Phase 1 · Complete (Apr – Sep 2026)
The first 500 readings produced two papers, both openly available:
Phase 1 · Part 1 collection closed on 14 Aug 2026; the human-side paper was accepted at Curieux Academic Journal. Phase 2 · Multi-AI Calibration is now underway. A Phase 1 · Part 2 wave is deferred — the definitive Phase 1 paper that would draw on it is not currently scheduled.
Section 1 · Data accumulation
The first target of 500 responses was cleared, and both Paper 1 and Paper 2 are now deposited on Zenodo (2 Sep 2026). Weekly totals below, cumulative, as the dataset keeps growing toward the definitive Phase 1 paper.
Cumulative readings
Phase 1 · Complete · 500+ responses (Apr–Sep 2026)
Two preprints published on Zenodo (2 Sep 2026)
P1 accepted · Curieux Academic Journal (Sep 2026)
Phase 2 · launches 1 Oct 2026
504 / 500(100%)
Phase 1 · Complete · 500+ responses (Apr–Sep 2026)
Two preprints published on Zenodo (2 Sep 2026)
P1 accepted · Curieux Academic Journal (Sep 2026)
Phase 2 · launches 1 Oct 2026
504 / 500readings
100%
Section 2 · Paper readiness
Phase 1 splits its pre-registered hypotheses across a sequence of papers. Each paper waits until its data has real weight. The three below sit in different places on the runway.
Paper 1
Published on Zenodo · September 2, 2026
H1 · Plurality claim
H2 · Verbal dependency ↔ agreement
H3 · Social complexity ↔ variance
43 adult readers, 33 short social clips, 442 responses. Most clips do not draw a single agreed reading, and about one response in four explicitly reports more than one emotion — even at the highest confidence.
Read on Zenodo →Paper 2
Published on Zenodo · September 2, 2026
H5 · AI ↔ human divergence patterns
Gemini 2.5 Pro on the same 33 clips: top pick differs from the human modal on about 40% of clips, and when the AI hedges its listed emotions overlap the human distribution 84% of the time.
Read on Zenodo →Paper 3
Current chapter · running · target close 31 Dec 2026
H1 · Cross-model plurality tracking
collecting
H2 · Prompt-condition sensitivity
collecting
H3 · Modal vs plural comparison
collecting
H5 · Within-cell reliability (N = 3)
collecting
The current chapter — the four-model, five-condition, thirty-two-clip Multi-AI Calibration Study is running now. Live analysis on /phase-2 shows per-cell progress as it accumulates; the frozen pre-registration lives at /phase-2/preregistration. The definitive full-Phase-1 paper (cross-cultural · reader-style · difficulty validity) depends on a Phase 1 · Part 2 wave that is currently deferred.
"Ready" means the data required for that hypothesis is in place; formal test unlocks and effect sizes will appear on the pre-registration page once v0.2 is committed.
Section 3 · Timeline
2025
Originating paper — Curieux Academic Journal
Artificial Intelligence in Emotional Intelligence Training for Autism (Kim, 2025). The paper that named the dataset-first pivot.
See the paper →2026 · summer
Phase 1 · Part 1 · data collection
First-target threshold of 500 responses cleared on 13 Aug 2026. Analytical cutoff 14 Aug 2026; dataset then frozen as Phase 1 · Part 1 · v1.0.
2026 · Sep 2
Preprint 1 published on Zenodo
Paper 1 — Plural Emotion Readings in Short Social Video Clips: A Descriptive Study. DOI 10.5281/zenodo.22251335.
Read on Zenodo →2026 · Sep 2
Preprint 2 published on Zenodo
Paper 2 — Where One AI Diverges from Plural Human Readings on Short Social Video Clips. DOI 10.5281/zenodo.22253369.
Read on Zenodo →2026 · Sep
Phase 1 · Part 2 · deferred
A Part 2 collection wave was outlined (randomized within-wave clip order and gated parental-consent verification), but is deferred while Phase 2 · Multi-AI Calibration is the active chapter. Not currently recruiting.
2027 +
Definitive Phase 1 paper · deferred
Paper 3 (cue mediation · reader style · difficulty validity · cross-cultural) depends on the Phase 1 · Part 2 wave being revived first. Not committed to a date.
Section 4 · Data availability by hypothesis
Factual data-collection state only. Effect sizes and formal test unlocks appear on the pre-registration page once the v0.2 amendment is committed.
Plurality claim
Every live clip has enough readings for a distribution.
Verbal dependency
Requires curator-coded verbal_dependency and enough responses per clip.
Social complexity
Requires curator-coded social_complexity and enough responses per clip.
Cue mediation
Needs deeper per-clip N so cue-partitioned sub-samples become usable.
AI ↔ human divergence
Needs an approved AI annotation per clip, verbal_dependency coding, and per-clip response coverage.
Cross-cultural
Requires substantial per-country participant volume — the tallest ask in Phase 1.
Reader style
Individual-level: participants need at least six answered clips.
Self-rated difficulty
Needs the four-level difficulty scale answered by enough participants; newly moved to the session-summary flow to boost coverage.
Section 5 · Open science
Pre-registered protocol (v0.1)
Hypotheses, thresholds, and analysis plan committed before substantive data collection.
Read the protocol →Phase 1 · Part 1 dataset (v1.0) — Zenodo
The frozen snapshot behind the two 2026 preprints. Cutoff 14 Aug 2026. CC BY 4.0.
Open on Zenodo →Live dataset — Part 2 collection
Every new response joins the running dataset. No behind-the-scenes selection.
Explore the data →Analysis code — public
Reproducible analysis scripts released on GitHub alongside the two preprints (MIT license).
GitHub repository →No conflicts of interest
MindLens Lab is a non-commercial research project. No funders, no product ties.
Section 6 · Research notes
A running log of the small research decisions this project makes as it grows. Kept open so the reasoning is on the record, not just the result.
Sep 2026
After the two 2026 preprints went out, a post-hoc look at the Part 1 response counts turned up a systematic pattern: clips presented later in a wave's fixed order picked up meaningfully fewer responses than clips at the start of the same wave. Across the three collection waves, the within-wave rank correlation between clip position and response count was consistently strong and negative (Spearman ρ = −0.78 to −0.92; p ≤ 0.005 in each wave; ρ = −0.43 pooled, p = 0.012).
Two things follow from that. First, the Part 1 dataset used in the two preprints has a real dropout confound: later-position clips carry less statistical weight than their earlier neighbours, and any comparison between clips on the tail of a wave and clips on the head of a wave is standing on uneven ground. Second, this is a fixable problem, but only for future data — the fair move is to lock the Part 1 dataset exactly as the papers used it and to change the mechanism for the next round.
What we changed: The Part 1 dataset is now version 1.0 as of 14 Aug 2026, published as a Zenodo deposit (doi.org/10.5281/zenodo.22265192). Phase 1 · Part 2, opening in September 2026, keeps the same 33 clips and the same response instrument, but (a) randomizes clip order per session within each wave and (b) gates parental-consent verification at session entry rather than as a post-hoc exclusion. A subsequent definitive Phase 1 paper will draw on the Part 2 corpus.
✓ Phase 1 · Complete (Apr – Sep 2026)
Phase 2 · Multi-AI Calibration is now running against the frozen Phase 1 dataset. Live analysis on the home page; the frozen plan lives on the pre-registration.