Why we're testing four AIs against the same people they're supposed to read.

The Phase 1 human study (Kim, 2026a, accepted at Curieux) found that adult readers of short social clips are more plural than the one-emotion-per-clip evaluation frame assumes. The companion AI study (Kim, 2026b) scored Gemini 2.5 Pro against that plural distribution rather than against a canonical “correct answer.”

That is a striking result — but it's one model. What Phase 1 couldn't answer is whether this calibration is specific to Gemini or a general property of frontier vision-language models. If different labs and open- vs closed-weight approaches all track the plurality similarly, that changes what we can say about “where AI is” on this task. If they diverge sharply, model choice becomes the story.

Four vision-language models chosen to span major labs and both proprietary and open-weight approaches:

Each model reads the same 32 short clips from the Phase 1 corpus, under five prompt conditions that vary how much scaffolding it gets — from a bare “What is this person feeling?” to an explicitly plural “Return a probability distribution over these fifteen emotions.” Every cell is run three times so we can separate real signal from within-cell variance. Responses are scored per clip against the plural human distribution; the primary metric is Jensen–Shannon divergence (lower is closer), modal agreement secondary.

Every choice — which models count, which prompt conditions are “fair”, which metric is primary, what counts as a hypothesis being supported — is fixed before the first API call. That commitment lives on the Phase 2 pre-registration, which was posted on 26 September 2026 and is time-stamped on Zenodo. Once the runs finish (target: 31 December 2026), the Phase 2 findings note will report against those hypotheses in the order they're listed. Nothing gets a hidden edit.

Phase 2 runs entirely on the frozen Phase 1 human dataset (already released on Zenodo under CC BY 4.0) plus AI API outputs. No new human subjects are recruited, contacted, or observed. The dataset's original ethics posture (documented in the Phase 1 preprints' Ethics Statement) governs all human data used here; no additional IRB review is required for the AI runs themselves.

A future Phase 3 is planned that would introduce small-N human validation of the best-calibrated Phase 2 configuration. That work will require its own consent framework and ethics review and is not part of this pre-registration.