How claims relate
Support, contradiction, refinement, correction and supersession between records
Because a record never changes once it is admitted, the only way a Room’s knowledge moves is by accumulation, and most of what a reader needs to know about a new record is how it stands to the old ones. A paper says this in prose: “unlike [12], we find”, “this corrects our earlier estimate”. A reader arriving long afterwards with a version id and a question about one number cannot follow that, and neither can an agent choosing which of forty results to build on.
A typed link puts the relation in the record rather than in the sentence: this record, that record, this kind, this explanation, this author. The question “has anything happened to what I am about to cite?” then has an answer that does not require reading everything published since.
Two kinds of link, doing different jobs
The first kind is written into a record when it is published: its references to other exact versions, each with one purpose from a fixed list and an explanation. The same purposes apply when a publication is reused in another Thread or Room Reuse into Threads: Live. A reference is part of its record and changes nothing about its target.
| Purpose | The author is saying |
|---|---|
| premise | This work takes the target as a starting assumption |
| supports | In my assessment, this supports the target |
| contradicts | In my assessment, this contradicts the target |
| extends | This goes further than the target in the same direction |
| refines | This narrows or sharpens the target |
| derived_from | This was developed from the target |
| related | Relevant, with no stronger relationship claimed |
The second kind is a record in its own right: a research link Research links: Live, attributed, explained and immutable, asserted after the fact by someone who need not be the author of either side. These do act on their target.
| Kind | It says, and how the target then reads | Asserted by |
|---|---|---|
| corrects | A source record fixes an error in the target, which reads as corrected | Any member who can post in the Room |
| supersedes | A source record replaces the target as the current statement, which reads as superseded | Any member who can post in the Room |
| retracts | The target is withdrawn, should not be relied on, and reads as retracted | The target's author, or the Room's owner |
| disputes | The asserter disagrees with a record, which reads as disputed, or with a link, which counts it on that link's notice | Any member, once per target |
A correction and a supersession each need a source record; a retraction has none, because nothing replaces the withdrawn record; a dispute may name one but does not need to. Where a record carries links of more than one kind the strongest reading wins: retracted over corrected, over superseded, over disputed — worked out from the links each time the record is read, and stored nowhere.
What a link does to its target
Not what the word “correction” suggests. The target’s content is never touched: it keeps its wording, its version id, its address and its label, every reference pointing at it still resolves, and it can still be cited Exact versions: Live. What changes is what a reader sees around it: a notice Notices and record status: Live naming the link, its kind, its author, the explanation they wrote and the source record, together with the derived status.
The notice travels outwards one step. A record whose own exact references name a version that has since been retracted, corrected or superseded carries a warning saying so Citation warnings: Live, naming that version and the link, whatever purpose the reference was given. It is advisory: it never refuses a write and never alters the citing record. Somebody has to decide what it means, and that somebody is a person.
A link is not a verdict
Every one of these is one author’s judgement put on record, not the platform deciding who is right: a contradiction says that this member, on these grounds, reads the two results as incompatible. Nothing ranks two corrections of one target; both notices show, and two records can sit contradicting each other with no resolution, which is the honest state to be in while a question is unsettled.
The types invite three mistakes worth naming. Transferring a result to a new domain and finding it does not hold there is not contradicting it. Failing to show a difference is not showing equality. Naming a record as a premise is not asserting that it supports you, only that you are standing on it.
Relating a finding to the experiment it came out of can therefore carry two verdicts rather than one Findings linked to experiments: Live: whether it supports, contradicts, is mixed on, is inconclusive about or does not apply to the hypotheses, and separately whether it answers, partly answers, does not answer or does not apply to the experiment’s own question. Those are different facts, and a link that carried one number for both would lose one of them.
Growth by addition
A correction is a new record plus a link; a retraction is a link; a disagreement is a link that can itself be disputed. Let knowledge grow by addition, so nothing true is ever quietly overwritten and nothing false is ever quietly forgotten. The price is that a reader has to read the notices instead of trusting a headline, and that a mistake stays visible for as long as the record does. One guard keeps the shape sane: a chain of corrections or supersessions that would lead from the target back to the source is refused, so no record replaces itself through a loop.
Two records, one Room
A link speaks about its target, and about the source where it has one, and about nothing else: not the author’s other work, not the state of the question as a whole. Links stay inside one Room too: the source, the target and any disputed link all belong to the Room the link is asserted in, and a link that would cross Rooms is refused. Reuse works the other way round, needing membership only in the Room you are writing in. So a Room can build on another Room’s record, and say in its own record how it reads it, but cannot place a notice on it.
Third-party scientific links: IdeaThird-party scientific links
Scientific links could reach across Rooms, and out to results published elsewhere, so that a Room stating its finding contradicts another Room’s could leave the same notice on that target.
A constructed example
A constructed example, in a Room comparing two training recipes. A finding reports that on the held-out split recipe B reached 4.1% error against 5.2% for recipe A, citing the metrics file the run produced. Weeks later its author rereads that file: 4.1% came from the validation column, the held-out figure was 5.4%, and the comparison goes the other way. They publish a second finding with the corrected numbers, then assert a correction from it to the first, explaining which column was read.
What each reader meets:
- On the first finding. The original wording, unchanged and at the same address, with the status corrected and a notice giving the link’s label and kind, its author, the explanation and the record that corrects it. The 4.1% is still there to read: anyone who quoted it can see what happened to it.
- On the second finding. An ordinary finding. Nothing in its assertion announces itself as a correction: the link is the statement, and it carries the explanation and the attribution.
- On a hypothesis published earlier that named the first finding as a premise: a warning that it cites a corrected version, naming the version and the link. The hypothesis is not edited, not withdrawn and not refused; its author decides whether the prediction still stands.
This corrects a transcription; it does not dispute the science. Reading the wrong column and doubting a result take different kinds. A member who thought 4.1% right after all would have disputed the correction itself, which is a target like any other.
Further reading
- David Shotton, CiTO, the Citation Typing Ontology, on giving a citation a stated nature instead of leaving it to prose
- NISO, RP-45-2024, Communication of Retractions, Removals, and Expressions of Concern
- Crossref, Crossmark, on showing a reader the current status of what they are reading
In Substrate
- References and reuse. The purposes, and how to reuse an exact version in another Thread, are on Reuse and exact versions.
- Corrections and notices. The four kinds, their authority rules, and how notices, status and warnings appear in the browser and over the API, are on Corrections, disputes and retractions.
- Verdicts on an experiment. Relating a finding to the experiment and the hypotheses it bears on is on Findings and cited claims.
- Nothing is computed from the links. Substrate derives a status from the kinds that target a version and counts disputes, and nothing more: no explanation is weighed, no two records are ranked against each other, no target is edited or hidden. A link is author-curated like the records it joins.
Open question
Typed links show how two records stand to each other and nothing about the shape of a research programme: which line of inquiry a contradiction closed, or what a run of findings is converging on. Whether that belongs in the record as an object of its own, or can only be written by a person in prose, is on the laboratory’s research agenda.