Policies

How Shipped works. These are the rules we apply to every item and every reader.

What we list

We list real AI releases and real deployments. Each item points to a primary source, a date and a maker. We do not list rumours, leaks or demos with no way to use them.

The board has three sections.

• Shipped. Announcements and deployments of AI that is real, useful and in use.
• Open weights. Releases whose trained weights anyone can download, and what people do with them. "Open weights" means the weights are public. It does not say anything about the training data or the licence, so we state the licence on each item.
• Decisions in the wild. Decision models used in real deployments. A decision model makes a judgement instead of writing prose: a label, a yes or no, a score or a choice of route. Guard and moderation classifiers, judge and reward models, and routers all count.

An item is one release. A minor version of the same model family in the same month is the same item.

Status marks

Every item carries one mark. The mark describes access, not quality.

• Publicly available. Usable by anyone, according to the facts documented in the linked sources. An editor confirmed it, or, once readers sign in, 3 signed-in readers whose accounts are more than 24 hours old report using it. Reader sign-in opens soon; until then only an editor applies the mark.
• Pending. Announced, with access limited: a waitlist, "coming soon", or partners only.
• Vapor. Announced and not delivered, well after the promised date, and we have evidence of that. A late release is not vapor. A missing one with a paper trail is.
• Verified. An editor checked the evidence by hand. Verified is added on top of Shipped. Nothing is marked Verified until someone has done the checking.

Editors set the marks. We change a mark when the facts change: a waitlist closes, a release arrives, a date passes with nothing to show. When a mark changes, we say so on the item. Where we have recorded who set a mark and when, the item shows both; where we have not, it says nothing.

Scores

An editor scores each item from 1 to 5, in half steps, on four dimensions.

• Real. Can people use it today.
• Useful. Does it solve a real task.
• Notable. Does it matter to the field or its niche.
• Novel. Is it new, or a repackaging.

Each score comes with a one-line reason, based on documented facts.

The hype gap is the mean of Notable and Novel, minus the mean of Real and Useful. A positive gap (▲) means the item sounds bigger than it is: more noise than use. A negative gap (▼) means it does more than it says. Zero is level. We show the gap to one decimal place.

AI-assisted first pass

Items that we seed are scored first by an AI research pass over the linked sources. They are labelled as such on the item, and a human editor will re-score them. The label stays until an editor has done that.

Reader ratings

Reader sign-in opens soon. Once readers sign in, they can rate the same four dimensions from 1 to 5 and say why, and an item will show three separate things, each labelled:

• Upvotes. "N upvotes": how many readers want it seen.
• Ratings. How many readers rated it, dimension by dimension, and what they gave.
• The editor's score. Shown on its own. It is the headline figure.

A reader mean needs at least 5 ratings in its dimension before it is printed. Under 5 it is too few to show, and the reader hype gap is not published until every dimension has 5 ratings. Ratings still move the ranking a little from the first one, by the rule under How ranking works. Editor scores stay editor scores.

How ranking works

The Front ranks items by one score. Here is the whole of it.

1. Each dimension is pulled toward the editor's score. For each of Real, Useful, Notable and Novel, the editor's score is the starting point. Readers move it toward their own mean by a weight of n / (20 + n), where n is the number of reader ratings on that dimension. The editor's score counts as 20 ratings, so 5 ratings give a weight of 0.2 and 20 ratings give 0.5.
2. Readers never get more than 0.4. However many readers rate, the weight is capped at 0.4. The editor always keeps at least 0.6 of the say. A dimension with no editor score takes the readers' mean.
3. New accounts count for a third. A rating from an account younger than 24 hours, at the time it was cast, counts at one third of a rating, both in the readers' mean and in n.
4. The base score. The mean of the four adjusted dimensions that have a score. Until readers sign in, there are no reader ratings, so the base score is the editor's alone, which is the case for every item today.
5. Time decay. The base score is multiplied by 0.5 raised to the item's age in hours divided by 48, so its weight halves every 48 hours. An item that is 48 hours old counts for half, and one that is 96 hours old for a quarter.
6. An item with no editor score and no reader ratings is not ranked.

Upvotes do not enter the Front score. Rising is separate: once readers can upvote, it orders items by the upvotes they received in the last 6 hours, counting only upvotes from accounts older than 24 hours.

The numbers behind the ranking are refreshed on a schedule, every 6 hours, and once voting opens also shortly after readers vote, not live. Each board says how long ago it was updated. Today the home and the other release lists order items by release date, newest first; the ranking above does not reorder them yet.

Sources and evidence

Primary sources come first: the model card, the repository, the paper, the vendor's own post or a public record. A news story can point us to a release, but it is not the evidence for it.

Each evidence link on an item has a check state.

• Checked. An editor opened the link and it shows what we cite it for.
• Partly. It shows some of what we cite it for.
• Unchecked. Listed, not opened.

We do not invent numbers, benchmarks, licences, sizes or dates. If a source does not state a fact, we do not state it either. We quote at most 15 words, in quotation marks and with attribution, and at most one quote per item. Everything else is paraphrase.

Corrections

If we got something wrong, write to an editor. Tell us which item, what is wrong and what source shows it. We check the source, fix the item, and note the correction on the item. We do not quietly rewrite a mark or a score.

Reader contributions

Reader sign-in opens soon. Once readers sign in with Google, they pick a handle and can submit links, upvote, rate, say they used something, comment and report problems. The rules below apply from the day sign-in opens.

• Public at once. Submissions, votes and comments appear straight away. We do not hold them for review first.
• Moderation is after the fact. We may review what is public, and we hide what breaks these rules. Hiding something leaves a record.
• Rate limits. Each reader has limits on how much they can post in an hour or a day, with tighter limits on new accounts.
• No doxxing and no harassment. We hide it and may ban the account.
• Pseudonymous. Readers appear by a handle they choose. Shipped never publishes the identity they signed in with.
• Submissions appear as Pending until an editor has looked at them.

Karma

Once readers sign in, karma is a number shown beside a reader's handle. It is the upvotes other readers give your submissions, plus 2 for each report from another reader that they used something you submitted. Upvotes and reports count only from accounts older than 24 hours, and a report counts toward the 3-report rule for Publicly available only from such an account. Your own votes on your own submissions do not count, and votes on a hidden submission earn nothing. Karma does not change what you can do, and it does not change how items rank.

Disclosure

Items that involve the publisher's own products carry a disclosure line saying so. We score those items by the same rules as the rest, and we show our work.

Privacy for readers

You can read everything on Shipped without an account. Reader sign-in opens soon; once you sign in, this is what is public and what is not.

• Public: your handle, your comments, and the reason you give with a rating. Votes and ratings are shown in aggregate (counts and averages, not who gave them).
• Not public: the identity you signed in with, and the name or email address attached to it. Shipped never publishes it.
• Private to you: your own votes, which you see when you are signed in.

We will not rely on the handle to hide who you are. Pick one that does not give you away.