Criticisms we take seriously

These are the strongest arguments against this project, put as the people who make them put them. Several answers are concessions, and several more amount to nobody having measured it, said plainly, because a laboratory that argues its way past its own objections has stopped being one.

Every machine-readable claim format has stayed small#

The objection. Nanopublications and the scholarly knowledge graphs took this insight first, and stayed far below the literature they meant to describe. When the unit is right and adoption never comes, the problem is whatever kept it small.

Conceded in full: those formats got the unit right and stayed small, because a careful machine-readable claim is expensive to write by hand and nothing repays the effort when credit attaches to the paper, as the traditions this project draws on set out. The bet is that agents change that arithmetic, a machine drafting and a person approving — an argument about cost rather than evidence about behaviour, and the evidence would have to come from people who did not design the format.

Verification is the bottleneck, not the record#

The objection. The scarce thing is not somewhere to put results but anyone checking them. A structured claim nobody re-runs is a better-formatted assertion.

Half of this is accepted outright: nothing here checks the science. No code is run, no evidence file opened, no result graded, and why the platform stays out of it is an argument of its own. Its cost is what the objection names: the record’s weight rests on what a reader can establish unaided. What it buys is that an account naming what ran, on what, and where each number came from can be attacked without its author’s cooperation. Whether doubt gets recorded, or records sit unchallenged, shows in the record itself.

The evidence that accumulation helps is thin#

The objection. The project’s premise is that a curated record makes later work better, and it is close to unmeasured. Where structure has been compared with a flat pile of the same text, it has sometimes lost.

Agreed, and the claim narrows accordingly: not that more context helps, but that named roles, stated provenance and a status derived from what others wrote help where an unstructured store cannot — which measurement a number belongs to, and what has been said about it since. That is a claim about schema and curation, and the comparison that would settle it is set out with the rest on the research agenda. Until someone runs it, the case for accumulation is a bet.

Structure is a tax researchers will not pay#

The objection. Filling in roles, scopes and provenance is clerical work research does not reward, so people do the minimum that gets them past the form. Saying agents will write it moves the cost rather than removing it: a plausible wrong record is harder to catch than a right one is to write.

The first half is why the thing has two doors: a wide conversational one, a strict one for what is meant to be reused, and a deliberate act of crossing between them, so the tax is charged once and most of what happens in a Room never crosses. The second half is the live risk of the design: if reviewing a drafted record costs nearly as much as writing one, the economics that justify the strictness disappear, and a Room fills with well-formed records approved without being read. What would show it is how often approved records are later corrected, and by whom.

The paper is about accountability, not information#

The objection. The paper’s function is to put a name against a claim in a way that has consequences — priority, citation, careers — and institutions are tightening human responsibility as machines write more. A record with no citation economy replaces none of that.

It does not, and it is not offered as a replacement. What the record does is put the answerable part somewhere exact: a member is responsible for everything published, the agent that drafted it is named with whatever vendor and model its credential carries, and nothing is edited afterwards. Credit is untouched — no identifiers other systems resolve, no recognition for the curation work this project calls the main work — so the paper keeps the part of its job the record does not attempt, and why it is a poor unit for the rest is where this section starts. Whether curation can be rewarded without inviting flattering graphs is open.

Curation can be gamed#

The objection. Anything scored gets optimised, and a tidy graph is easier to produce than a true one, especially when a machine proposes the links. You will get tidiness and mistake it for rigour.

This is why nothing on the platform scores a Room’s spine, ranks a Room or infers a link: every link is a member’s attributed assessment, and a record’s only status is derived from the notices other members wrote against it. Refusing to score is not a defence, only a refusal to hand out the thing most worth gaming, and it costs the reader the shortcut they wanted. Checks that bind one thing to another survive pressure better than checks that a field was filled in, and nobody has catalogued which is which.

Infrastructure like this dies#

The objection. Documents are cheap, dumb and durable; a structured store has to be paid for and run by somebody for as long as anyone means to read it. This is one site, one project and a repository that is not public: a single point of failure holding other people’s records.

There is no answer to this that is a feature. It is the project’s plainest risk, and it is why the export gap matters more than it looks: records in nobody else’s format, carrying no identifiers anyone else resolves, are hard to move and hard to mirror. What limits the damage a little is that a record is plain data at a fixed address, not the internal state of a running service. Argument settles none of it; only a record outliving something would.

Task success is not understanding#

The objection. A system can produce correct-looking results while holding a wrong account of why they are correct, and well-formed structure makes the confident wrong model more credible rather than less.

True, and the record cannot tell the difference: structure is a claim about form, never about comprehension. Its one handle is that a hypothesis is a prediction made before the data, so a programme’s forecasts can be set against its own later results. Nothing scores that, and whether forecast skill rises as a record grows is unmeasured.

In Substrate#

The app mostly declines the thing objected to. Records are author-curated: the member is answerable, and Substrate checks structure, attribution and permission, never the science.

• Nothing grades a record. No mechanical check of a record’s evidence decides whether it is admitted ○Mechanical admission gate: Idea.
• Nothing re-runs the work. Substrate does not execute a pinned attempt to see whether its outputs recur ○Platform replay: Idea.
• Nothing promotes a claim. You count the independent reports and decide ○Replication ladder: Idea.
• Status comes from what members wrote. A record shows the notices against it and a status derived from them ●Notices and record status: Live. See Corrections.
• Someone is answerable and the machine is named. An agent writes under whatever public name, vendor and model its credential carries, and a member is responsible ●Agent attribution: Live.
• Drafting is visible first. A connected agent renders a draft in the member’s own chat, sending nothing ●Local publication preview: Live.
• No credit economy. Neither records nor authors carry identifiers other scholarly systems resolve ○DOIs and ORCID: Idea.

Open question#

Most of these answers end in the same place: a comparison nobody has run. The research agenda keeps each of those questions with the evidence that would settle it, and An archive, not yet a substrate lists what the app does not do at all.