6,288,938 outbound dials, and which of the two levers actually moves

The attempt curve, with the median campaign beside the pool

The median campaign out-answers the pool at every attempt, and that is not decoration. It says the largest campaigns run below-median connect rates, so a pooled rate understates a typical campaign -- which is exactly the error a reader makes when a vendor publishes one number without the dispersion beside it. Both series are drawn on the same axis so the gap reads as a gap.

The last two columns are the disclosure gate showing its work, not padding: they are the two clauses a reader cannot otherwise check. No published bucket rests on a thin cell -- the smallest holds 182,292 dials.

Finding two

Inside the hours already being worked, the clock barely moves anything

This is a negative result and it is the point. The hypothesis going in was that the answer rate varies strongly by hour. Across the 9 publishable hours it spans 4.918% at hour 11 down to 4.151% at hour 19 -- a factor of 1.185, against 1.994 for the attempt number on the same corpus. Each hourly bucket holds between 120,684 and 791,683 dials, so this is not a small-sample flat line.

Methodology

The grain and the label

one row per outbound dial attempt, read from vicidial_log joined to vicidial_statuses and vicidial_campaign_statuses. The numerator is human_answered='Y', campaign dictionary first, then system dictionary -- the campaign's own status dictionary is consulted first because a campaign may redefine a status code, then the system dictionary. An UNMAPPED bucket is emitted even when it is zero: 535 rows (0.0085% of the corpus) matched neither dictionary. They are counted as not-answered and they stay visible, because a status nothing describes must never be silently folded into a class.

Why whole weeks

Monday to Sunday, whole weeks only, so no weekday is over-represented. The window runs 2026-05-25 to 2026-07-26 inclusive, 9 complete weeks, and it is trimmed inside the replica's range at both ends so no partial day is counted. An hour-of-day cut over a ragged window measures the calendar as much as the clock.

Where the two cuts do not reconcile, and by how much

The hour cut closes exactly: its published and withheld buckets sum to 6,288,938 dials, the whole corpus. The attempt cut does not. Its buckets sum to 6,288,796, leaving 142 dials -- 0.002% of the corpus -- carrying no usable attempt counter and therefore sitting in no attempt bucket at all. It moves no published rate and it is stated because a page whose argument is that distributions should reconcile has to reconcile in public.

The disclosure gate

Every dial was placed for a client. A bucket one campaign dominates is that client's operating number, not a pooled statistic. Withheld buckets are listed rather than deleted so the published curve is not mistaken for the whole curve. The thresholds are 75% maximum single-campaign share, 5 campaigns at 1,000+ dials, and a 100,000-dial floor. They are enforced in the generator and re-checked in the page's data module at build time, so a bucket that violates one cannot be rendered even by hand.

The window is fixed

Nothing refreshes it. A dataset that silently re-derives itself lets a published figure change underneath a citation, so re-running the queries produces a new dataset with a new window rather than an update to this one.

Privacy

Aggregate only. Every query aggregates in SQL and returns counts; no dialled number, lead identifier or free-text comment is in the shipped dataset, and the generator sweeps every string it emits for phone, email and identifier shapes before writing. No campaign is named. No bucket one campaign dominates is published. The same three rules govern the sibling inbound answer-rate benchmark.