gtmjosh

Interactive guide

Make the Decision Without Faking Certainty

A decision system almost never fails by returning an answer somebody can immediately see is wrong. It fails by returning an answer that looks exactly like a good one.

A required field was empty, so the calculation treated the absence as a bad signal instead of a missing one. A score of 20 gets read as “a bit unhealthy” when it actually means “no meeting in four months and these are all email points.” A confidence figure of 0.9 gets treated as a nine in ten chance of being right, which is not what it measures. Every one of those produces a plausible number, and none of them announces itself.

You will build a Decision Design: the shape the decision takes, deterministic rules doing the work they are good at, a model used only where interpretation genuinely helps, an explicit answer for missing and conflicting evidence, an honest confidence policy, one consistent output shape, and a test set of awkward records that has actually been run. The lab saves your progress in your browser.

Start with a broken answer

Pick the Right Decision Shape

Chapter 1 of 8 · 12 min

Start guide →

The learning path

Every chapter starts from a plausible wrong answer and works backwards to whatever produced it. An empty field read as a bad signal. A score of 20 that means two unrelated things. A confidence figure that measures something other than what the person reading it assumes.

Choosing how the answer gets made

  1. 01Pick the Right Decision ShapeChoose between classifying, scoring, ranking, and recommending, and know what each can and cannot answer.12 min · exercise
  2. 02Apply Simple Rules FirstDo the deterministic work first, and let one condition short-circuit the rest.14 min · exercise
  3. 03Use AI Where Interpretation HelpsDraw the line between what a rule can check and what needs reading, and put the model on one side of it.14 min · exercise
  4. 04Handle Missing and Conflicting EvidenceStop absence being read as a bad signal, and decide which source wins when two disagree.14 min · exercise

Surviving contact with reality

  1. 05Treat Confidence CarefullySay what a confidence number does and does not mean, and keep a composite score’s parts recoverable.12 min · exercise
  2. 06Return a Consistent ResultReturn the same shape every time, carrying what the next guide needs to explain it.12 min · exercise
  3. 07Test Plausible Edge CasesBuild the test set from real failures, and give a population that behaves differently its own rules.14 min · exercise
  4. 08Capstone: Run the Qualification SetRun every design section against a full set of records at once.15 min · exercise