Synaptic Protein Zoo

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.

The reference combines expert-accepted volunteer consensus categories and geometry. Individual volunteer scores use paired green/magenta responses against that consensus, so they do not provide an independent expert comparison. A separate majority baseline is not defined.

Each paired volunteer (a user who classified both the green and magenta channel of a subject) is scored with the campaign's end-to-end channel scorer directly against the frozen benchmark target. A volunteer's item score is the mean of their green and magenta channel scores. A subject's baseline is the mean over its paired volunteers, 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 — Synaptic Protein Zoo

## Method

Direct target scoring (`direct_target`) is computed by the campaign's declared
generator, `zoobench.baseline`, which emits `baselines.json` with a fingerprint
of the exact cohort, references, volunteer responses, and scoring configuration.

A paired volunteer is a user who classified both the green and magenta channels
of the same subject. Each channel response is scored directly against that
channel's frozen category-and-geometry target with the same end-to-end scorer
used for model answers. The paired item score is the mean of the two channel
scores.

## What counts as one item

One item is one paired volunteer on one subject. Paired-item scores are averaged
within each of the 100 subjects and then across subjects.

The final cohort is 100 subjects sampled with seed 42 from Chloé Morin's 108
accepted examples. Expert screening validates the retained references but does
not make their volunteer-derived coordinates independent of the volunteers.
All reported scores use 10-pixel tolerances; 15 pixels was only the historical
review-pool eligibility threshold. Exact within-cohort permutation is also
computed. A separate majority baseline is not defined for these aggregate
category-and-geometry targets.

## Caveats

- The frozen category and geometry targets were aggregated from these same
  volunteer classifications. Direct scoring therefore includes a volunteer's
  category and marks in the target they are scored against and can inflate both
  categorical and geometric agreement.
- Only volunteers who classified both channels form paired items. Every subject
  must contain at least one such volunteer; subjects are never silently dropped.
- The channel scorer gates on category first: a wrong category scores zero for
  that channel regardless of geometry, exactly as for model answers.
- Geometry credit inherits the frozen tolerances in the manifest scoring block.