> Story drafts about what is rising or cooling in a publication's weekly map, where every sentence cites the counts and releases behind it and code checks the citations. Built; drafted by AI, steered by readers, decided by editors.

Canonical: https://p.apyr.us/cited-trends.html

# Cited Trends

Cited Trends are story drafts about what is moving in a publication's own weekly [Knowledge Map](/public-knowledge-base.html): which topics and entities are rising, which are cooling, and which releases are behind the movement. An AI drafts them. Code checks that every sentence points at evidence. They land in the [Proposals](/proposals-and-editor-override.html) queue, where readers vote and an editor decides.

This is built and tested. It has run on [Shipped](https://shipped.anth.us/proposals/), our open-code demonstration site about machine-learning releases, where its example drafts sit in the Proposals queue. The weekly run is off by default and is switched on per site. Approving a draft does not yet turn it into a published story; that step is not built.

## What makes them different

A model is good at fluent prose and bad at staying inside its evidence. Cited Trends does not ask anyone to trust the prose. After the model drafts, a deterministic check rejects the draft unless:

- every sentence, the headline and the dek included, cites at least one signal or release it was given;
- at least two signals back the story as a whole;
- every number and date in a sentence appears in the signals and releases that sentence cites;
- the section exists on the site and the banned words are absent.

A draft that fails is not filed. The run record keeps it, with each problem and the sentence it came from. A draft that passes keeps its sentence-by-sentence citations. The check does not prove that a sentence reads its evidence correctly, so an editor reads every draft.

The chart is drawn by code from the real weekly counts of the cited signals. The model names which signals to chart and writes the caption; it never supplies a number for the chart.

## How a draft is made

1. Monitors collect items, and the shared relevance Cyclotron decides which publications want them.
2. The weekly knowledge run builds the Knowledge Map: Biblicus topic-model clusters, typed entities and relations such as "built on", an AI review, and a weekly snapshot with a history archive.
3. Code picks the rising and cooling signals from the snapshot and attaches the releases behind each one.
4. One GPT-6 Luna call drafts. In our runs so far it costs a few tenths of a cent.
5. The citation check runs, then a duplicate check against recent stories and earlier drafts.
6. The draft is filed as an AI story draft Proposal with its chart, and one soft [Review Request](/review-requests.html) asks a person to look. The weekly step runs after the knowledge run, so the drafts follow the week's map.

## Run it

```
papyrus ideas trends --site <site> --snapshot https://<site host>/_data/knowledge/ --store proposals --dry-run
```

`--dry-run` calls the model and runs the citation check, then prints each passing draft and each failing draft with its problems. It files nothing. The details are in the Papyrus repository, in `docs/idea-pipeline.md` under Trend stories.

See the [cookbook recipe](/cookbook/cited-trends/) for a walk-through with Shipped.