Research methodology
How we decide what we are allowed to say.
Our standards for evidence labels, source selection, verification and corrections. Every HUTE publication is held to them.
HUTE publishes research and explanatory writing about a field where claims are cheap and evidence is hard. This page states the rules we follow so that readers can judge our work by them.
1. Every publication has an evidence status
Each research report carries one of these labels at the top, with a note about what it does and does not contain.
| Label | Meaning |
|---|---|
| Literature review | A synthesis of published work. It contains no new experiments. Findings are attributed to their sources. |
| Conceptual framework | A proposed way to organise or measure something. It is argued from the literature and has not been experimentally validated. |
| Experimental result | New data collected and analysed by HUTE, with methods, sample, limitations and (where permitted) materials. None is published yet. |
| Editorial review | A draft that is not yet verified to our standard and is therefore not published. |
Articles in Insights are explanatory and cite their sources, but they are not reports of new experiments.
Our first three reports are a literature review (Report 001), a conceptual framework (Report 002) and a literature review with analysis (Report 003). None presents original experimental results, and none should be read that way.
2. How we select sources
- Peer-reviewed and archival first. Journals, conference proceedings and recognised preprint archives such as arXiv. Preprints are labelled as preprints and treated with the caution that implies.
- Primary over secondary. We cite the study, not a news report about it.
- Authoritative legal texts (for example EU regulations) are cited at their official source.
- Breadth matters. Where a claim is contested, we cite work on more than one side. For example, theory-of-mind results in language models are mixed, and we cite both the positive and the sceptical studies.
- Dated. The field moves quickly. Each publication states when it was published and last updated.
3. Every reference is machine-verified
Before a source can be cited on this site, a script checks it against its registry: Crossref for DOIs, arXiv for preprints, and the official page for legal texts. The published bibliography is generated from that registry data (title, authors, year), not typed from memory. If a source cannot be verified, the build fails and the page does not publish. The verification record is dated in our repository.
Verification confirms that a source exists and is correctly identified. It does not by itself prove that we have summarised it correctly, so we also read the abstract and, for central claims, the paper, and we keep claims close to what the authors report.
4. Established findings, interpretation and proposal are kept apart
In reports, sentences that report what a study found are attributed to it. Our own interpretation is signalled as interpretation. Proposals are labelled as proposals. We avoid numerical claims we cannot trace to a source, and we never present a proposed evaluation method as if it had been run.
5. What we do not do
- We do not fabricate citations, results, participants, customers, partnerships or metrics.
- We do not use model-generated text as evidence.
- We do not treat benchmark scores as proof of understanding; see Research Report 002.
- We do not publish a claim as established because it is widely repeated.
6. Limitations of our own reviews
Our literature reviews are not systematic reviews in the clinical sense: we do not search every database with a registered protocol. They are structured, source-verified syntheses. Selection of studies reflects our judgement and the date of writing, and readers should treat them as maps of a field, not as complete inventories.
7. Corrections
If you find an error, an out-of-date statement or a source we have misread, tell us. We correct the text, update the "last updated" date, and add a note describing what changed when the change is material.
8. Independence and conflicts of interest
HUTE builds products and infrastructure that depend on the research questions we study. That is a conflict of interest, and we state it. It is why we publish our standards, cite work that complicates our own view, and label proposals as proposals.