The shift
For most of its history, science was bottlenecked by production. Experiments were slow, data was scarce, a good hypothesis was hard to come by. AI has loosened all three. A model drafts a protocol in minutes, reads more literature than any lab, and proposes more candidate experiments in a month than a field can run in a year. Autonomous labs are starting to run the bench.
Production was never the part that made science cumulative. A field compounds through the slower work that follows: selecting what survives scrutiny, remembering what failed so no one repeats it, and carrying corrections back to the claims they overturn. That half has not gotten cheaper. So the rate-limiting step has moved. The hard question is no longer how to generate the next idea. It is how to decide what becomes inherited knowledge, and how to get a correction to the work that depends on it before the next decision is made.
The record
Today that work happens in prose. A paper carries a claim and its author’s reasoning, not the change to what a field believes or what depends on that change. Corrections travel by rumor. Failed runs stay in a thread. An agent’s synthesis disappears when the conversation ends, and the next worker rebuilds the map by hand.
The missing layer sits beneath the paper: a shared, writable record of frontier state. What is known, what failed, what changed, what depends on what, and what would move the field next. Software already has this. Git gave code a memory, and AI writes code at scale because the work lives in objects an agent can inherit, test, and merge. Science has no equivalent for changes to what it believes, so confidence travels farther than correction.
Whoever owns that record shapes what later intelligence treats as true. If it defaults to closed silos with AI wrappers on top, corrections will need commercial cooperation to travel, and a field’s memory becomes a private asset. The record has to be open and forkable from the first deposit, held by a foundation kept separate from any company that runs the dominant client. Open code with a closed registry is captured infrastructure with a permissive license file.
There is a defensive reason to build it and a generative one. Without a shared record, AI accelerates the failure to inherit: more output, narrower attention, faster monoculture. With one, a field gains the corpus a model learns from. AlphaFold followed the PDB; the reason there are not a hundred more like it is that there are not a hundred more PDBs. A shared record is how a field manufactures one, and where an answer is cheap to check, capability compounds as fast as the record accretes.
The wager
The claim is testable on something small. One bounded frontier, blood-brain-barrier dysfunction in early Alzheimer’s. One chartered registry. A few labs. A weekly review window. Grant money tied to signed deposits of failed protocols and corrections. It passes only if one accepted correction changes a decision a funder, reviewer, lab, or regulator would otherwise have repeated. If that loop closes once, the rest is engineering.
Vela is the protocol that makes scientific state writable. Canopus is the open foundation that keeps it public. The terafactory is the physical body that reads from the record and writes back to it.
The three essays make the case at length. Constellations of Borrowed Light is the moral and epistemic argument for a shared record. The Discovery Engine is the operating loop that turns activity into governed state. The Terafactory Age is the physical body that acts on it. The whitepaper is the protocol itself.
