The horizon

Every other page in this section argues from something the app already does. This one sets out what becomes possible if the record holds — if claims written with their evidence and provenance turn out to be worth keeping at scale — and none of it is built. Nothing here is a plan, a sequence or a promise, and the project may be wrong about which of these matter.

What they share is that each begins as an engineering problem and turns into a question about who decides something.

Asking the record a question#

A record answers the questions its reads were built for. Ask one they were not — which measurements of a quantity exist across every Room, and which have been disputed — and the answer is assembled by hand out of a dozen reads. The edges are in the record; nothing walks them.

Two capabilities would change that: traversal, following the typed edges out from any record through the attempts, materials, hypotheses and corrections around it in one read; and a structured question asked of every Room at once rather than of one object at a time. A third would let the record be asked where it is silent — which pairs of concepts nothing connects, which claims nobody has tried to reproduce.

The difficulty is not the query engine. An answer gathered across Rooms asserts that the things gathered are comparable, and the record does not know whether they are: twenty findings that agree may be twenty reports of one run or twenty groups converging — the distinction reproduction, replication and independence turns on — and that difference is most of the answer. A query language also decides what gets written: whatever it can express is what people shape their records to fit.

Two Rooms, one meaning#

A concept means what the record defining it says, and equal labels never merge, so one idea splinters across records with no way to gather it. The horizon is a layer above that rule rather than an exception to it — two Rooms declaring their definitions the same, as an attributed and reversible act with its grounds written down.

Who decides is the whole problem. A committee agreeing a vocabulary is what defining concepts locally was meant to avoid, and it does not scale. A matcher deciding automatically scales and is wrong quietly, which is worse than fragmentation, because a wrong join leaves no mark: two definitions become one and nobody sees the seam. No rule yet separates the joins a careful author would accept from the ones they would not, so the safe rule stays, and the fragmentation with it.

Checking without trusting the author#

Every record is its author’s account, and writing it carefully makes that account attackable rather than true. The rung that does not rest on the author is somebody else running the work.

Three shapes of that are missing: a service that re-executes a pinned attempt and compares what comes out against what was reported; receipts attested by the environment that ran the work, so an execution is witnessed by something other than the person reporting it; and review records, an attributed assessment of a claim against its evidence.

Re-execution costs compute on somebody’s machine and somebody’s budget, and the results most worth re-running are the most expensive to run, so whoever pays decides what gets checked. It has to resist the author too: if an author can predict which attempts are re-run, the check selects for what is safe rather than what is doubtful. The other two shapes turn on who is willing to stand behind them — an environment that attests to what it ran, and an assessor whose judgement is worth the name against it — and neither is something a mechanism supplies.

A ladder that is computed, not declared#

Between what an author asserts and what someone unrelated found, there are rungs a reader can weigh when nobody can verify every result, several of which look mechanical: a plan accepted before the attempt that cites it, a second member reporting the same values. A ladder would derive a record’s standing from facts like those instead of from anyone asserting it.

Two things make that dangerous rather than merely hard. A published rung is a specification for the cheapest way to reach it, so the ladder gets climbed rather than earned. And a rung reads as a verdict: a label from the platform is cheap to produce and easy to take on trust, the signal this project argues against everywhere else, so a ladder would have to stay legible as a list of checks that passed and never as a judgement that the science is sound.

Records the rest of scholarship can read#

A record here is JSON at its own address, in none of the graph formats scholarly systems read and with none of the identifiers they resolve, so nothing written here can be cited where a durable identifier is required.

Mapping onto those formats is the obvious move and the lossy one. Roles are open: an author defines the roles a statement needs, which makes an unusual statement expressible and is exactly what a shared vocabulary cannot absorb, so a mapping either drops the roles or asserts correspondences nobody agreed — the concept problem again, in a larger room. Identifiers are hard differently: one is worth what its resolver is worth, so issuing them is promising to be answerable for decades, an institutional commitment rather than a piece of software.

What it would add up to#

Together these describe a record other people’s systems can build on: something an agent nobody here configured can query, correct and extend, that another tool can read without asking permission, and that outlasts the software it was written with. That is a different ambition from a successful website. A successful site has readers; a successful record is one whose readers mostly never see the site, where the question is no longer whether the pages are good but whether what is written in them is.

Measured against that, the honest description is an archive rather than a substrate: records kept carefully, and little done with them. Which of these gaps matter is not obvious from the page, and some may turn out not to.

In Substrate#

Everything above is listed below with the registry’s mark against it. All of it would build on a record whose admitted versions stay as written, at addresses that do not move ●Exact versions: Live.

• Walking the graph. A read that follows the typed edges out from an arbitrary record ○Neighbourhood read: Idea.
• One question of every Room. Structured questions asked across Rooms, not object by object ○Cross-Room queries: Idea.
• Where the record is silent. Questions raised from what the record connects and leaves out ○Literature-based discovery: Idea.
• Declared sameness. A store in which two Rooms say their definitions mean the same thing ○Cross-Room concept identity: Idea.
• Links that cross Rooms. A correction or dispute naming a record of another Room ○Third-party scientific links: Idea.
• Re-execution. A pinned attempt run again and its outputs compared with what was reported ○Platform replay: Idea.
• Witnessed executions. Receipts attested by the environment that ran the work ○Signed receipts: Idea.
• Assessment with a name on it. Review records weighing a claim against its evidence ○Referee review: Idea.
• A computed ladder. A record promoted towards confirmed as independent replications accumulate ○Replication ladder: Idea.
• Export. Records serialised as RDF nanopublications ○Nanopublication export: Idea.
• Identifiers that outlive the site. Records and authors resolvable by other scholarly systems ○DOIs and ORCID: Idea.

Open question#

Which of these a record needs before other work can build on it — and whether this page has the list right — is what it is least able to answer about itself. The research agenda keeps the question with the evidence that would settle each part.