How Firsthand promotes and decays

A pattern seen once stays in the journal. Seen three times, it becomes a durable fact. The migration between tiers is itself recorded, so a reader can find the moment a pattern stopped being a coincidence.

PROMOTION_THRESHOLD  3
DECAY_DAYS           90

Promotion

Three occurrences of one pattern inside the decay window promote it to a WARM entity carrying first_seen, last_seen and n. Firsthand then writes a journal event recording the promotion.

Occurrences are counted on each observation's own timestamp, not on the journal's write time, so backfilling an index cannot promote a pattern whose evidence is already stale.

Three, not two, because in this dataset two is coincidence.

Contradiction

When a new observation disagrees with a durable fact, the entity is overwritten. Uniqueness on (tenant, category, name) means it cannot drift into two versions. Both the old and the new value go to the journal, so the change stays auditable even though the entity keeps no history of its own.

Demotion

A fact whose supporting observations have all aged past the window is archived with reason="evidence aged out", not deleted. We stopped seeing it is not the same claim as it stopped being true, and a reader should be able to tell those apart.

The session baton

write_event(evaluated=..., acted=..., forward=...) carries what the next session must pick up: counterparties due for re-check, claims awaiting corroboration, claimants whose weight is still provisional. Firsthand's first act on boot is to drain it.

The baton carries news. An unchanged verdict hands nothing forward, because re-handing identical work on every sweep grows an append-only journal without bound for no new information.