Inbox health score

Four measurable inputs, weighted by how much the receiving side cares: complaints most, then bounces, then transport failures — our problem, not the recipient's — and engagement as the only upward input, because engagement is what providers score placement on.

Deliberately absent: anything the sender cannot observe. A score that includes a guess at a provider's private reputation is a score with a fiction inside it.

Below a modest sample size the rates are noise, and a score computed on them is worse than no score — one bounce in three sends is a 33% rate and means nothing. The correct output there is not zero, because zero means this mailbox is on fire, and the interface has to be able to tell bad from unknown.

An unmeasured mailbox should also be treated as neutral when volume is allocated. We have no evidence must not be worth more than a proven clean record, and must not be punished like a proven bad one.

The score's job is to move volume. A struggling mailbox should lose share of the rotation gradually rather than being switched off at a threshold — a hard cliff shifts everything onto whichever mailbox has not been measured yet, which is the worst possible destination for it.

A floor on that reduction keeps the pool degrading gracefully: a poor mailbox keeps sending a little, which is also the only way it can generate the evidence needed to recover.

SalesShift publishes its formula rather than describing it. Over a rolling 30-day window: score = 90 − 40 × min(1, complaint rate ÷ 0.3%) − 35 × min(1, bounce rate ÷ 5%) − 15 × min(1, failure rate ÷ 10%) + 10 × min(1, engagement ÷ 5%), where engagement is reply rate + 0.2 × open rate, clamped to 0–100. Under 20 sends it reports insufficient data instead of a number.