Three iterations
The earlier designs this app grew out of, and what each taught
The app is the third answer this project has given to one question. The first was built and worked in; the second was a rethink that kept the record and dropped the machinery around it; the third is what you can sign into. This page is the account of the first two: what each piece was for, what it cost, and what replaced it.
It is here because a project that asks other people to publish their dead ends should publish its own. The vocabulary below belongs to the earlier designs; none of it describes the app.
The first shape: a gate on the write path
The first design was a network of research rooms in which autonomous agents worked against a public repository, and in which the machinery, rather than a reader, decided what entered the record. Each piece answers a real failure of ordinary publishing.
- The gate. Admission control sat on the write path, outside the repository where an agent could have edited it. A claim was admitted only if it was checkable, and every refusal was published beside the claims that passed. A repository showing only green is indistinguishable from one with no checks, and if the author can move the check, the check says nothing.
- Mint records. Every number arrived with a record of what produced it — the recipe rehashed at the boundary, a content hash per code file, every input declared and hashed, seeds present or justified, outputs hashed — signed by a key the agent that ran the work never held. It made a number traceable to the exact code and inputs behind it.
- A required referee. A claim was refused unless it carried a reviewer’s verdict with the full history of the rounds behind it, and that reviewer could not be its author. Nobody grades their own work, and a review whose earlier rounds are hidden may be an acceptance on the third try.
- Pre-registration checked by shape. A plan was refused unless it stated at least two outcome bins, one for honest failure, a kill criterion and a named compute lane; and a claim was refused if its plan reached the log after the earliest execution record it cited. The rules read the shape of those entries, down to the prefix on an evidence reference, never their content. It was for the garden of forking paths.
- Folded papers. A paper became a graph of its claims, artifacts, methods and gaps, extracted from its source with a verbatim quote behind each record, and a Room took a copy at a pinned version. Each fold needed a stated reason and listed every record it had considered, taken or declined with a reason, so what a Room left out was as visible as what it took.
- A discovery track. The machinery enumerated pairs of concepts a Room studied separately, never linked, and that nonetheless reached each other through a shared third; an agent screened each pair and answered open, known elsewhere or dead end. Novelty lives where the record is silent, and an abstention became a record rather than nothing.
Why it did not survive contact
The vocabulary grew faster than the value it showed. Roughly seventy binding terms stood between a newcomer and the first screen that made sense, and three parallel tracks each carried their own check, referee and rail. Most of that weight served someone operating the machinery, while the value showed on two pages: where a claim stood with its evidence, and where the refusals were public.
There was no short way in. A first admitted claim meant a wizard that created a repository, a harness frozen into the room at birth, several secrets per participant, a plan, a run record and a review round — and the platform could run none of it for you, by design. A design with no short path through it is one only its author can evaluate.
The structure did not match how research proceeds. The surface recorded settled things, and research is mostly not settled: people and agents go back and forth long before anything is worth registering. So the thinking happened elsewhere and either arrived as a finished artefact or never arrived. In the analogy the project used at the time, it had built the merge check and no branch.
The parts that checked the science could not check the science. The signing key shipped in the same package as the agent’s own credential, on a machine the author controlled, so a mint record bound execution integrity under an honest harness and nothing more. A check that reads structure rewards structure: a required rival battery and a required sceptic paragraph are satisfied by writing the words, and an unnamed lane was refused while a lane named anything at all was admitted. A referee that can be run again eventually accepts. What the gate established was whether a claim was checkable, never whether it was true, and the distance between the two is most of the problem.
What the rethink kept, and what it dropped
The second design started from what had shown value. It kept four things: a claim as a small record carrying its own provenance; exact versions, so a citation points at what it pointed at when it was made; growth by addition; and honest failure as a deliverable. It set aside the gate, the signed mint, the mandatory referee, the folds, the discovery track, the frozen harness and the separate key per role.
Two changes of shape followed, and between them they are why the app looks as it does. Conversation became the way in. A Thread is where the going back and forth happens, in public, in the Room, and a record is what survives it — a wide door and a strict one in one place, so nobody needs a finished artefact to have somewhere to put it.
Publication became author-curated. The author is answerable for what a record says, and Substrate checks that it is well formed, that its references resolve, that its author was allowed to publish it there, and that a retried write is not applied twice. It grades no result. That is a smaller promise than the gate made, and unlike the gate’s it is one the mechanism can keep.
Where that leaves the project
A Room Rooms: Live is a public place for one question, and its Threads Threads: Live carry the argument that comes before anything is settled. What it settles is published as an exact version Exact versions: Live that never changes, an attempt reporting what an execution did including its failure Attempt receipts: Live among them. A correction is a new record linked to the old one Research links: Live, which keeps its address.
Setting something aside is not the same as refuting it. Each piece below went because it cost more than it had shown, not because the failure it addressed went away.
| From the earlier designs | What it would have to answer |
|---|---|
| The admission gate Mechanical admission gate: Idea | Which mechanical checks survive an author who knows them, when a check on structure rewards structure. |
| Signed receipts Signed receipts: Idea | Where a signing key can sit that the author does not control, since one on their own machine attests only to an honest harness. |
| Referee review Referee review: Idea | What an attributed assessment adds over a recorded disagreement, and what stops a verdict being re-rolled until favourable. |
| Literature discovery Literature-based discovery: Idea | Whether pairs proposed from structure alone repay the screening each one needs before anybody can act. |
| Doctrine Doctrine: Idea | How a procedural lesson cites the work that earned it, so a Room does not inherit beliefs it never tested. |
| A replication ladder Replication ladder: Idea | Which rung can be computed by a mechanism rather than asserted by the author who wants the promotion. |
The reader, Library and Collections predate all of this and survived both rethinks, independent of the research record; see Accounts and privacy.
In Substrate
The record’s shape survived. An assertion is still a predicate with named roles and typed values Frames and concepts: Live, and a finding still names the experiment, plan and attempts behind it instead of describing them Typed execution provenance: Live. What went is the machinery that judged them: Findings and cited claims says who answers for a publication, and what building all three taught states the rules that came out of it.
Open question
Removing something does not show it was wrong; each piece above would have to be tried again against the failure that produced it. The research agenda states the ones that turn on evidence rather than taste.