Questions we keep returning to

String theory needs the Ricci-flat metric on a Calabi–Yau manifold, and there is no closed form for it. Neural networks can approximate one metric for one shape — but the landscape holds a vast number of shapes. This note asks whether a network trained on a neighbourhood of similar shapes (a "village") transfers to shapes it has never seen.

• Inside a village, one epoch of fine-tuning from the village specialist matches what training from scratch reaches after roughly 185 epochs — and at convergence the warm-started network is still 1.4× more accurate. Scratch does not catch up.
• The village edge is a slope, not a cliff: accuracy degrades smoothly with distance, and warm-start beats an equal-budget scratch run at every radius tested.
• One network for the whole space does not work at this scale. A router plus a set of specialists is the working generalist — so covering the landscape is a tiling problem, not a modelling problem.

Replicate on a second family (a quintic or a CICY), locate the village boundary, then push downstream to Yukawa couplings on a three-generation quotient — where the metric’s accuracy actually cashes out.

Most wellbeing apps nudge on a schedule. Animo asks a harder question: when does a nudge actually help, and when is it just noise? It frames the intervention as a cost-penalised contextual bandit — a system that learns, per person and per moment, whether to say something at all.

• Built as a just-in-time adaptive intervention (JITAI): the decision is not what to say but whether this is the moment to say it.
• Every interruption carries a cost. The model is penalised for nudging, so it has to earn each one with a measurable, proximal benefit inside the app rather than a distant promise about screentime.
• Android is the research platform — iOS keeps Screen Time data away from third-party apps — and ethics review comes before any data collection.

Finish the intervention design, clear the ethics board, and run a small in-the-wild study. Animo is treated as research first; whether it becomes a product is a question for after the data.