In brief
- Independent, sourced content.
- Scoped LLM use: speed up formatting, not decide.
- Systematic human validation before publication.
- Facts and hypotheses are kept separate.
- Publication and revision dates visible on every article.
Why an explicit method
Research on language models moves fast. Claims circulate, get distorted, get repeated without sources. This site documents its editorial pipeline for three reasons: (1) traceability — every published claim can be traced back to its original source; (2) sourcing — an article is published only after cross-checking facts; (3) reliability — articles carry an explicit evidence level and can be revised as the literature evolves.
This page describes exactly how a raw source — a research paper, technical documentation, a practical experiment — becomes a published article, along with what is guaranteed and what is not.
The five-step editorial pipeline
Each article follows the five steps below, in this order. None is optional.
Step 1 — Collection
Sources come from four complementary channels: (1) academic sources — arXiv, Semantic Scholar, ACL Anthology; (2) automated monitoring via Python scripts watching repositories and research-centre feeds; (3) GitHub signals — releases and significant commits on reference projects; (4) manually curated bookmarks drawn from technical reading and conferences. Each source is stored in its raw form, timestamped, and linked to the topic it documents.
Step 2 — Extraction
Structured extraction scripts isolate verifiable assertions in raw sources: facts, figures, experimental results, exact quotes. Extraction explicitly separates what is factual from what is interpretation. Only factual elements enter the rest of the pipeline as publishable material; interpretations may be cited, but are flagged as such, with their author.
Step 3 — LLM processing
A language model analyses, structures and rewrites the extracted content. Its role is explicit: speed up formatting and the connection between sources, not decide on content. The model proposes a candidate draft from the extracted factual elements; it does not choose what gets published. This step is a writing tool, not an editorial decision-maker.
Step 4 — Human gate
Every published assertion is manually validated before release. The model proposes, the author decides. No content is auto-published. At this step, contradictions between sources are arbitrated, the evidence level of each fact is set (see below), passages based on a single source are explicitly flagged in the text, and wording is checked against the original source to prevent any drift in meaning.
Step 5 — Publication
The article is dated, versioned, and published with its links to the primary sources cited. Each article carries a publication date and a last revision date. An update is published explicitly if a source is modified by its author, if a later publication contradicts an established fact, or if the literature evolves significantly on the topic. Revisions are timestamped and the article version is incremented.
Evidence levels
Each published fact carries an explicit evidence level, from weakest to strongest:
- Speculative — unconfirmed hypothesis. Not published publicly; mentioned only when explicitly flagged as speculation.
- Single source — claim from one source only, without cross-confirmation. Explicitly marked as unverified in the text.
- Replicated — confirmed by at least two independent sources, or by an experiment whose raw data is preserved and accessible.
- Triangulated — accepted by the research community, with no active controversy; cited by sources of different kinds (papers, documentation, field feedback).
A public article relies primarily on facts at replicated or triangulated level. Passages based on a single source are explicitly flagged in the body of the text.
What this site does not do
- No unsourced content. Every published claim is linked to an identifiable source. No unsourced claim is published publicly.
- No paraphrase without attribution. Reformulating an idea does not exempt from citing its author. Non-trivial ideas are explicitly attributed.
- No promotional content. Articles are not used to promote a product, model, or player. Strengths and limitations are presented side by side.
- No claim to exhaustiveness. This site covers what is read and verified, not the entirety of the literature. An absent topic is not absent by editorial choice, but by time constraints.
- No guarantee of zero errors. Residual errors exist. They are corrected through the revision mechanism, and any significant correction is dated.
Report an error or a missing source
Does an article contain a factual error, a missing source, or a claim you can support with a better source? Reports are welcome. Please include: the article URL, the relevant passage (exact quote or section title), and the source that contradicts or clarifies the claim. The correction is published with an explicit mention in the article's revision history.
What this method does not guarantee
- Full exhaustiveness on a fast-moving field.
- Absence of any error.
- Stable conclusions — revisions remain possible as the literature evolves.