A database that learns
Most databases forget. They hold what they were given and answer what they are asked, and every question starts from zero. EquationDB now keeps something more: what it has learned, and how sure it is, written down every night.
Claims, tested again and again
A claim is a belief the database holds with evidence. Today there are three kinds:
- the 259 theses on Thesis: breakouts, fear, earnings drift, the calendar, gaps, weather;
- every findings detector in every market, with its record of directional calls scored after the fact;
- discoveries: beliefs users tested on Thesis that came out supported or mixed. The statement joins the ledger; the user is never named.
Every night the learn job re-tests them on the data that arrived since. Twenty-five theses a night, oldest first, so each is re-tested every couple of weeks; every detector's record, every night. Each result is appended to the claim's history. Nothing is overwritten.
Confidence over time is evidence
A claim's status comes from its history, not from one test:
- established: supported now and in the two tests before;
- emerging: supported, but not yet three times running;
- weakening: was supported, now only mixed;
- flipped: was supported, and now the opposite happened.
What changed each night goes to a log: a new claim, a verdict that changed, an edge that grew or faded. Read over months, it is a record of which market beliefs hold and which wear out, measured the same way every time.
Why this matters
An edge that worked for twenty years and stopped two years ago looks fine in a twenty-year backtest. A rule that only ever had a small sample looks great until it meets a bigger one. Re-testing on a schedule, and keeping every result, catches both. It is what a careful researcher does by hand, done every night, for everything the database believes.
It also compounds. Each day adds tests. Each user who finds something real adds a claim. The ledger gets more useful the longer it runs, which is the opposite of how a query box ages.
Open, in SQL
The ledger is a table like any other:
SELECT title, verdict, t, edge_pct, tests, t_change FROM knowledge WHERE status = 'weakening' ORDER BY t_change
SELECT * FROM knowledge_log ORDER BY date DESC LIMIT 50