Module VI of VI

Bayes in the wild

Every module so far has run the same engine: start from a prior, weigh the evidence, multiply. That engine has been running the world’s plumbing for decades, mostly without credit. Here is a short tour of where it lives, followed by a chance to operate one of its oldest installations yourself.

The spam filter

Every word of an incoming email is a clue with a likelihood ratio — “WINNER” screams spam, “our meeting” whispers legitimate. The filter multiplies them all against the base rate of spam. The Bayesian move: many weak clues, honestly multiplied, beat one clever rule.

The A/B test

Version B converts at 5.1 % against A’s 4.8 % — after 400 visitors. Module V says both rates are wide curves that still mostly overlap; declaring victory now is reading the peaks and ignoring the widths. The Bayesian move: ask how much the curves overlap, and keep collecting until the answer could pay for the decision.

The screening policy

Module II at national scale: screen millions for a rare condition and the false positives arrive by the stadium-load, each one costing money, follow-ups, and sleep. That is why screening guidelines target high-prevalence groups — it’s the prior that makes the test useful. The Bayesian move: a test’s value depends on who you point it at.

The forecaster

The best geopolitical forecasters in tournament after tournament share a style: start from the base rate, update in small multiplicative nudges as news arrives, and keep score. Philip Tetlock’s “superforecasters” are, in effect, walking Module III. The Bayesian move: track your calibration like a golfer tracks a handicap.

Now run the filter yourself. This panel is a hand-sized version of the classifiers that guarded inboxes for years — a naive Bayes filter, so called because it multiplies every token’s ratio as if the tokens were independent. They aren’t (“100% free” and “act now” travel together), and the theory says that should hurt it. In practice the naivety costs little and the multiplication does the work.

Panel VI·1 — The one-line spam filter

prior 1 : 1 · one ratio per token

Half of all incoming mail is spam, so every message starts at 1 : 1. Compose an email by toggling what appears in it — each token multiplies the odds, spam words up, everyday words down.

P(spam | tokens) = 67 % · odds 2 : 1 · filter says: deliver to inbox
Try this Compose an email with your first name, an invoice number, and 'act now.' Which way does the filter lean? Now add 'WINNER' — one loud word against three quiet ones — and watch the verdict flip.

Notice, in passing, what the panel demonstrates about ordinary arguments: a verdict is rarely defeated by attacking one clue. Evidence multiplies, so the case survives losing a token or two — and, symmetrically, a mountain of weak insinuation can add up to something that deserves an answer.

That is the tour. If the six modules leave behind a residue, let it be four habits rather than a formula:

  1. Put a number on it. “Probably” is a hiding place (Module I).
  2. Ask for the base rate before admiring the specific detail (Modules II–III).
  3. Grade every clue by how surprised you’d be to see it if you were wrong (Module IV).
  4. Give quantities error bars, and remember what precision costs (Module V).

None of this requires software or Greek letters. It requires the willingness to be checkably wrong, and the small daily discipline of moving your numbers when the world tells you to.

Retell itOne engine runs all of it: where you start, times the strength of what you see. Everything else is costume.
The glossary collects the vocabulary; the about page says how this site was made and where to read further. The instruments stay switched on — come back and turn the dials whenever an argument smells off.