Better tools for people who find emotion hard to read.

I started this project because reading emotion is harder than the tools that claim to help with it usually admit — and the people most affected by that gap are often the ones already finding emotion hard to read in the first place. Autistic learners. Anyone for whom social cues don't land as cleanly as the textbooks describe. The original intent was simple: build something that helps.

I wasn't sure yet what that something would be. A wearable, an app, a classroom tool. The shape would come later. The point was the people.

Before building anything, I wanted to read what already existed. I spent most of a year on a systematic literature review of AI tools designed to help autistic users build emotional intelligence. The result was published in 2025 in the Curieux Academic Journal.

Four tools across very different designs — a wearable, a classroom platform, a recognition engine, a social robot — kept running into the same four problems. The annotation bottleneck. Weak transfer to the real world. Overdependence on AI feedback. Unresolved ethics around emotion data. The recommendation: AI should complement human-mediated learning, not replace it.

The obvious next step was a fifth tool. I started sketching: a wearable bracelet that would pick up physiological signals and translate them into emotional state — I called it Emo-Sense in my notes. A classroom app for structured exercises. A conversational AI tutor.

I wasn't pretending to be done with the paper's question yet. I was trying to take the next swing at it. The sketches got more elaborate. I started writing them up like product specs.

A few weeks into the spec, I noticed something I couldn't unsee. Every sketch I drew was going to inherit the exact same data substrate my own review had spent thirty pages critiquing. The annotation bottleneck. The transferability gap. The single-label assumption. None of those go away when you add a new interface on top — they get inherited.

I'd been about to build the fifth tool on the same broken foundation as the previous four.

So the project changed shape. Not abandoned — restructured. Instead of building a single tool right now, I committed to a multi-year project that begins by building the substrate the field is missing. Plural human readings of the same moments, captured carefully enough that the tools that eventually come out of this work can rest on something solid.

That's what MindLens Lab is. The wearable can come later, when there's a foundation under it. For now: the data.

In late 2025 I gave the project a name and a structure. The pre-registration went up. The site came online. The first clips went through consent and curation. The invitation to read them went out.

Everything that came next depended on strangers on the internet — and, eventually, adult readers in seven countries — actually taking the ten minutes.

Emotion reading is plural, not private.

Phase 1 · Part 1 closed on 14 August 2026 having demonstrated the load-bearing claim from the originating paper: adult readers of the same short social clips converge on one answer only about 70% of the time. The disagreement isn't noise — it's structural.

442 adult readings across 33 short social clips, from 7 countries. Two preprints came out of it: one on how humans read those moments plurally, one on how a single contemporary AI reads them by comparison. The human-side paper was accepted at Curieux Academic Journal.

The dataset is frozen on Zenodo under a permissive licence. Anyone can read it, re-analyse it, or build on it. That was the point of Phase 1 — not the papers, but the substrate underneath them.

Phase 2 · Multi-AI Calibration is where we are now. The Phase 1 preprint left one question unresolved:

Was that pattern — agreement where humans agree, hedges inside the human plurality — specific to one AI, or how frontier vision-language models behave in general?

Phase 2 answers that concretely: four models, five prompt conditions, the same 32 clips, three runs per cell, all scored against the plural human distributions Phase 1 produced.

The plan is pre-registered before analysis. The runs are visible live on the home page as they complete.

This is a multi-year project. The concrete milestones — papers, timelines, and thresholds — live on the progress page. Here I only want to name what stays the same, even as everything else moves.

The arc bends back to the original question: helping the people the project was designed to serve. Just by a longer, more honest route than I first imagined.