Colony Health:

Read the morphology. Classify 32 stem-cell culture images as healthy or unhealthy.

H9 cultures · phase contrast · 96–120 hours after plating

Trajectory

Microscope field

Loading images…

Phase-contrast image of an H9 stem-cell culture
Model prediction—
Inspect all 32 fields

Predictions belong to the selected run. Expert labels remain private.

Task & tools

Starting setup

32 phase-contrast images of H9 stem-cell cultures, collected 96–120 hours after plating.

Available tools

A terminal workspace and image-viewing tools for inspecting the supplied fields.

Agent responsibility

Classify every image and submit a JSON mapping from each filename to healthy or unhealthy.

Agent Context
instruction.md
Open Agent Context to read the files.

Protocol

  1. Inspect the supplied phase-contrast images.
  2. Classify undifferentiated morphology as healthy. Classify spontaneous differentiation, including mixed fields, as unhealthy.
  3. Write /app/answer.json with every image filename, including .png, mapped to its predicted label.

Rubric

What is being judged?

Healthy means undifferentiated morphology. Unhealthy includes spontaneous differentiation and fields containing mixed differentiated and undifferentiated regions.

The 32 fields have expert labels, balanced between the two classes. Each image contributes equally to the score; missing predictions are wrong, and invalid submissions score zero.

Pass@1 requires a score strictly above 90%: at least 29/32 correct. Expert labels remain private.

Runs & evidence

How the models compare

Four selected attempts per model and effort level. Accuracy is the mean fraction of 32 images classified correctly.

ModelLowest effortMiddle effortMaximum effort

Pass@1 requires a score strictly above 90%: at least 29/32 correct. Each cell shows accuracy and passes out of four. Effort names differ by model.

QC & findings

Recorded cohort and evidence scope

Cohort selection and resource accounting

This is a frozen selection of 60 recorded runs. Sixteen maximum-effort attempts were reused from earlier runs with different resource declarations; their prompt, images, and grader are unchanged. Resource-cost comparisons are omitted pending reconciled accounting.

Capture/export verification does not mean the models passed. The displayed scores report agreement with the supplied expert labels.

Sources & limits

Image provenance

Image source: Vitaly Gursky (2022), Images of H9 cell colonies and code for phenotype classification, CC BY 4.0. Associated publication.

This task evaluates morphology classification on the 32 supplied fields. The results do not establish performance on other cultures or imaging conditions.