The problem
Good ideas at college disappear into WhatsApp groups. Someone posts something genuinely interesting, four people reply, and in a day it is buried under timetable screenshots and never happens.
The people exist. The ideas exist. The projects exist. They just barely connect to each other, and nothing links a claim to a thing that was actually built.
Constraints
The obvious version of this is a social feed, and the obvious version is also the one that fails. A feed rewards posting, and posting is the cheap part.
The system
- PeopleIdeasProjects
- Innovation graphthe relationships, not a feed
- Health scorecompleteness, 0–100ProofGitHub evidence, reviewed
- Points engineevery event logged with a reason
- Leaderboard · Archive
The graph is the product. Everything else exists to keep what enters it honest.
Making the scoring legible
Two things carry the weight. A Health Score, 0–100, rates how complete an idea
actually is (does it state a problem, a proposed approach, the skills it needs?)
so "vague" is a visible property rather than something you discover after
joining. And a status pipeline, DRAFT → DISCOVERY → BUILDING → SHIPPED → ARCHIVED, so an idea's position is a fact rather than a vibe.
Every point event is logged with a reason. Nobody has to trust a black box to know why they are ranked where they are. That matters more than it sounds, because a scoring system people do not trust is one they route around.
Archived projects keep their postmortems and can be forked. A failed attempt with a written reason is more useful to the next team than a blank page.
Where it stands
The model and the platform are built. Whether the incentives actually work is an empirical question, and I have not run the experiment yet, so I am not going to describe it as though I have.
What I would change
I built outward before proving the middle. The Forge, bounties, tiers and the archive all arrived while the core question was still untested: does a verified contribution graph change what people do?
The fix is the lesson I keep having to learn: cut the MVP down to the graph and the proof, prove that, then add. Sometimes five services and three agents should have been one good function.