Skip to main content

0015 · Salience had no reader

Memory namespace · version 15 · source

A column that existed because it was expected​

fact.salience was written into the first migration and has been zero for every fact ever stored. Nothing computes it. Nothing orders by it. Nothing reads it.

Why it goes rather than getting a computation​

Salience is a ranking signal, and recall deliberately ranks nothing yet — a ranking stage introduced before the plain path is measured makes it impossible to say afterwards which stage is doing the work and which is doing harm. So salience arrives with the measurement that justifies it, or it does not arrive.

That measurement needs a corpus at scale on hardware that can be held still, which is not available. Keeping the column until then means shipping a schema where one column is a promise and every other is a fact — and a reader who finds it zero for the whole history learns that this schema contains things that are not true, which is the more expensive lesson.

What this does not undo​

The distinction it was created for stands: a confidence is REPORTED by the extractor and a salience would be COMPUTED from behaviour, and collapsing them would make a weakly-extracted fact mentioned often indistinguishable from a confidently-extracted one. confidence stays and keeps meaning exactly what it meant.

Adding the column back is one migration. Deciding it was needed is the part that requires evidence, and that is the part being deferred.

What it runs​

  • ALTER TABLE fact DROP COLUMN IF EXISTS salience

SQL​

{schema} is replaced with the validated namespace name when the migration is applied.

28 lines
-- ── A column that existed because it was expected ─────────────────────────────────────────────
--
--  `fact.salience` was written into the first migration and has been zero for every fact ever
--  stored. Nothing computes it. Nothing orders by it. Nothing reads it.
--
--  ── WHY IT GOES RATHER THAN GETTING A COMPUTATION ────────────────────────────────────────────
--
--  Salience is a ranking signal, and recall deliberately ranks nothing yet — a ranking stage
--  introduced before the plain path is measured makes it impossible to say afterwards which stage is
--  doing the work and which is doing harm. So salience arrives with the measurement that justifies
--  it, or it does not arrive.
--
--  That measurement needs a corpus at scale on hardware that can be held still, which is not
--  available. Keeping the column until then means shipping a schema where one column is a promise
--  and every other is a fact — and a reader who finds it zero for the whole history learns that this
--  schema contains things that are not true, which is the more expensive lesson.
--
--  ── WHAT THIS DOES NOT UNDO ──────────────────────────────────────────────────────────────────
--
--  The distinction it was created for stands: a confidence is REPORTED by the extractor and a
--  salience would be COMPUTED from behaviour, and collapsing them would make a weakly-extracted fact
--  mentioned often indistinguishable from a confidently-extracted one. `confidence` stays and keeps
--  meaning exactly what it meant.
--
--  Adding the column back is one migration. Deciding it was needed is the part that requires
--  evidence, and that is the part being deferred.

ALTER TABLE {schema}.fact DROP COLUMN IF EXISTS salience;