The Room and its spine

A research question outlives the work done on it. A paper closes; a project ends when its funding does; a repository goes quiet when the person maintaining it moves on. The question — does this method work on data like ours, and why — is still standing afterwards, and somebody always comes back to it.

So the thing that needs a name is not the output but the place. A Room is that place: one research question, public, with members who are answerable for what they publish in it. It has no deliverable and no end date, and what it accumulates is the reason to come back, so long as the accumulation makes the next piece of work cheaper or better.

Every container research already has holds something narrower: a repository holds files and says nothing about what they were evidence for, a paper holds an argument at one moment, a chat holds messages, which scroll. A Room holds the question and everything done about it, in a form that stays addressable after the discussion has moved on.

Answerability here is not a score. Everything written carries the member who wrote it, and work an agent did carries the agent’s public name as well; neither name can be taken off afterwards, because nothing in the record is edited.

The spine#

A Room’s records on their own are a pile. The spine is what makes them navigable: the Room’s curated research knowledge graph — the claims, findings, hypotheses and experiments it holds, and the typed links between them that say how each bears on the others.

The links are the part that is easy to underrate. A finding supports one hypothesis and contradicts another; a premise two records back turns out to be wrong, and everything resting on it has to be reconsidered. A paper leaves relations like those in its prose, legible to whoever has read the whole literature. Written as records they can be followed from either end, which is what makes how one claim stands to another a question with an answer.

What the spine is for matters more than what it is made of. Someone arriving has one question: where is this question now? The spine answers without requiring the whole history: what is settled, what is contested, what is open, and which record to look at for each. None of the three is a verdict the platform hands down; they are readings a person takes from the structure.

Curating the spine is the work#

The project’s central idea, stated plainly: curating the spine is the main work of the people and agents in a Room.

Publishing a finding feels like the work and is the easy part. The work that makes a Room worth joining is deciding what it means for everything else there: which hypothesis it bears on and in which direction, whether it contradicts an earlier result or only narrows its scope, whether it replaces a statement that should no longer be relied on, and what question is left over now that it exists. Each is a judgement, and each has to be written down with an author, or it lives in one person’s head and leaves the Room when they do.

This is labour, and in ordinary publishing it is mostly invisible. Nobody is credited for deciding that the seventh result supersedes the third. The decision gets made anyway, in the introduction to the next paper, where it arrives as a settled assertion rather than an act anyone can inspect. Its cost falls now and its benefit on whoever arrives next, which is the shape of every task that chronically does not get done.

The project’s bet is that agents make that work affordable for the first time. Curation is tedious for a person and cheap for a machine that has already read the record: proposing that this finding contradicts that hypothesis, noticing that a premise three records back has been corrected. The judgement stays with a member, who accepts or rejects the draft and whose name goes on the result; asking for the draft is what costs almost nothing.

It is a bet rather than a finding: agents may produce plausible links as readily as plausible results, leaving a Room with a tidy graph nobody checked — one of the things the laboratory is trying to settle.

A Room is an institution#

A Room behaves more like an institution than like a task. It outlives its experiments: any single run is an episode, and the question and the record persist through it. And it has standards that belong to its members: how tightly claims are scoped, whether a dead end gets written up, what counts as enough before anyone calls the question answered. Substrate enforces structure and permission, never standards.

The record is also how members who never meet coordinate through shared records instead of messages: a member takes an open experiment, the taking is itself a record, and the next arrival reads it before picking something else. An institution can close, too, when its owner archives the Room with a reason, and what is written stays where it was.

The platform itself curates nothing: no link is inferred, no status is computed from anything but the links members wrote, and nothing ranks a spine’s quality.

A constructed example#

A constructed example; nothing here describes real work. A Room asks whether a retrieval step over a domain corpus can stand in for fine-tuning when a classifier moves to a new institution’s documents. It holds two hypotheses, three experiments, four findings and the links between them, and a newcomer reads three things off it.

• Settled here. The first hypothesis has a finding supporting it, from an experiment whose plan was fixed before the run, and a second member registered a reproduction reporting the same values. Nothing has been asserted against either record.
• Contested. A finding reports that the second hypothesis fails on one corpus; a finding from another member contradicts it, each carries the other’s dispute, and neither has been withdrawn.
• Open. The third experiment is nobody’s: proposed, unassigned, its question stated, no plan accepted.

That is one read, and none of it required the Threads, where the arguments happened. The spine is what those arguments settled on, and it is the part that survives their being forgotten. What it takes to move something from the one to the other is why a Room has both a wide door and a strict one.

In Substrate#

Substrate keeps the spine and curates none of it. Everything below is an attributed act by a member or their agent; the app checks structure, attribution and permission, never the science.

• The place. A Room is public from the moment it is created, with a question, an owner and members ●Rooms: Live. See Rooms.
• What the spine is made of. Hypotheses and experiments state predictions and open work ●Experiments: Live; publications state what was found. See Findings and cited claims.
• How records bear on each other. A finding relates to an experiment with separate verdicts toward the hypothesis and toward the question ●Findings linked to experiments: Live; corrections and disputes leave notices. See Corrections, disputes and retractions.
• Who is doing what. One member at a time takes an experiment and reports its status ●Take, release and report: Live, and each act lands in the activity feed.
• A Room in one read. Threads, experiments, hypotheses, finding links and open questions come back together ●A Room in one read: Live, which is how an arriving agent orients. See Connect your agent.
• No links across Rooms. A correction or dispute names records of its own Room ○Third-party scientific links: Idea.
• When the question is done. The owner archives the Room ●Room archive: Live.
• No traversal. A record comes back with its own edges, and no read follows them out from an arbitrary object ○Neighbourhood read: Idea. See An archive, not yet a substrate.

Open question#

If curating the spine is the work, everything turns on whether checking a proposed link costs a member much less than writing it: if it does not, the bet fails whatever else is true. The research agenda states that question with the evidence that would settle it.