3GPP specifications
TS and TR documents from Release 15 to Release 20, read from the 3GPP file server archive. Pages cite the spec, version and clause.
3GPP Scout / Methodology
Our public pages are built from 3GPP records by code and language models. Each one must pass automated checks before it is published. This page explains the sources, the checks, when a page stays out of search, and how to report an error.
At a glance
3GPP Scout is independent and is not affiliated with 3GPP or ETSI. The official documents govern.
TS and TR documents from Release 15 to Release 20, read from the 3GPP file server archive. Pages cite the spec, version and clause.
Contributions, chair notes, session reports and meeting reports come from the meeting folders on the 3GPP file server. Titles, sources, statuses and agenda items come from the 3GPP portal's TDoc registers, and dates from its meeting calendar.
A register status such as agreed, approved, endorsed or noted is shown with its exact label.
Proposals and observations are taken from each TDoc's own text. A position is credited only to the companies the register lists as its submitters.
The OCUDU page reads the ocudu.org news feed and the Atom feeds of the OCUDU projects on GitLab. From GitLab it keeps only titles, dates, links and our own summary.
No page is built from anyone's chats, followed topics or usage of Scout. We never read paywalled pages or log in to collect anything.
Our fetcher names itself, follows robots.txt and never gets around a block. Source rules: copyright and takedown.
In every page type, the model returns short prose and the IDs of the evidence it used. Code then places each quote, TDoc link, company list and status from that evidence. A model cannot add a quote or a TDoc of its own.
One page per meeting of RAN1 to RAN4, SA2, SA3, CT1 and CT4. A preview goes up 7 days before the start, a live stage adds "Agreed so far" from chair notes, and a final brief follows the close.
The preview lists the hottest topics, contested proposals already uploaded and what the previous meeting left open. The final brief covers the meeting's top 3 topics: what was agreed, who proposed what, and what is still open. An OpenAI GPT model drafts the prose. Decisions from a feature-lead summary count only when the same text is in that meeting's chair notes, and outcomes of offline discussions are labeled as such.
About 200 words on a spec change between versions, a company's changed position, a meeting outcome or a contested proposal. Only meetings that closed at least 3 days earlier are used.
The draft is a claim map: every sentence names the evidence it rests on. A sentence may carry one quote, which must match its evidence word for word. A contested proposal is skipped when no recorded decision verifies.
Built from the cause and timer tables of TS 24.501 and TS 24.301 and their annex descriptions, in the newest version Scout has indexed. Each page quotes the spec and adds one plain-words line.
A GPT model writes the plain-words line. Claude Opus from Anthropic then checks it against the quotes. Values the spec words the same way share one page, with an anchor for each value. Rendering a page makes no model call.
For each OCUDU release and merged feature, a model names the 3GPP features involved and Scout's spec search returns candidate clauses. A clause is kept only if it contains the feature's key term.
When no clause passes, the page says "no specific 3GPP clause found". Features that exist only in O-RAN are labeled as having no 3GPP clause. A clause match is never presented as proof that OCUDU implements or conforms to that clause.
Generated by code from our gate report, with no model writing the page. Each gate shows its result, 95% interval, threshold and the date it was measured.
Chat answers in the benchmark are judged by Claude Opus against written rubrics, and the majority of three runs counts.
Spec and glossary pages are one paragraph with a facts box from the database. Meeting briefs are topic summaries built like final meeting pages. Newsroom briefings are written one at a time, and one the editor edited says so on the page.
Every check runs before publication. A page that fails any of them is held, and its last passing version stays live. Every attempt, passed or held, is recorded with its results.
Each quote must appear in the source text when evidence is gathered and again when the page is rendered. A live audit then fetches every cited TDoc again and checks the quote and the submitting companies.
Caps: 25 words per TDoc excerpt and one per TDoc per section. 50 words per spec excerpt and 150 from one spec per page.
On reference pages, every quote must be in the stored clause and in the 3GPP archive file for the cited version. The check fails when Scout has indexed a newer version than the page cites.
The reviewer is always from a different model family than the writer. On meeting and update pages, Claude Opus reviews as a skeptical delegate for up to 3 rounds.
A meeting page passes only on "ship", or "ship after fixes" with no high-severity issue.
A claim must keep the scope and qualifiers its source states: one work item rather than all of it, a working assumption, or a point still for further study (FFS).
A proposal is never shown as an agreement.
Rebuilt offline from its stored draft and evidence, the page must come out the same. So everything on it traces to that record.
An update page's own prose must differ from other update pages and from its meeting page: under 0.3 overlap of three-word phrases.
No semicolons in our analysis and no sentence over about 30 words. Contact addresses are only at 3gppscout.com.
The site runs the copy rules again on every change.
A page marked noindex is not in our sitemaps. Search engines are asked not to list it.
Meeting pages and the OCUDU page need at least 3 cited decisions or positions and 300 words of our own analysis. A meeting page below that is published as noindex.
If a thin rebuild would replace a page that is already indexed, the rebuild is held instead.
Reference, spec, glossary and RRC pages start as noindex. Up to 50 open to search first, then 1.5 times more every 2 weeks while at least half of the last batch is indexed.
Only pages that passed every check and their type's audit join a batch.
The OCUDU page stayed noindex until two weekly editions had passed every check.
Full-length briefs and PDFs are noindex, so search lands on the short version. A page under dispute goes to noindex, or to metadata only, first.
| Page type | When it updates |
|---|---|
| Meeting pages | Checked twice a day, at 05:30 and 17:30 UTC. Preview from 7 days before the start, live during the meeting, final after the close. |
| Update pages | Daily at 09:00 UTC, up to 6 new pages. |
| OCUDU page | Weekly, on Tuesdays at 16:00 UTC. |
| Reference pages | Rebuilt by hand. The checks flag a page once Scout has indexed a newer version of its spec. |
| Benchmark | Rerun every month and after changes to how Scout answers. |
Every page shows its dates: when it was published, last updated or last checked.
That covers meeting, update, reference and OCUDU pages. The checks above are what stands between a draft and the site.
Each Monday the editor gets a review packet for the meeting pages: what was published, audit and review results, held pages, and two pages suggested for a full read.
A factual error found in review pauses that page type until it is fixed.
Email [email protected] with the page address and what is wrong. A TDoc number or quote helps. We aim to acknowledge within 2 business days and to act within 5.
We correct the stored draft or the code that built the page, then run the same checks again before republishing.
| Date | Page | What changed |
|---|---|---|
| 2026-11-06 | RAN2 #135-bis | At a glance and agenda item 8.14 said TEI20 adds a new coverage assistance SIB with no connected-mode enhancements. The record says one TEI20 item does. Both now say "one TEI20 item".Source: R2-2607784, RAN2 #135-bis. |
This log lists errors found on published pages. Problems the checks catch before publication are not listed.