Software-defined democracy.
A state you can read. Every rule it runs on written down, every change to those rules recorded, every outcome checked against what was promised. None of that is a new idea in government. It has just never been cheap before, and now it is. India is where we test it; the design assumes nothing about India.
Written down, and diffable
A rule that lives only in a file note is a rule nobody can check. Policy, the reasoning behind it and the decisions it produces sit in one versioned record. You can pull up what changed, when, and who argued for it. The history is the accountability. Nothing else has to be taken on trust.
Every policy ships with a kill switch
Before it goes live, a policy names the number it is trying to move, the window it has to move it in, and what happens if it doesn't. Most government programmes name none of the three. Missing a target should be ordinary and survivable, not a thing to be denied for a decade.
People who have done the job
Air safety rules written by people who have investigated crashes. Hospital procurement designed by someone who has run a hospital. It is not a demanding standard, and almost nothing meets it, because the qualification for office is winning an election and the qualification for the work is having done the work.
No single point of failure
Decisions belong at the level that holds the information: mostly local, sometimes national, rarely both. A district that gets something wrong should be able to fix it without waiting on a capital. A capital that gets something wrong should not be able to break every district at once.
A democracy that learns
Reform usually arrives as one large launch and is never measured again. These are the parts that make a second version possible.
Live instrumentation
Where the money went, how far the work has got, and whether the number the policy promised to move has moved. Published as it happens, rather than compiled into an annual report that lands eleven months late and is read by nobody.
A pipeline before deploy
Nothing reaches everyone first. It gets read by people who know the domain, argued over in the open, then run somewhere small enough that being wrong is cheap. It scales only if the pilot says it should.
Models that show their work
Forecasting is not the hard part. It is the part that gets oversold. A model earns its place here by making its assumptions legible enough to argue with, and when it disagrees with someone who has done the job, that disagreement is the useful output.
None of this exists yet.
It is a design, and designs are cheap. What it needs now is people willing to specify one piece of it properly and then defend the specification.
Join the lab →