Methodology

How something becomes publishable here, and what stops it when it is not.

The claim is the unit, not the paper

Papers bundle many claims of very different strength behind one title, one date and one impression. Here the claim is the object that has an identity, a version, a status, evidence, falsifiers and a citation. A paper, when one exists, is an artifact attached to claims — not the other way round.

Each claim is decomposed into components, and each component carries its own support state. This is what makes it possible to say the honest thing that a paper abstract cannot: this part is directly measured, this part is inferred, and this part is carried by nothing at all.

The publication state machine

A claim moves through states, and each transition has a requirement that is checked mechanically:

DRAFT
  → SCIENTIFICALLY_APPROVED   evidence attached, falsifiers registered, coverage map complete
  → SECURITY_REVIEW           secret scan, private-provenance scan, classification check
  → LICENSE_IP_REVIEW         every artifact has a licence; ambiguous ownership blocks release
  → PUBLICATION_BUILD         record frozen, canonical digest computed and sealed
  → PREVIEW                   rendered and checked, not reachable from public indexes
  → PUBLIC                    authorised release
  → SUPERSEDED | WITHDRAWN    new version supersedes; withdrawal propagates, deletes nothing

Some transitions are forbidden outright and are property-tested in the build:

WITHDRAWN → CURRENT is impossible. A withdrawn claim can only be replaced by a new version carrying new evidence.
FALSIFIED → QUALIFIED without new evidence is impossible.
A replication state cannot be asserted without a replication object that actually supports it.
A material scientific change without a new version, or without updating the material-change timestamp, is impossible.
A published version's digest may never change. Editing one in place fails the build.

Reproduction is not replication

Four different things are routinely reported as if they were one:

What happened	What it establishes
Artifact verification	You received the bytes we published. Nothing about the science.
Computational replay	Our code is deterministic on our data. A property of us.
Experimental reproduction	We can re-run our own protocol and get our own result.
Independent replication	Someone else, with their own implementation and data, met criteria fixed before they started.

Only the last one is independent replication, and a claim may only assert it when a replication object exists whose people, implementation and data are all independent. Replications record their independence across eight dimensions and are not ranked against each other — the dimensions are more informative than any ordering we could invent.

Challenges are contracts

Criticism that cannot be settled is noise. A challenge here is a machine-readable object naming its target version, its attack surface, the criterion that would falsify the claim, the apparatus required, and what evidence would count as an answer — all fixed before anyone runs it. Agreement and disagreement change nothing; execution does. There is no voting anywhere on this site.

Evidence debt

What a claim still owes is published alongside it as typed objects: a missing replication, an unexecuted falsifier, a measurement gap, an unresolved contradiction, a release restriction. Debt is never summed into a score. A missing replication and a known contamination pathway are different problems, and an average of them describes neither.

Negative results are first-class

A falsifier that executes and fails is a result, and it is published with the same identity, versioning and citation machinery as anything else. The demonstration claim on this site ends in withdrawal, and that withdrawal is the most informative record in it.