01
Why is it like this
Why a country's inflation, growth or jobs sit where they do: the engine's sum of named terms for that year, then what the network trained on every country's figures says, split exactly into what each figure adds.
The analyst
Ask anything. See every step it took.
A question in ordinary words goes in. What comes back is the answer, the plan it made, every figure it used as a numbered row of evidence, where inside the instrument each row came from, how sure it is and why, and what would make it surer. Nothing is asserted that a row does not carry.
One question, worked through
The working, in the order it was done
Every answer is built the same way, and every stage of it stays on the page.
-
01 / 05
It reads the question
It finds the countries, blocs and regions named, the figures meant (inflation, jobs, debt, stability and some forty more, in the words people use for them), the years, whether it is about today or the years ahead, whether it is about your organisation, and whether a scenario sits inside it. A question in several parts is split and answered part by part, in order, so a what if asked first is the run every later part reads.
- Overview, lookup, comparison, ranking, explanation, relation
- Forecast, what if, risk, today, sources, reliability
- Your company, and what to upload next
-
02 / 05
It makes a plan
Each kind of question has its operators, and the plan is the list of them in the order they run. An explanation opens the engine's own arithmetic first, then the pattern across countries, then what the simulation leans on. A what if weighs the ways the sentence could go, runs it against the baseline on the same seed, and reads the result.
- Every step is listed with the rows it produced
- A question it cannot place is said to be one, not guessed at
-
03 / 05
It gathers evidence, with a source for every row
Every figure it uses becomes a numbered row: what it is, its value, its kind, and where it lives inside the instrument, down to the file, the feed or the equation. A row from a patterned estimate carries less weight than a row from a sourced measurement, and says so.
- Measured, your upload, curated, part curated, patterned
- Derived, live feed, the simulation, the brain, your company
-
04 / 05
It concludes, and says how sure
Each conclusion cites the rows it stands on, and is as strong as the weakest of them. The one that answers the question as asked leads. A neural network that fits a figure poorly on countries it has not seen is reported as fitting poorly rather than leaned on, and a pattern across countries is never called a cause.
- Confidence per conclusion, and for the whole answer
- What would change it: the coefficients it is sensitive to and the nearest thing that has happened
-
05 / 05
It writes it up, and every sentence is checked
Where a model key is configured, the conclusions and their rows are written up as prose. The server then reads every sentence: it must cite a row, every row it cites must exist, and every figure in it must be a figure in those rows. A sentence that fails is removed, and the page lists what was removed and why.
- Figures may be rounded, never invented
- The key never leaves the server
What it can be asked
No template, no example to copy, no list of approved questions. These are the kinds of question it takes apart, alone or in one sentence together.
02
What if
A sentence about a change is read, weighed into the ways it could go, run against the baseline on the same seed, and read back: world output and prices, who is hit hardest and who gains, the channels it travels through, how much hangs on which way it goes, and the signs in the world that would show which way it is going.
03
What does it mean for us
With your files in the data room, every run is put against the company line by line: revenue by country, costs by supplier, the wage and energy bills, debt and cash by currency, and every site rated against the run. Then across every future, not only the median one: the spread, the share of futures that lose, the result in each way the sentence could go, and other countries with the capacity for any flagged input.
04
What would change it most
The shocks a company's own files or a country's own ties are most exposed to, each run alone against a world with nothing changed and ranked by how far it moves the answer, with the supply a shock puts at risk named even when the model cannot price it.
05
What is happening now
Today, for a place, for the company's own countries or for the world: the Live desk's pressure and what makes it up, the headlines naming it, the chokepoints near it, and the model's structural view beside the news.
06
Where is the risk
The scored hotspots, instability on its own, where the run sees stability falling, the latest register of organised violence, and which of your own countries sit among them.
07
How will it look
The baseline first, with its band, and the scenario on screen beside it, with the reliability of each run and the coefficients each is most sensitive to.
08
Where does this come from
How each country's figure was made, what that means, which file or feed holds it, and what would firm it up.
09
What should we upload
The least firm figures the brain leans on hardest, weighted toward your own countries, each with where to get it and where to put it; then the company files still missing.
10
How do these go together
Two figures across every country, as a rank and a linear correlation, reading a figure the record does not carry (growth, say) from the baseline run and saying so.
Every row says what kind of thing it is
Ten kinds, each with a weight that caps how strong a conclusion resting on it can be. The weight is shown beside every row.
| Kind | What it means | Where it lives |
|---|---|---|
| Measured | Fetched from the World Bank by the refresh and stamped with its source and year. | the country record, source and year on the row |
| Your upload | A world figure an administrator supplied, marked as the organisation's own wherever it appears. | the intake, stored beside the record |
| Curated | Entered by hand from published statistics. | inc/data/countries.php |
| Part curated | Population and income entered by hand, the structure of the economy patterned from region and income tier. | inc/data/countries.php |
| Patterned | Filled in from the region and income tier, not measured. | inc/data/countries.php |
| Derived | Computed here from other rows: a composite index, a correlation, a count of provenance marks. | the function that computed it, named on the row |
| Live feed | Read by the server from a named live source this session. | the live tier, with the time it answered |
| The simulation | A median and band across the futures of a run, or a term of the engine's own equations. | engine.js, with the seed and the number of futures |
| The brain | An estimate from one of the two neural networks, with its fit on countries it did not see. | the Brain tab, network A or B |
| Your company | A line from your own files put against the run. | your data room file and its row number |
What it will never do
Say a number it cannot point to
Every figure in every conclusion is in a row, and every row says where it came from. The written up version is checked sentence by sentence on the server before it reaches the page.
Lean on a weak fit
A network that scores poorly on countries it has not seen is reported as scoring poorly, with the score, and its pattern is left out of the answer.
Call a pattern a cause
What goes with what across countries is labelled as that. The only causal story told is the engine's own arithmetic, and it is labelled as the model's.
Take the working with you
Every answer can be kept (in your account when signed in, in this browser otherwise), shared with the team, opened as a report laid out for print or PDF, handed on as a link that asks the same question again on the reader's own run, and downloaded as its evidence table in CSV or the whole analysis in JSON. A kept answer reopens with the figures of its own day and says so, and watches the countries it turns on: when one of them turns high on the live desk, the answer is flagged to be asked again. A what if can be put back to the scenario that was on screen before it.