How is the mathematics of a decision engine built and validated?

What does it take to build the mathematics of a decision engine?

Building the mathematics starts on paper. The team names the question, lists the factors that move the answer, and writes one equation that combines them. Weights and thresholds are fixed before any case is scored, so no result can bend the rules afterwards. The EQOS Asia decision engine follows this logic with a universal equation, φ(S), computed on 9 structural dimensions (EQOS Asia method page, as of October 2026).

Six steps carry the build, from framing the question to stress-testing the score. Each step leaves a written trace: the question, the sources, the equation, the test cases, the recomputed values, and the input ranges. A reader who holds those six traces can judge the engine without trusting its author. The figure below shows the sequence.

Which inputs feed the equation of a decision engine?

Inputs fall into three groups: facts that can be measured, judgments that experts rate on a scale, and context that sets the rules of the game. Every input carries its source and its date, so a score can be refreshed when one of them changes. A short list of well-defined inputs gives the equation firmer ground than a large pile of loose data.

In the EQOS Asia method, structured interviews of 2 hours at the Standard level, or 3 days at the First level, collect the inputs, and documents are collected and vectorized alongside them (EQOS Asia method page, as of October 2026). The guide on the inputs a decision engine uses details the three groups with examples.

How is a decision score read once the equation has run?

The equation returns one normalised number between 0 and 1. A threshold turns that number into a zone, and the zone points to an action. On the EQOS Asia scale, a φ(S) score below 0.618 falls in the STOP zone, and a score above it falls in the OK zone. The Audit pillar also lists the 3 priority risks and a map of fragility zones (EQOS Asia method page, as of October 2026).

A number without a rule for reading it invites argument. The rule should be public, simple, and fixed in advance, so two readers reach the same conclusion from the same score. The companion guide explains how to read a decision score, from thresholds to next steps.

How is a decision engine back-tested before anyone relies on it?

A back-test replays cases whose outcome is already known. The engine sees only the data that existed at the time, produces its score, and the result is compared with what happened next. A model that cannot separate past successes from past failures has little to offer on a new case. Dated, public inputs keep the replay honest.

EQOS Asia publishes retrospective applications of its engine to Theranos with public data from 2013, WeWork with public data from 2017, and the 2008 financial crisis with public data from 2006 and 2007 (EQOS Asia back-test pages, as of October 2026). The guide on how to back-test a decision engine explains what a retrospective can and cannot show, since hindsight is a known risk.

How are the calculations verified and reproduced?

Verification asks whether the arithmetic is right. Reproduction asks whether a second person obtains the same number. A deterministic core makes both checks possible, because the same inputs always return the same result. The EQOS Asia method page states, as of October 2026, that every verdict can be recomputed identically by any data scientist with access to the Codex, under NDA.

The guide on the verified calculations methodology lists the checks any engine should pass: independent recomputation, comparison with hand-worked cases, and a log of every engine version. To reproduce a decision engine result, a reviewer needs the frozen inputs, the engine version, and the date of the run.

How far can one input move a decision score?

A sensitivity analysis moves one input at a time across its plausible range and records the change in the score. Inputs that swing the score widely deserve the most careful sourcing. Inputs that barely move it can rest on rougher estimates. The test also shows whether a verdict sits close to a threshold, where a small change flips the zone.

The EQOS Asia simulator lets a reader try this directly: each of the 9 dimensions has a slider from 0 to 100, and an influence view shows how a move in one dimension drives the others (EQOS Asia simulator page, as of October 2026). The guide to sensitivity analysis of a decision score applies the method step by step.

Where does the mathematics of a decision engine stop and judgment begin?

An engine answers the question it was built for, with the inputs it was given. It cannot see a factor nobody entered, and it cannot settle a question of values. The guide on the limits of a decision engine lists these boundaries, so a reader knows when to trust the score and when to widen the question.

The score then meets expert judgment. The calculation gives the debate a fixed starting point, and an experienced executive adds what the inputs left out. A 30-minute pre-diagnostic by video, scheduled in Singapore time (SGT), applies the EQOS Asia equation to a real case (EQOS Asia contact page, as of October 2026).

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile