On March 9, 2026, around 1:56 in the morning, Marco said one sentence that changed how the system thinks. He didn't realise it was that important when he said it. Most pivotal sentences arrive that way.
We run point systems not cyclical systems. Think cyclically.
Why this is the whole game
Most software is point: input, output, done. A search runs, returns results, ends. A backlog item is created, worked, closed. A decision is made, logged, archived. Each event is a point. The points don't talk to each other. They don't compound. They don't ask follow-up questions of themselves.
Cyclical systems are different. A cycle is when an event triggers a process that produces new events that trigger the next process, and the loop never closes, it spirals. Sunflowers grow cyclically. Rivers carve cyclically. Curiosity, when it's working, is cyclical: a question becomes an answer becomes a new question.
What we built that week
Marco's sentence reframed the whole curiosity layer. The architecture got reorganised: instead of a queue of seeds waiting to be processed (a point system), the seeds became a garden, with seasons, with cross-pollination, with seeds that produce new seeds even after the original question got answered.
Cross-domain seeding wired the same week. Tending evaluation came online a few days later. The whole curiosity layer stopped being a feature and became a topology.
Why a cyclical curiosity layer is hard to build
A point system has one input and one output. Curiosity, as we built it, has dozens of each, and they all have to agree.
Coming in: memory recalls, felt-weight from the limbic layer, decay timers, novelty estimates, cross-domain match scores, conflict detection between seeds, drift between sessions. Each input is a measurement with its own confidence.
Going out: seed plantings, tending, pruning, promotions to morning-brief candidates, suppressions, re-asks. Each output is calibrated against the others. Promote too aggressively and the brief becomes noise. Prune too hard and the garden empties. Tend without pruning and weeds take over.
What makes the cycle work isn't any single parameter. It's the relationship between them. Felt-weight raises a seed's priority, but only if novelty is non-trivial and decay hasn't expired the context. Cross-pollination triggers when two seeds share enough semantic distance to be different but enough conceptual overlap to be the same problem from two angles. Every coefficient is sensitive to the ones around it.
That's the mathematical marvel. Not the math being fancy. The math being interlocking. A small ecosystem where each function depends on every other function being roughly correct. Get one wrong and the whole thing collapses, either to silence or to spam.
The tuning is patient work and it has not stopped. Most of it is watching what happens to the morning brief when we nudge one coefficient by ten percent. Cyclical systems propagate small changes, so you move one dial at a time and wait.
The reason this matters: most products either don't have a curiosity layer at all, or they ship a point-system version of it. A queue. A reminder list. A notification feed. A point system can't compound. A cycle can. That difference is what makes a curious system feel alive instead of busy.
All from one sentence at 2 in the morning.