Oracle

Why it comes last

Prediction was deliberately not the first thing built. A forecast is only as good as the observation beneath it, and an unverifiable forecast is worse than none — it converts a guess into something that looks like knowledge, which is harder to argue with and easier to act on.

So the order was: collect and attribute the observations, make the dependencies auditable, build a harness that can tell whether a signal actually preceded anything, and only then start forecasting. That harness, the corridor-signal backtests on /record, is a separate and older ledger: 24 scored runs, 0 of which established skill. It is not the Oracle's record, which is section 04 below.

What it will do

• Specific, checkable outcomes. A corridor's traffic falling below a threshold inside a stated window — not a vague sense of elevated risk.
• Recorded before the fact. Every forecast is written down before its outcome is known, with the time it was issued; one issued after its window opened is marked late. It is scored once the outcome can be read, and never quietly revised afterwards.
• Willing to say nothing. Where the evidence is too thin, it declines to forecast rather than producing a confident-looking number nobody can check.

What it inherits

Oracle does not start from scratch. The signal miner, the point-in-time replay, the ground-truth ledger and the scoring harness are all live today, and every one of them was built so that a forecast could later be checked against them without rewriting history.

The bar it has to clear

Until then Oracle is a pre-release. Everything in the Terminal marked as model output is labelled as such, and nothing there is presented as a forecast. The families read each other's state through the dependency graph; on the evidence so far that adds no skill, and the links are kept to explain a forecast, not to improve it.

The accuracy record so far

Oracle 0.19.0 is a pre-release. Its record covers fifteen questions, about shipping lanes, ports, rivers, power grids, oil supply, natural hazards, markets, commodities and currencies; two of them are controls. The production run asks every one of them each day, and the record shows no question it does not ask. Each question was scored on years its models had not seen, with every input read as late as it really arrives on that day of the week, and each year's alert level was chosen from the earlier years alone.

Each question is judged against two rules anyone could follow without a model, each given the same number of alerts: the usual places (alert where the event has been most common) and tomorrow like today (the question's own persistence score). A question holds up when it beats both. Three of them do.

A new event is one that began while the place was not already in that state; alerted ahead means an alert came before it began. The two naive rules are shown as the share of their alerts that came true and the share of new events they alerted ahead. Conflict is not on the record: no production question asks it since its model input was withdrawn on 10 October. The controls ask what no public data should answer: when an earthquake will strike, and which way a price will move. The earthquake control ranks seismic countries above quiet ones, as a map would, and shows no skill at the timing it is asked about. Under the stricter tests registered on 8 October and corrected on 9 October, the price-direction control's small margin over a coin is a movement shared by each group of assets on a day, which a model of that alone matches. On labels shifted by whole years, the engine beat the registered comparison for ports and for price direction. A further test, registered before it ran on 10 October, gave the comparison everything a shift by whole years keeps; against it the engine showed no skill on shifted labels in any of 26 runs. Improvements measured on this engine are acted on again, and each is still judged on the forward record.

The record above is version 0.19. It adds three questions: whether a strait running normally collapses within 14 days, whether a collapsed strait recovers, and whether a serious cyclone alert names a country within 30 days. It reads three market series (the VIX, the Nasdaq and the 10-year breakeven) two days behind, as their publisher posts them, where earlier versions read them a day early. Every question was scored again on the same inputs as 0.18; the seven these changes do not touch came out the same to the last digit. Version 0.18 corrected inputs earlier versions read before they were published: the river questions had read a model reanalysis weeks before it existed, the port and strait lists were chosen with later years in view, and some outcomes were learned from before they were known. The earlier record overstated those questions.

The forecasts of 25 September 2026 were committed to a dated log before their outcomes were known. That log holds one day: the daily run depended on one desktop session, which stopped running it after that day. Forecasts are now issued by a scheduled production run that does not depend on any one machine, and each is kept in an append-only record from the moment it is published, resolved when its outcome can be read and scored beside its baseline. When this page was built, on 11 October 2026, the production run had published 3,019 forecasts, abstentions counted, for fifteen questions, the newest issued on 11 October 2026; 0 had resolved and 0 had been scored. The forward record is the test this backtest cannot replace, and it is the one Oracle 1.0 will be judged on. Section 05 reads it live.

The forward record, read live

Each question is a contract: what is counted, over which window, against which baseline and from which series are fixed before a forecast is issued, and a changed definition is a new contract. Every contract has a standing. Research means its forecasts are issued, kept, resolved and scored in this record and are never presented as forecasts with skill. A contract is supported only after two gates: one that it runs without a person and covers its places, and one that it beats its baseline on past years with an interval that excludes zero. It is validated only once its forward record does the same over at least sixty resolved forward issue dates and, for a question about an onset, over enough forward onsets to judge. No contract had passed a promotion gate when this page was built, so every forecast in this record is research.

Forecasts counts each published forecast once, abstentions included: its first published revision, so a later recomputation can never improve the record. Resolved counts forecasts whose outcome has been read from the series the contract names; scored, those given a Brier score. Gate C counts the issue dates of the contract's newest release on which every forecast that was not an abstention has its outcome, at least one of them resolved. A place in brackets is the kind of place the question is asked of; the API names each one.

The markets question reads the VIX, an index of Cboe Global Markets, through FRED (Federal Reserve Bank of St. Louis); FRED lists it as copyrighted, citation required. No VIX value is republished here, only the question's forecasts.