VIDECRA

Methodology

Where every number comes from

VIDECRA is a deterministic scenario engine, not a forecasting model. This page is the public version of the model card: how a result is produced, what the evidence grades mean, what the engine refuses to do, and which ruleset produced what you see.

The chain

From a stated shock to an implied level

  1. 01

    A stated shock

    You choose the shock — a US policy-rate move in basis points, a chokepoint capacity loss, an insured-loss severity — its horizon and its severity mode. The engine never estimates whether the shock will happen.

  2. 02

    A versioned ruleset

    Each family reads one JSON ruleset with a version number. Rules change only by a human edit, a version bump and a re-run of the golden scenarios; no process edits them automatically.

  3. 03

    A graded channel

    Every transmission channel carries a direction, a sensitivity per 100 bps, a lag in months, a confidence, its transmission path and the source it rests on — plus an evidence grade.

  4. 04

    A range per indicator

    The result is a low–high change per indicator, with the horizon and severity multipliers applied. The same inputs always produce the same run_id and the same numbers.

  5. 05

    An observed anchor

    Where an official series exists, the latest observed value is shown beside the modelled change and the implied level. Observed data is context for the reader; it never feeds the engine.

Evidence

Three grades, shown everywhere

Empirical
The channel and its sign are supported by a cited public source or a validated historical episode. The source is printed with the channel.
Expert
An illustrative domain assumption. The direction is defensible; the magnitude is a scaling that an economist must validate before client use.
Hypothesized
A plausible relationship that is not yet validated. Kept visible so it can be challenged, never hidden inside a total.

An empirical grade certifies the channel and its sign. It does not certify that a magnitude is calibrated to any real portfolio.

What the engine does not do

  • It does not forecast. Every output is conditional on the shock you specified.
  • It does not estimate the likelihood of a rate move, a disruption or a catastrophe.
  • No number, rule or causal link is ever produced by a language model.
  • It gives no investment, trading, hedging, legal or policy advice.
  • The default book is synthetic. Your own exposures enter only through the custom-exposure upload and stay with your organisation.

In force now

Rulesets and their evidence mix

Read live from the same files that produce the runs. A version here is the version printed on every result and every PDF.

FamilyRulesetChannelsEvidence mix
UAE rate transmission………
Saudi rate transmission………
Euro-area rate transmission………
China rate transmission………
US rate shock (portfolio)………
Trade chokepoints………
Natural catastrophe………

Source: GET /rulesets on the production API — counts of channels per evidence grade in each ruleset file.

The world beside the model

Observed series

Official series from FRED, the ECB, the IMF, the BIS, the World Bank and central-bank releases, with history for charting. Shown beside modelled indicators as anchors, and in Market Context as a reading of the day. Never an input to a rule.

Corporations and holdings

Banks, national companies and sovereign holdings appear as nodes of the transmission map with their own published figures — each taken by hand from the entity's own report and cited to the document. They move indicators the way states do; the source names the page.

Evidence pipeline

A scheduled job collects new working papers and official notes from public research sources and tags them to families and indicators. Every candidate is reviewed by a person. A candidate can lead to a rule change only through a human edit and a version bump — the pipeline itself cannot touch a ruleset.

AI reading

A narrator that may only cite

The AI brief and the analyst note explain an already-computed run in plain language. They are bounded by construction:

  • The model receives only the run's facts — numbers, sources, run_id, ruleset version. It sees nothing else.
  • Every sentence with a number, a direction or a causal claim must carry the run_id and the key of the source that supports it.
  • A server-side check rejects any reply that cites a source not present in the facts, or that breaks the note's fixed structure. The reader then sees no text rather than an unverified one.
  • Model scratchpad is filtered out; a reply that is visibly reasoning instead of answering is discarded.
  • The model is a hosted provider configured by the operator, with a fallback; the model name is returned with every reading. Its text is never written back into a calculation, a ruleset or a report figure.

Evidence coverage — not accuracy

Each run carries a coverage figure between 0 and 1. It weights the evidence grades of the channels used and whether each modelled sign matches published history. It is an audit aid for a reviewer, not a backtest, not an error metric, and never an input to the calculation.

per channel: min(base + direction_bonus, 1) × confidence_weight · base: empirical 1.0 / expert 0.7 / illustrative 0.4 · direction_bonus: +0.25 when the sign matches an official truth point · overall: confidence-weighted mean

A high figure means the channels are well documented. It does not mean the scenario will happen, and it does not replace the economist validation every expert-graded channel still requires.

Human accountability

Decisions remain with the people who make them. Every result, brief and PDF shows its sources, assumptions, evidence grades, uncertainty band, scenario inputs and ruleset version, so that a reviewer can trace each number back to the rule and the document behind it.

© 2026 Senfina Creative Studio FZC LLC, UAE. All rights reserved.