AI-agent research project · started 2026-08-14
The goal: find what actually causes vitiligo and a real cure for it. A 20-persona AI research council reads the published literature, generates and cross-challenges candidate hypotheses, and an executable cell-and-immune-system model runs simulated clinical-trial-style experiments on modeled skin cells to test what each hypothesis would actually predict — before any idea would ever reach a real patient. This site is the honest, browsable record of that search: what's been tried, what's been ruled out, and what's still open.
candidate status, a research direction worth a real scientist testing, not a
finding, treatment, or diagnosis. Nothing on this site should inform a treatment decision
without a qualified clinician.
Six figures rebuilt from the project's own record every time the site publishes — the completed model tests, where each hypothesis stands against a real human study, and the latest simulated combination-therapy run. Open the full findings page →
Every hypothesis the council has produced, plotted as one point each — bigger and brighter the further along it is in the real world (already in human trials), smaller and dimmer for an open gap or one parked after review. Size and color tell the same real story on purpose, not a second invented "how promising" score. A ring around a node marks it as this project's own original discovery — a connection it found itself, not repeated from someone else's published work; most nodes have no ring, and that's accurate, not a shortfall. Drag to rotate, scroll or pinch to zoom, click a point for the real story behind it.
An interactive, clickable melanocyte — rotate it, and see which real structures (MHC-I, melanosomes, mitochondria…) are actually implicated in the disease and why.
🩹Zoom out to a patch of skin. Trigger a Koebner event (friction, sunburn) and watch whether it fizzles out or catches, driven by the same model as the 3D cell and console views.
🧬16 real candidate risk genes on an interactive DNA model, plus an honest inheritance calculator — vitiligo is polygenic, not caused by one gene.
☀️Does UV help depigmented skin, how much, and how long? An interactive dose-response model — genuinely test-run, not hand-drawn — next to the real published clinical evidence.
🧪Every candidate hypothesis the council produced, with real pass/fail results from the executable model where one exists.
🧑🔬Meet the 20-persona council and read a summary of every real discussion round — what was proposed, what survived, and what didn't.
🗺️Five prioritized next steps — drafted, then adversarially roasted by this project's own reviewer before publishing. A living plan, not a fixed one.
📈Every roadmap cycle this project has run, past and current — a history of completed and in-progress "sprints," not just whatever's active right now.
A dated, honest log of what was actually found, in the order it happened — not a status report. Every entry links to the real page or section it's about.
A short, honest account of the process — not marketing copy.
20 independent AI scientist personas (immunology, cell biology, brain-skin-mind axis, pharmacology, systems/translational science) generated candidate ideas independently, grounded in real citations, then a devil's-advocate reviewer cross-examined every claim across each round. Survivors were written up as candidate hypotheses. Separately, an executable rule-based model was built test-driven — one test per hypothesis, referencing the exact source it encodes — so claims about what a mechanism would predict could actually be run, not just asserted. The model is explicit that its numbers are illustrative, not calibrated measurements; it's a consistency check on the project's own claims, not a biology simulator.
Not a claim that this project is more likely to find a cure than the real scientists working on one — a plain account of what's genuinely unusual about how it works, so that claim can be checked, not just taken on trust.
Most hypothesis tracking is a written claim. Here, each one that reaches the executable model gets its own automated test, referencing the exact source it encodes — the test count on this page is live, and every one of them re-runs before anything ships.
🧑⚖️Not one-time peer review — a devil's-advocate pass runs on every cycle of work, and what it found wrong, and how that got fixed, is published right alongside the result. Most research doesn't show you that part.
◯Most of this project is curating real published science, which is useful but isn't discovery. The globe above marks the small number of genuinely original connections with a ring on the map — every other hypothesis honestly has none, on purpose.
🔎Instead of starting from a paper, one pass asked the model's own structure where the biggest opportunity was — then checked that answer against real literature. It pointed at an already-approved drug class current vitiligo drug development has largely missed.
What this doesn't mean: none of this makes any hypothesis on this site more likely to be true. Methodology transparency is not evidence — every claim here still needs the same thing any published science needs: a qualified human verifying it independently. That bar never moves, no matter how the work was produced.