Lab notebook · rebuilt daily from a redacted export

DARWIN

As of —. Export generated — UTC.

What this is

DARWIN is an autonomous system that runs a small business's learning loop: it locks predictions with a probability and a resolve-by date, grades them when they resolve, proposes changes to its own parameters, and is only allowed to keep a change if a frozen set of already-graded cases does not get worse. A separate model, the Challenger, must write an accept, reject or revise verdict on every proposal before it can reach that gate. It also runs experiments and hunts for new ventures against a fixed bar.

This page is the public record of that loop. Numbers come from the same code that writes the operator's morning digest. Every field is on a named allowlist and the whole export is scanned for email addresses, links, postcodes, money and names before it is written; if either check fails, nothing is published and the previous day's page stands.

Headline

In the margin

Learning window, —

One row per day the loop ran unattended. The digest is the operator's morning report; the gate is the only thing that can change a parameter.

Date	Dispatcher (parameter proposals)	Learning cycle (new evidence)	Gate/harvest loop	Graded	Gate	Ledger rows	Challenger p/a/r/v	Digest

Calibration by regime

Brier score, lower is better; 0.25 is what always saying fifty-fifty scores. A regime with no graded case has no score yet, and this page says so rather than pooling.

Regime	Locked	Resolved	Graded	Mean Brier

Plumbing audit

Verdicts

Hunt	Verdict

This week

Changelog

Rulings are decisions the operator made and committed. Fixes are changes DARWIN made to itself. Plain English, dated, nothing identifying.