--- title: "GEDmatch Oracle explained: what the distances mean and how to read Oracle-4" description: "Oracle ranks reference populations by how closely their calculator percentages match yours — not by shared DNA. How the distance is computed, what single and mixed modes tell you, and the misreadings to avoid." canonical: https://www.ancestrify.io/blog/gedmatch-oracle-explained date: 2026-08-31T08:28:00+00:00 author: "Andi Thomaj" --- # GEDmatch Oracle explained: what the distances mean and how to read Oracle-4 Run any GEDmatch admixture calculator and the percentages arrive with a button marked Oracle — and sometimes Oracle-4. Most people click it, see a ranked list of population names with numbers like `3.24`, and take away either too little ("Finnish?? I'm Irish") or far too much ("I'm basically Tuscan"). The Oracle deserves better: it is the most interpretable output a GEDmatch run produces, once you know what the number is. Prerequisite reading if the percentages themselves are unclear: [what an admixture calculator is](/blog/what-is-an-admixture-calculator) and the [map of the GEDmatch projects](/blog/gedmatch-admixture-calculators-explained). ## What Oracle actually computes Oracle does **not** compare your DNA against anyone. It compares your *calculator percentages* against the stored percentages of the project's reference populations, and ranks the references by distance — smaller is closer. If [Eurogenes K13](/blog/eurogenes-k13-explained) gave you 41% North Atlantic and 18% West Med, Oracle scores every reference population by how far its own thirteen numbers sit from yours, and sorts. That indirection matters. Two people can land near the same reference for different reasons; a population can rank first because it is genuinely similar to you or because it happens to be a *mixture that averages out* near your profile. And the whole ranking inherits every property of the calculator upstream — its components, its [training biases](/blog/why-admixture-calculators-disagree), its 2012-era references. Oracle on a poor calculator is a precise ranking of the wrong thing. The utility descends from the original Oracle written for the [Dodecad project](/blog/dodecad-k12b-explained); each GEDmatch project carries its own reference list, which is why the same kit gets different Oracle answers in different projects. ## Single mode: the ranked list The default output ranks single reference populations. Reading rules: - **Read neighbourhoods, not winners.** The gap between rank 1 and rank 5 is often smaller than the noise in your own percentages. The first *cluster* of entries — usually a coherent region — is the signal; the exact winner is not. - **The distance scale is calculator-specific.** A distance of 3 in one project and 3 in another are not comparable, and no threshold ("under 5 is good") transfers across calculators. Within one run, only the *relative* spacing carries information. - **Expect mixed-population artefacts.** Profiles from genuinely admixed people often rank references that match the *average* — someone half northern, half southern European can see central European populations at the top that match neither parent. That is arithmetic, not ancestry. ## Oracle-4: the mixed fits Oracle-4 extends the search to combinations — pairs, triples and quadruples of references, with weights — and returns the best-fitting mixtures at each complexity. It was designed for exactly the case single mode fumbles: recent admixture, grandparents from different regions. The reading rules sharpen accordingly. A four-way fit has more free parameters, so it will *always* fit better than a single reference — a smaller Oracle-4 distance is not evidence that the four-way story is true. Combinations of closely related references trade off almost freely (Irish + West Scottish versus Cornish + Orcadian is not a distinction your percentages can make). And nothing tests the fit: like every tool in this family, Oracle [cannot reject anything](/blog/qpadm-vs-admixture-software#is-qpadm-worth-it). It reports the best available arrangement of what it was given, full stop. ## The spreadsheet behind it The same results page usually offers a spreadsheet: the full table of every reference population's component percentages. It is the single most underused artefact on GEDmatch — it is where you learn what a component *means in that project's own terms* (which references are high in "East Med"; what "North Atlantic" is anchored to). Ten minutes with the spreadsheet prevents most of the classic misreadings of both the percentages and the Oracle. ## The modern equivalent of the question Oracle's question — *which reference populations sit closest to me?* — is a distance question, and it now has a direct answer that skips the component detour entirely: Euclidean distance on [Global25 coordinates](/blog/what-are-global25-coordinates) against dated reference panels. The free [G25 distance tool](/lab/g25-distance) ranks 1,535 curated populations era by era in your browser; the [PCA viewer](/lab/g25-pca) shows the neighbourhood; the [admixture calculator](/lab/admixture) does what Oracle-4 does, with a stated fit distance and your choice of curated panels. Same question, current references, no 2012 components in between. The habits transfer: read neighbourhoods rather than winners, treat mixture fits as descriptions rather than findings, and when a ranking is about to become a claim about descent, take it to a [method that can reject a model](/qpadm) rather than one that can only rank. Terms used here are defined in the [glossary](/glossary).