and serve workflows rather than endpoints (Writing effective tools for agents). It also recommends keeping intermediate results out of the model's context by letting agents filter data in code (Code execution with MCP).

What exists today

The institute already runs an AI assistant inside its cell atlas, and it documents what that assistant does not do. One of those things is "provide detailed provenance or methodology explanations for the data displayed" (technical documentation). Its neural-dynamics team runs its own MCP server over experiment metadata. A community MCP server that wraps the older Atlas API was published in early 2026 and drew little public use. Chapter 2 covers all three. None of them resolves cell types, brain structures and genes against pinned releases, with a provenance record on every answer.

What the series argues

  • Chapter 2 shows that existing MCP servers lack version-pinned identity and provenance.
  • Chapter 3 collects documented cases where a small change to a default, a reference, an identifier or a platform changed results on identical inputs.
  • Chapter 4 collects AI deployments in science and medicine that failed on what could not be checked: invented citations, numbers with no lineage, and accuracy that did not survive new conditions.
  • The design review turns the evidence into requirements and a design that joins mouse brain datasets at compatible versions, with a record on every answer, and specifies the experiment that decides which parts of it stay.