Skip to content

Add correlation and voxel members, and count passive conditions as task states - #58

Merged
jdkent merged 2 commits into
mainfrom
claude/enum-free-text-audit-83b601
Nov 5, 2026
Merged

jdkent merged 2 commits into
mainfrom
claude/enum-free-text-audit-83b601

Conversation

@jdkent

@jdkent jdkent commented Nov 5, 2026

Copy link
Copy Markdown
Member

From the open-enum free-text audit of the 27B astra-distill records (64,589 papers, set-20261101). The audit dashboard is a private claude.ai artifact; ask James for the link.

Changes

  • ConditionKind.task_state now covers active and passive states: "performing a trial and responding, or viewing, listening to or receiving stimuli with nothing to do". It said "an active state" before, so the model refused it and wrote "passive viewing / listening / stimulation" 6,665 times in 2,538 papers.
  • ModelFamily.correlation (new): a correlation between the measure and a variable that is not another brain region's signal. That covers brain-behaviour correlation, including one on connectivity strength. The model wrote it as free text 10,543 times in 4,624 papers.
    • Where it stops: a seed-based or ROI-to-ROI correlation is a connectivity analysis, and its model family is the model that tested the effect on the connectivity estimates. The seed_based definition, AnalysisType.mass_univariate, MassUnivariateDetails and Analysis.details now say this.
    • Overlap measured: of the 9,880 analyses with a correlation-worded family, 5,749 are mass-univariate (5,183 cross-subject regressions) and 3,124 are connectivity analyses. About 1,100 of the connectivity ones were the time-series correlation itself.
  • RegionType.voxel (new): a single voxel or coordinate with no surrounding volume. "peak voxel", "voxel" and "seed voxel" were written 4,535 times in 1,532 papers. A sphere or box stays coordinate_sphere, which takes the 840 cube and box values.
  • schema-tutorial.md: the region_type and model_family member lists are updated.

Generated files

The extraction tree is regenerated with python3 -m pondie.schema.generate (with PONDIE_SCHEMA_DIR pointing at this checkout), and the models and JSON Schemas with tools/generate_models.py. The parse schemas pick up ModelFamily through storage.

Checks

  • tools/generate_models.py --check passes.
  • pytest: 232 passed, 16 skipped (pyarty and setuptools are not installed locally).
  • pondie's storage_parameter_priorities, field_provenance and extraction_to_storage_map checks pass.

🤖 Generated with Claude Code

…sk states

An audit of the 27B astra-distill records (64,589 papers, set-20261101) counted what the
model wrote as free text in each open enum. Three of the findings change the vocabulary:

- ConditionKind.task_state said "an active state", so the model refused it for passive
  conditions and wrote "passive viewing", "passive listening" and "passive stimulation"
  6,665 times in 2,538 papers. task_state now covers active and passive states alike.
- ModelFamily had no member for a correlation; it was written as free text 10,543 times in
  4,624 papers. `correlation` is a correlation between the measure and a variable that is
  not another region's signal (a score, a rating, age), so a brain-behaviour correlation is
  this family, including one on connectivity strength. A seed-based or ROI-to-ROI
  correlation is a connectivity analysis, whose model family is the model that tested the
  effect on the connectivity estimates. seed_based now says so, and the lists of
  mass-univariate families name brain-behaviour correlation and send region-to-region
  correlation to ConnectivityDetails. Of the 3,124 connectivity analyses with a
  correlation-worded family, about 1,100 were the time-series correlation itself.
- RegionType had no member for a region that is one voxel; "peak voxel", "voxel" and
  "seed voxel" were written 4,535 times in 1,532 papers. `voxel` is a single voxel or
  coordinate with no surrounding volume; a sphere or box stays coordinate_sphere.

Extraction tree regenerated with pondie.schema.generate; models and JSON Schemas with
tools/generate_models.py. The parse schemas pick up ModelFamily through storage.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

@jdkent jdkent left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

One change I wanted for wording

@jdkent
jdkent merged commit 93cfc88 into main Nov 5, 2026
1 of 5 checks passed
jdkent added a commit that referenced this pull request Nov 5, 2026
Regenerate after #58's task_state description edit
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant