Back to the AppDeen ArchiveBlue Papers, research from the AppDeen AI lab

Blue Papers · BP-002

The Archive method: how a record is scored

What makes an event a record, how sources are weighed, and why a record is never rewritten.

No.BP-002
Rev.0
Date6 Nov 2026
AuthorsOsama Al-Hasan, Mr. AppDeen
StatusWorking draft
LabAppDeen AI, Amman
Drawing. The method as a machine, in plan and in section A–A. Four hoppers feed a funnel pulled every 15 minutes; ingest parses each item, the ledger keeps every item, the scorer reads four gauges (T tier, C corroboration, E engagement, N novelty) and sieves near-duplicates, and the gate passes only items at or over the threshold θ into numbered, permanent records. The weights and θ are clouded: they arrive with Rev. 1.

Abstract

The AppDeen Archive is a ranked record of what happens in AI. Nothing enters it unranked: every item pulled from the field is scored first, and only items above a threshold become records with a number and a permanent address. This paper describes the method as designed: the four layers the stream is pulled from, the four marks each item receives (source tier, corroboration, engagement and novelty), the line between "on the record" and "reported", and the rules that keep records permanent and corrections public. It is a working draft: the weights and the threshold are being calibrated and will be published with Rev. 1.

1The stream

The stream is built from four layers, pulled by machine every 15 minutes and scored before anything is shown:

LabsThe labs' own feeds and news pages: launches, research and policy, first-hand.
ResearcharXiv in cs.AI, cs.LG, cs.CL and cs.CV, and Hugging Face daily papers with their upvotes.
TrackersEpoch AI's models data, METR time horizons, Artificial Analysis and a funding tracker.
Wide netGoogle News queries, GDELT's coverage volume in 65 languages, and Hacker News.
Figure 1. The four layers of the stream. Each is pulled every 15 minutes; an item from any layer goes to scoring, never straight to the page.

Papers with Code, the old research tracker, shut down in July 2025 and now redirects to Hugging Face Papers, so Hugging Face is the research layer rather than a backup. Newsletters are the editorial benchmark, not a source: they set the bar for a five-minute read an engineer trusts.

2The four marks

Every item receives four marks before it can become a record.

Table 1. The marks an item receives
MarkWhat it measuresHow it is set
Source tierWho said itA lab or a tracker scores 3, major press 2, an aggregator 1
CorroborationWhether others say it tooIndependent sources within 24 hours; GDELT's coverage volume breaks ties
EngagementWhether the people who build careHugging Face upvotes and Hacker News points
NoveltyWhether it is newNear-duplicates are found by embedding and merged into one record

The marks are combined into one score, which every record carries. Items over the threshold become records with a number and a permanent address. Everything else stays in a searchable ledger: it is not shown as a record, but Walter can still answer from it.

Not yet published: the weight of each mark in the score and the threshold itself. They are being calibrated against the first weeks of the stream and will be stated, with the data used to set them, in Rev. 1 of this paper.

3On the record, or reported

Every record says where its central claim comes from. A record is on the record when its primary source is the lab, the company or the paper itself. It is reported when the claim comes from the press and the primary source has not confirmed it.

The same distinction runs through each record's sources, which are listed as primary or reported. When Walter answers about a record, every claim carries a record number or a source link, and his own reading is labelled as his, not as the record.

4What a record holds

A record can be read at three depths on one page. Glance: the headline, three lines and why it matters. Read: Walter's piece in plain words, 400 to 700 words. Deep: the sources, the timeline of its thread, and the numbers in a fixed box, each with its source.

Every record belongs to a thread (one model family, one lawsuit, one company), and every thread has a timeline, so a record about a known story opens with what changed rather than the background.

5Permanence and corrections

A record's number and address never change. Record 0412 is always at the same URL and always says what it said when it was filed.

When a record is wrong, the correction is filed as a new record, and the original lists it under Corrections with its date and number. Nothing is edited out of sight. A reader who cited the original can see exactly what changed, when, and why.

Record 0409 is filedNumbered, dated, sourced. Its address is permanent.
A fact in it proves wrongThe correction is filed as a new record with its own number.
0409 links to the correctionUnder Corrections, with the date and the new record's number. The text of 0409 is kept as filed.
Figure 2. How a correction is made. The record numbers are illustrative.

6What this draft does not settle

  • The weights of the four marks and the threshold, which Rev. 1 will publish with the data used to set them.
  • How the score treats languages other than English. Google News and GDELT reach Arabic coverage, but the marks have not yet been checked on it; BP-003 studies Arabic separately.
  • An outside reader. Every Blue Paper is meant to name one reader from outside the lab; this draft has not been read outside it yet.

References

  1. AppDeen Archive, API contract v0.1: the Card and Record shapes, including on_record, score, tier and corrections. Kept with the Archive's site code in archive/README.md.
  2. Hugging Face daily papers. huggingface.co/papers
  3. The GDELT Project, DOC 2.0 API. gdeltproject.org
  4. Hacker News search by Algolia. hn.algolia.com/api
  5. BP-001, The Tempo: measuring the speed of AI. appdeen.ai/archive/papers/bp-001

Cite this

Al-Hasan, O. and Mr. AppDeen (2026). The Archive method: how a record is scored. Blue Papers BP-002, working draft, Rev. 0. AppDeen AI, Amman. https://appdeen.ai/archive/papers/bp-002