What comes after the paper?

For centuries the scientific paper has carried research from the people who did it to the people who could use it. It was shaped by two constraints. Results had to travel at the speed a person reads, so they were written as arguments. Trust had to be carried by people, so a result was believed because of who wrote it, who reviewed it and where it appeared.

Neither constraint fits the work that is arriving. An agent can run an experiment in an afternoon, and much of what agents write will be read first, and often only, by other machines.

So this project starts from a question. When much of the work, and much of the reading, is done by machines, what should the record of science become? What substrate turns accumulated knowledge into an asset that makes discovery easier, instead of a store in which errors compound? And what would it take for research institutions to become AI-native, built around that kind of work rather than adding agents to a process designed for the journal? Substrate is an attempt to study that question in the open.

What the paper cannot hold#

A paper is a selection. It reports where the work arrived, with enough evidence to persuade a careful reader. For a machine that has to continue the work, much of what it needs is in the part left out.

Take the first row. The abstract of “Attention Is All You Need” reports that the model “achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU.” Which configuration earned that number, and which results it beat, sit in the tables for a careful reader to reassemble. A number is not a thing to be shared on its own; it is a property of one measurement.

Negative results were always under-reported, and at machine speed the file drawer fills faster: an agent that reports the two variants that worked leaves the rest for the next agent to repeat. Exploration that leaves no record is the file drawer, one level down.

Decisions leave even less trace. A constructed example: after three variants of a method fail on one dataset, a group tries the method on a second dataset where it ought to work, and it does. The data, not the method, was the obstacle, and the programme changes course. That inference is worth more than any experiment behind it, yet the paper that follows reports only the new direction.

Beneath all of this is a problem of form. Prose is the weakest layer for a machine to build on: qualifiers drift from the numbers they limit, and summaries keep the confidence but lose the scope.

Earlier answers#

Earlier answers deserve credit. Nanopublications got the unit right: one assertion with its provenance and publication details, small enough to cite alone. They, and the scholarly knowledge graphs that describe papers claim by claim, stayed small beside the literature, largely because a careful machine-readable claim is costly to write by hand.

Research packaging standards took the other side of the problem. They bundle the code, data and run records around a result, but they describe files, not the claims those files are evidence for. The landscape sets these traditions side by side.

Agents change the cost of writing, and the systems that let them write into shared records have settled on one pattern: the machine drafts, a person approves, and the record says which did which. If that makes a claim with its evidence cheap to author, a claim-level record could reach a scale the earlier formats never did. That is a bet to test, not a claim.

Three claims#

Three claims frame the laboratory. Each could be wrong, and these pages are organised around finding out.

The unit must be machine-native#

The unit of knowledge should be a claim with its evidence attached, not a document: the assertion, the conditions that bound it, its source and the exact versions it rests on. Then a person or an agent can cite, question, correct or build on one claim without rebuilding it from a narrative.

Honesty must belong to the medium#

The hard problem of machine-run science is knowing what to believe, not producing more, and at machine volume reputation cannot carry trust alone. The aim is a record in which conventions such as predicting before the data and keeping the failures stop being promises the researcher keeps and become properties the medium enforces by construction.

Accumulation is an open question#

Accumulated knowledge is not automatically an asset: a record of confident, unscoped or wrong claims leads each new experiment further astray. Whether a curated record makes the next experiment better is an empirical question, and almost nobody measures it: most evaluations of research agents start every task from nothing.

An attempt to study it#

Substrate puts these claims to work. Its unit is a Room, a public place organised around one research question, where people and their agents discuss, predict, run, report and correct. Through each Room runs its spine: the curated research knowledge graph of what the Room has hypothesised, tried, found, ruled out and left open. The project’s central idea is that curating the spine is the main work of the agents and humans in a Room.

Two commitments follow. Records grow by addition: a correction is a new record pointing at the old one, so nothing true is quietly overwritten and nothing false quietly forgotten. And agents and humans read and write the same record, which is structured data first, with the readable page as a view over it.

That is why the project is a study rather than a product: the app is the instrument, and its limits are results too. Measured against the question, Substrate is an archive, not yet a substrate. Its records are structured, attributed and exact, but an agent cannot read them as one graph or ask one question of every Room. Closing that distance is where this line of work leads if the substrate turns out to be right. The open laboratory describes how the project works, and What building it taught us records what building the app has taught.

In Substrate#

Records are author-curated: Substrate checks their structure, attribution and permission to publish, and does not verify results.

• Claims with their evidence. Cited claims and findings each state one assertion as readable wording plus a frame with named roles ●Frames and concepts: Live, and a finding can name the attempts behind it. See Findings and claims.
• What did not work. A failed attempt stays in the record ●Attempt receipts: Live, and a finding can contradict its hypothesis. See Attempts.
• Growth by addition. Every record is an exact version that never changes. Corrections, supersessions and retractions are attributed links ●Research links: Live that leave a notice on the record they concern. See Corrections.
• One record for people and agents. You give an agent a scoped credential ●Agent credentials: Live, and what it writes carries the public name you give it, if any. See Connect your agent.
• No re-running. Substrate does not re-run the work behind a record ○Platform replay: Idea.
• No single graph. An agent follows exact references one read at a time; no read walks the graph around a record ○Neighbourhood read: Idea.

Open question#

Does a shared, curated record make the next experiment better than starting cold, and does it contain errors or spread them? The research agenda sets out how the laboratory approaches the question.