What we do, why, and how
The framework is by Dimitrios Thomakos: Ars Prognostica — The Origins, Evolution, and Formalization of Prediction (2026) and Intelligence Without Wisdom — Practical Judgment and the Epistemology of AI Risk Claims (2026), both published at the Cybernetics and Artificial Intelligence Laboratory of the National and Kapodistrian University of Athens. The Oracle Hunters applies it; any error in the application is ours. Each step below says what we do, why, and how - once.
Also from the laboratory: Symbiosis & AI Risk - Dimitrios Thomakos, CAIL, 8 March 2026.
Understanding the Oracle
Twelve questions on Ars Prognostica and Intelligence Without Wisdom, each with an interactive answer: the nine Greek words, the three modes, Bayes, entropy, the oracle's ambivalence, deception versus repetition, quantum decision theory, the phronesis score, the four tiers of AI-risk claims, the six speech acts, the POMDP and its quantum generalisation. Everything runs in your browser.
Open Understanding the Oracle →1. Collecting
What: speeches and statements from official, primary sources only (live list): central banks' own archives (ECB since 1997, Federal Reserve since 2006, every central bank's speeches via the BIS since 1996), governments (GOV.UK, the European Commission, the White House, the Greek Prime Minister's office and Ministry of Finance, foreign ministries), international bodies and AI companies. Why: a quote from a news report can be shortened or paraphrased; the official text cannot be disputed by the speaker. How: each document keeps the time it was said, the time we captured it (UTC) and a fingerprint of its text. A speech that reaches us through two official channels is stored once.
2. Choosing sentences
What: up to 6 sentences per document that speak about the future. Why: a speech is mostly description; the record is about claims that can turn out right or wrong. How: rules look for future tense, expectations, projections, deadlines and likelihoods; procedural sentences (“I will now turn to…”) are dropped. Statements in Spanish, Turkish, Portuguese, German, French, Russian, Chinese and some thirty other languages are translated offline into English (Argos, on our own computer - nothing is sent anywhere) and then go through the same steps; every translated statement says so and links to the original words.
3. Classifying
Kind of claim (six speech acts): forecast, scenario, speculation, commitment, call for action, or mixed. What it needs as evidence (four tiers): possibility, mechanism, forecast, policy. How the speaker claims to know (three axes): based on data (empirical), educated guess (abductive), stated on authority (oracular). Why: a scenario is not a forecast and a forecast is not a mandate to act - treating them alike is the commonest mistake in public debate. How: published rules (rules-2.0, in method.json); the reasons for each label are shown on every statement's evidence page.
4. Can it be checked - and how clearly?
Testability: four yes/no questions - is the event named, is there a firm deadline, is a likelihood given, is there a measurable figure? Four of four: can be checked. Two or three: partly. Vague time frames (“over time”, “in the medium term”) do not count as deadlines. Ambiguity (0-100): a practical version of the ambivalence function of Ars Prognostica (the entropy over the readings a statement allows): readings grow with hedging words (“could”, “might”, “maybe”) and with each missing element; ambiguity = log₂(readings), scaled to 0-100, published so anyone can recompute it. The public ledger holds every AI-risk statement and every forecast, commitment or scenario with at least two of the four elements; vaguer remarks stay in the archive. A statement that is fully testable and free of hedging is marked meets the standard - the same standard ATHENA's Forecast Desk holds its own calls to (one shared method, RIM-1); anything less is marked below.
5. Words to numbers
An explicit “x% chance” is used as stated. Otherwise the words are read with this fixed table. Each wording gives one number - the likelihood that the thing happens - and the call it implies: Yes above 50%, No below. How firm the call is gets its classical name, shown in Greek and English here and in English on the rest of the site: eikos - probable; hōs epi to poly - for the most part; anankaion - as good as necessary. For a No, firmness is read from the other side: 30% likely is a No held 70% of the time.
| Words | Implied call / likelihood | Classical term |
|---|---|---|
| “virtually certain” | Yes · 99% | anankaion - as good as necessary |
| “almost certainly” | Yes · 95% | anankaion - as good as necessary |
| “highly likely” | Yes · 90% | hōs epi to poly - for the most part |
| “very likely” | Yes · 90% | hōs epi to poly - for the most part |
| “highly unlikely” | No · 10% | hōs epi to poly - for the most part |
| “very unlikely” | No · 10% | hōs epi to poly - for the most part |
| “most likely” | Yes · 80% | hōs epi to poly - for the most part |
| “more likely than not” | Yes · 60% | eikos - probable |
| “unlikely” | No · 25% | hōs epi to poly - for the most part |
| “likely” | Yes · 70% | eikos - probable |
| “probably” | Yes · 70% | eikos - probable |
| “probable” | Yes · 70% | eikos - probable |
| “expected” | Yes · 70% | eikos - probable |
| “expects” | Yes · 70% | eikos - probable |
| “expect” | Yes · 70% | eikos - probable |
| “anticipate” | Yes · 70% | eikos - probable |
| “projected” | Yes · 70% | eikos - probable |
| “forecast” | Yes · 70% | eikos - probable |
| “going to” | Yes · 80% | hōs epi to poly - for the most part |
| “will” | Yes · 85% | hōs epi to poly - for the most part |
| “shall” | Yes · 85% | hōs epi to poly - for the most part |
| “may” | No · 40% | eikos - probable |
| “might” | No · 35% | eikos - probable |
| “could” | No · 35% | eikos - probable |
| “possibly” | No · 35% | eikos - probable |
| “possible” | No · 35% | eikos - probable |
| “remote” | No · 5% | anankaion - as good as necessary |
| “αναμένεται” | Yes · 70% | eikos - probable |
| “προβλέπεται” | Yes · 70% | eikos - probable |
| “πιθανόν” | Yes · 60% | eikos - probable |
| “πιθανώς” | Yes · 60% | eikos - probable |
| “θα” | Yes · 85% | hōs epi to poly - for the most part |
6. The nine words - the name for each way of speaking about the future
Where English has one word, “prediction”, the Greeks had a family of words, each naming a different way of knowing the future. Every statement is rated with the word that fits it, from the labels of steps 3-4. What we use: the vocabulary only - the mathematics in the monograph and in the companion explains the argument; it is not used to rate statements.
| Word | Meaning | Mode | When a statement is rated so |
|---|---|---|---|
| μαντεία manteia | Inspired knowledge | theocentric | A forecast resting on the speaker's authority or vision alone, with no data or reasoning offered. |
| χρησμοδοσία chresmodosia | Institutional delivery | theocentric | The institution itself, speaking through its official procedure (no named speaker), announcing what it will do. |
| προφητεία propheteia | Interpreting the will of authority | theocentric | An official interpreting what a leader or higher authority intends (“as the President has made clear…”). |
| πρόβλεψις problepsis | Seeing the signs | empirical | A claim built on signs already visible - early indicators, first signals, something “beginning to” happen. |
| πρόγνωσις prognōsis | Inference from evidence | empirical | A forecast inferred from data - figures, models, trends - the kind that can be checked against official numbers. |
| πρόνοια pronoia | Preparing for the future | empirical | Preparation rather than prediction: a scenario (“if… then”) or a call to build, plan or protect against what may come. |
| εἰκασία eikasia | Conjecture | speculative | An educated guess - a hypothesis about what may happen, offered without a firm basis in data. |
| προαναγγελία proanangelia | Public pre-announcement | speculative | A commitment announced in advance (“we will…”), meant to steer what others expect and do. |
| φρόνησις phronēsis | Practical wisdom | meta | A judgement that weighs conditions and trade-offs - data-dependent, balancing risks, ready to adjust. |
| προεξόφλησις proexoflēsis | Premature certainty | failure | The failure the vocabulary warns about: a guess or an assertion stated as settled fact, in a form that cannot be checked. Shown as a warning beside the word, never instead of it. |
Order of the rules: a statement that weighs conditions and risks is phronesis; one that interprets a leader's will is propheteia; a forecast resting on visible signs is problepsis; a commitment is chresmodosia when the institution speaks with no named speaker, otherwise proanangelia; a scenario or a call to prepare is pronoia; an educated guess is eikasia; a forecast resting on authority alone is manteia; a forecast inferred from data is prognosis. The word judges the sentence, never the person; every evidence page says which rule applied.
7. Checking the outcome (rules version 4, automatic)
Institutional forecasts (numbers with dates) are compared with the official outcomes on the scoreboard. Statements by people are settled when the deadline passes: the program looks up the official figure (the sources are listed on the Trust page) and applies these fixed rules without a human step. The rules prefer saying nothing to saying something wrong about a named person.
- Headline measures, each against its own official series: consumer-price inflation, real GDP growth and the unemployment rate as annual figures (World Bank); inflation at the end of a year, December on December (Eurostat for European countries, the US Bureau of Labor Statistics for the United States, the Bank for International Settlements for about 60 other countries); the government deficit and debt as a share of GDP (Eurostat, European countries); and central-bank policy rates at the end of a year (Bank for International Settlements, about 40 central banks). Core inflation, potential growth, GNP or GNI, wages, credit, structural or primary balances, company revenue, a quarter, a month or a fiscal year are not settled this way - no official figure would be a fair test. A claim about a place that has no official series stays unchecked.
- Only the speaker's own forecast: sentences that quote a survey, analysts, markets, the IMF, the OECD, the World Bank or a news agency are left out.
- The place named in the sentence: one country or the euro area is checked as named; none named means the speaker's own country; several, the world or a region means it is left out (world growth is quoted on different weights by the IMF and the World Bank, so neither figure is a fair test of the other).
- Each number with its own year: “1.9% this year and 2.0% next year” are two claims; the one that falls due is checked. A range is checked as a range.
- Verdicts: within 0.5 percentage points - on target; within 1.0 - close; further - missed. “Below” and “above” are checked as stated. “Will rise / fall” is checked against the change from the year before; a change smaller than 0.1 points counts as close.
- Published with the label “automatic check against official data”, the figure, its source and the arithmetic. A correction can be requested on the corrections page; every change is logged.
8. Scoring
Report cards show plain counts. Calibration uses the Brier score - the mean of (likelihood − outcome)², a proper scoring rule, so it rewards honest likelihoods rather than confident-sounding ones - and appears only after 10 checked forecasts. Speakers and ATHENA's Forecast Desk are scored by the same rules, so their records can be compared directly.
9. Rules we keep
Primary sources only · the same tests for optimists and pessimists, every party, every country · the original words always next to their label · labels judge the claim, never the person · every change logged · a right of reply · short excerpts with a link to the full source · daily seals so the record cannot be quietly rewritten.
10. Known limits
- The rules are written for English and Greek; other languages pass through an offline machine translation first, which can soften or sharpen a sentence - the original words are always one click away.
- An official figure can differ in definition from the one the speaker had in mind (national CPI versus HICP; the European Union's membership changed over time). The figure used is always shown.
- Ministries without a named speaker are scored as institutions.
- History before our first capture is sealed on the day we read it, so for old speeches the seal proves what the official text said when we read it.
11. Where each part comes from
| From the work of Dimitrios Thomakos | Our implementation |
|---|---|
| The kinds of claim (forecast, scenario, conjecture, strategic announcement, policy prescription, combinations); the three axes (empirical, abductive, oracular); testability and ambiguity as measures; the six questions for AI-risk claims; the nine words (Ars Prognostica; Intelligence Without Wisdom). | The rules that recognise each kind and each word; the 0-4 testability score; the 0-100 ambiguity scale; the table turning likelihood words into numbers; which statements reach the public ledger; the checking rules and their margins; the scoreboard, the Brier score display, changes of direction, seals and everything on this website. |