How Doxa is made
What comes from published research, what was generated, and what is neither. Read this before you put weight on anything here.
61
biases
610
prevention strategies
153
inline references
31
curated papers
The prevention strategies
Every bias in the library carries exactly ten prevention strategies — 610 in total across 61 biases. That uniformity is the tell: they were generated by a large language model, prompted from the debiasing literature to produce a fixed count of ten per bias. It is not the shape curation would have produced. The first set was then read and edited by hand, and anything wrong or useless was cut or rewritten. Strategies added or rewritten since were reviewed by a language model, not by hand.
They are not individually citation-backed. No single strategy traces to a specific paper, and none of them is a research finding. Treat them as prompts for your own thinking, not as evidence.
The references
There are two separate layers here, and one does not imply the other.
Inline references, on every bias page
All 61 biases carry two or three citations to the primary literature — 153 in total. These came from the same place the strategies did: a language model proposed them. What differs is what happened next — every one was resolved against the Crossref API and checked that the title and the year matched the paper it claimed to be, and the handful that did not match were corrected by hand or dropped. That verification is the whole reason these can be trusted where the strategies cannot. They are real, and you can check them yourself: each citation links out from the bias page.
Curated papers, on the Science page
The Science page holds a separate set of 31 papers with bilingual abstracts, linked to 30 of the 61 biases. So 31 biases have inline citations but no curated paper. A bias page with no "Scientific sources" section is not a bias without literature — it is a bias this second layer has not reached yet.
The AI-guided path
"Help me think it through" sends your text to a general-purpose language model, together with a prompt. The model is hosted by Groq; if Groq cannot answer, the request may be routed through OpenRouter to a fallback model hosted by Google. Untilt stores none of it, and each provider's own free-tier terms apply to what it receives. It is not a model trained or fine-tuned on neuroscience, and it knows nothing about you beyond what you just typed.
What it cannot do: invent a bias. The model answers with a bias identifier, and any identifier that is not already one of the 61 in the library is dropped before it reaches you.
What it can do: get the "Why this may apply" sentence wrong. That sentence is free text the model writes, and nothing checks it — not against the literature, not against your situation. Read it as something to verify, never as a verdict.
If something here is wrong
Doxa is a free side project, not a clinical or research instrument. If you find a broken citation or a claim that reaches further than the evidence, say so and it gets fixed.
Report a problem