Etch A Cell: Going Nuclear

All 100 current campaign subjects are shown. GPT-6 Astra received the campaign tutorial with low reasoning and the optional field guide off. These are direct API responses from one run on 2026-10-01.

Each individual volunteer drawing's raw line fragments are stitched into contours with the published CRIA endpoint-linking procedure, then scored directly against the frozen benchmark boundary using 5 px tolerance-boundary F1. A subject's baseline is the mean over its volunteer items, and the campaign baseline is the subject-balanced mean of the per-subject means, matching how the model's campaign score averages one score per subject.

# Volunteer baseline methodology — Going Nuclear

For `expert-c001-c006-five-slice-v1`, both volunteers and models are evaluated
against independent expert boundaries for the middle image of each subject.
The cohort is 50 C001 slices and 50 C006 slices.

## Individual volunteers

One item is one classification row's complete annotation on zero-based frame 2,
including all its nonempty line fragments. The existing CRIA endpoint-linking
procedure greedily stitches these fragments into complete contours. Model
answers already contain complete ordered contours and need no stitching.
Both are rendered as one-pixel OpenCV `LINE_8` boundaries on the 1000 × 1000
canvas and scored with five-pixel Euclidean tolerance boundary F1.

The expert reference comes from the matching 2000 × 2000 TIFF boundary page:
source z minus one, area resize, threshold at 128, and Zhang skeletonization.
Expert references are one-pixel skeletons, not OpenCV-rendered volunteer contours.

Scores are averaged within each subject and then equally across all 100
subjects. There are 2,755 included annotations. The production result is
`0.705745887348199` (70.5746%). Empty central-frame responses are excluded by
the existing drawing-baseline definition; this is a baseline for nonempty
annotations, not every submitted classification.

## Volunteer consensus

CRIA aggregation is unchanged and retained separately for comparison against
the expert labels. Its mean subject-level F1 is `0.9476671855382307` (94.7667%).
This aggregate is not the model's reference label. Unlike the former
volunteer-reference cohort, the current individual-volunteer score does not
include that volunteer's trace in the reference used to grade it.

## Interpretation

Boundary F1 with five-pixel tolerance is the sole score, shared with Correct
Mito. Exact boundary F1 and Hausdorff diagnostics have been removed. The
configured tolerance allows small tracing offsets. Empty or fully out-of-canvas
predictions score zero. The September 30, 2026 scorer version preserves the
existing five-pixel F1 values. This dataset represents two cells rather than 100 independent
cells, so slice-level variation does not measure between-cell generalization.

`baselines.json` contains the canonical individual-volunteer, volunteer-consensus
(`majority_volunteer`), and exact permutation baselines. `reference_validation.json`
independently records the consensus-expert comparison and migration parity checks;
the validator checks every consensus score against that record. Rebuild both through the preparer;
`python -m zoobench.baseline --campaign campaigns/goingnuclear` can also recompute
the canonical baseline artifact alone.