Moving from brand to process-based indicators of document quality.
In journalism and publishing, the way a document is edited, reviewed, and fact-checked before publication, and the way it is modified post-publication, are important indicators of thoroughness and trustworthiness. However, these indicators are not generally available, and when they are, are not generally machine-readable. It would be helpful to researchers, platforms, and readers if organizations published a discoverable, machine-readable “map” of changes alongside the original document to provide social platforms, aggregators, and users with a new first-party indicator of credibility. Ultimately, by analyzing large sets of review maps, we believe that researchers will be able to identify, classify, and validate the effectiveness of different types of reviews, leading to better practices across publishing communities and more transparency for readers.
MIT Knowledge Futures Group, PubPub Team
Travis Rich, Catherine Ahearn, Gabe Stein
Why do we trust documents that come from the New York Times or Science more than obscure news sites and journals? Because over time, the Times and Science have built trustworthy brands by repeatedly producing high-quality products using standard editorial and review processes. When the New York Times publishes something, we assume that its sources have been vetted, its claims fact-checked. And we assume that if it makes a mistake, it has processes it can use to correct it.
When publishing was an expensive endeavor, brand reputation built on the assumption of these processes was enough to differentiate between high-quality and low-quality sources of information. In many cases, merely looking at the quality of layout, typesetting, and printing was a decent indicator of quality. These indicators became a stand-in for what we were really trusting: the largely opaque process that went into producing the content.
Now that publishing online is inexpensive, more publishers are able to replicate the brand indicators that lend sources like The New York Times and Science credibility. And because many readers now discover articles on social networks like Facebook, Twitter, and ResearchGate, in many cases the work of these upstarts appears alongside more established brands, with identical formatting.
In some cases, this is positive, because it allows upstarts who employ robust processes and produce high-quality content to have a voice without needing to wait years to build brand equity. In other cases, however, upstarts who produce low-quality, or deliberately fake, material with no processes exploit this feature of modern information distribution to convince readers that they should be considered as credible as more established outlets.
Because of these weaknesses in the online distribution system, the brand quality signals built on top of good editorial processes are no longer useful to distinguish between high- and low-quality content. Instead of relying on brand signals, we propose that publishers peel back the layers on which reputation is built and make transparent what helped establish those brands in the first place: robust editorial and review processes. By publishing verifiable, machine- and human-readable maps of the processes that went into creating documents, high-quality publishers of all kinds can provide readers, researchers, and platforms with a new, more meaningful signal of trustworthiness.
In most kinds of publishing, documents go through many rounds of changes from draft to publication. And documents are sometimes updated after they’re published. Many of these changes are minor, such as spelling and grammar corrections. But some, like changes to facts and details after publication, can signal that a document was wrong or only partially researched when first published. Others, like peer review or fact-checking stages, can be a core signal of the credibility of a document. Taken together, all of the changes made to documents may reveal common review patterns that can be identified and used as credibility indicators by platforms and researchers.
For example: at what point does the frequency and type of a publication’s corrections to news articles go from a positive signal about its willingness to correct the record to a negative signal about the quality of its fact-checking? Is there a type of peer review that leads to fewer retractions? What is the ideal length of a public comment period for proposed regulations?
Review maps are intended to make asking these questions possible by exposing document change and discussion data as machine-readable, first-party credibility indicators that are currently inaccessible to platforms, readers, and researchers:
Changes between draft and published versions of an article
Identity of editors and fact-checkers (including computer programs)
Changes made to an article after publication, as demonstrated by Newsdiffs
Full or partial retraction of an article
Type of peer review an article received
Changes between submitted and published versions of an article
Reputation of reviewers
Community discussion that lead to changes in an article
Changes between preprint and published versions of an article
Full or partial retraction of an article
Changes in bills from proposal to submission to passage
Amendments to bills
Identities of amenders
Review Maps could be automatically generated by the systems already responsible for tracking changes and comments on documents: CMS systems. Developing a WordPress plugin and working with large publishers to implement review maps (potentially with funder support) would make adoption simple for many publishers. For publishers that can’t or don’t want to support review maps, third parties could create limited versions of them by periodically scraping articles and compiling their own maps that platforms and researchers could use (i.e. Newsdiffs). If maps became a commonly used signal, publishers would be incentivized to adopt their own, more complete maps rather than rely on incomplete third party versions.
We propose that review maps be published as separate, structured documents alongside original articles. They should be made discoverable to platforms, crawlers, researchers, and others via HTML meta tags, in a similar fashion to the open graph protocol.
Using this approach, maps can be published separately from their source documents, allowing browsers to request maps optionally, which decreases load time and data transfer. Separating maps and source documents also allows syndicated documents to point to maps on remote servers, avoiding map duplication and allowing for the creation of centralized review authorities and aggregators.
Badges, extensions, authorities. Developer vs. reader.
Similar to RSS, review maps need not be format-specific. They could be published in multiple formats and file-types designed for specific use-cases and rely on some common
<meta property="review-map" content="https://server.com/maps/article-map.json" />
const map = [
document: uuid /* e.g. versionId */
Map: A map is a series of objects (we need a name for these objects - stages?), each of which is anchored by a single document
Stage.document: A URI or uuid to a specific document. In PubPub’s case, this is a versionId
Stage.template: A template defines the parameters expected from the review process. The template is used to organize events for privacy, blindness, and meaning. A full review map may have single objects whose templates are never completed, due to either the introduction of a new version, or insufficient events. (There is some goofy language here. A template as described before is the completed set of details on how a review will go, with specific reviewers, etc. We may want template templates, that for example create 3 blind reviewers, and let you simply select the individuals. Maybe we call these ‘pre-built templates’ or something like that to avoid the confusing ‘template templates’ issue.
Stage.events: Events are objects that describe action taken within the context of a review. These can be notes, suggested changes, decisions, etc.
Should this be a rigid standard, a loose one, or merely an implementation, like a sitemap?
Who are the core “users” of review maps — platforms, researchers, publishers, or readers — and how do their uses differ?
Should events be signed in some way so that they can be authenticated against and confirmed?
How will bad actors attempt to game and manipulate review maps for their benefit?
How do we handle the identity question for reviewers and editors? In some cases, reviewers will need to be anonymous, but we’ll still want to aggregate their reviews. In other cases, we’ll want to federate identity with third-party tools.
Every element of an event (except its id perhaps) should have a variety of privacy settings, as groups won’t always want to expose details about events to the public. The decision, the notes, the identity of the reviewer can all be private, public, or restricted to a limited audience.
Can elements within the review template have privacy settings too? Perhaps we don’t want to show the due date assigned to a reviewer.
How do you make template[n].dependency complex? To support things like: ‘when there is consensus, invite HR to review’, or ‘unless unanimous reviewer approval, reject’.
How do we make event.decision diverse? How do we support binary, 1-10, etc.
Which of these features are core to ReviewMaps, and which are specific to individual platforms/organizations?