Recall Quality

Recall Quality

The House recalls justified information, never raw history.

Ordinary retrieval returns whatever is most similar to the query. That is how a system ends up citing its own disproven conclusion as authority — the wrong answer is still the most relevant one.

Recall Quality scores every retrieved memory on five axes before it is allowed to influence anything.

The five scores

Field	Asks
similarity	how close is this to the current question?
accuracy_score	when this was graded, was it right?
calibration_score	was its stated confidence proportionate to how right it turned out to be?
outcome_status	what actually happened afterwards?
validity	does it still hold now?

Similarity alone is what most systems use. The other four are what stop the House from repeating itself confidently.

The five validity states

A memory is not simply true or false. It carries a state:

State	Meaning
VALIDATED	graded, accurate, current, not superseded
PARTIALLY_VALID	mixed outcomes, or correct-but-aging (past 270 days)
OUTDATED	superseded by a newer ruling, or past the staleness horizon (540 days)
DISPROVEN	graded predictions resolved incorrect
UNKNOWN	no graded outcomes yet — usable, but never as authority

The rule that matters

A disproven memory is never cited as authority.

It is not deleted — deleting it would erase the evidence that the House was once wrong, which is precisely the most instructive thing it holds. It is retained, marked, and excluded from justification.

That distinction — retained but not authoritative — is the whole design. A system that deletes its errors cannot learn from them; a system that cites them cannot be trusted.

Where it is used

Recall runs before the council deliberates, so members reason on top of justified history rather than raw text. The House Mind is updated from the same graded pool.

→ Next: House Mind