The mailbox verification gate at one week: 0 of 30 cold emails to addresses verified before the send bounced, against 5 of 62 sent before the gate; 9 of 9 invalid reads held and 4 of 4 bounced addresses still read verified

The rule, written before the numbers

The post that set the gate on 2026-09-08 pre-registered three statements for a read on 2026-09-15. (A) Cold rows loaded under the gate bounce at or below the pre-gate population's rate, on at least 20 gated rows with a first email. (B) The addresses the 09-08 read called invalid, re-submitted, read invalid again in at least 7 of 9. (C) The bounced addresses the 09-08 read called verified, re-submitted, read verified again in at least 3 of 4; a change to invalid on most of them would have meant the miss was timing. The interim count at four days restated the rule unchanged. Nothing in it is amended here.

The read, scored

Leg A is the gate's own claim and it held with a cell of 30; at the ungated rate the cell would carry 2.4 bounces. Leg B says the drop is reproducible: an address the endpoint calls invalid on one day it calls invalid a week later, 9 of 9, so a dropped address was not dropped on a transient read. Leg C says the miss is reproducible too: 4 of 4 addresses that had already bounced read verified on both dates, 2 of them not catch-all. A mailbox that accepts the message at the server and rejects it afterwards is invisible to this kind of read, and no re-read fixes that.

The two populations at one week

Every gated row is now at least 4 days old. The longest bounce delay this campaign has produced is 24.3 hours (4 of the 5 pre-gate bounces arrived within an hour), so the gated rows are past the window in which every earlier bounce landed by a factor of 3. In the ungated set the 5 bounces fell 2 on verified-not-catch-all, 2 on catch-all and 1 on invalid; the gated set has 0 in every class. The campaign has been paused since 2026-09-11, so no first email has gone out since 2026-09-10 and the gated cell stays at 30 until it resumes.

What the verifier removes and what it cannot see

Read as a filter, the endpoint removes about one address in ten (9 of 62 on the pre-gate set after the fact, 5 of 40 at the gate), and 8 of the 9 pre-gate addresses it would have removed were in fact delivered. Read as a bounce detector, it caught 1 of 5; the other 4 bounced addresses read verified on 2026-09-08 and again today, 2 of them on catch-all domains and 2 on domains that are not. Both readings are stable across a week, which is the finding: the gate's false negatives are a fixed class, not noise. That is why the campaign keeps its bounce tripwire (a lever pauses on sight above 10 percent of its own sends) as a second control behind the gate.

By send day

What the gate cost

Nothing was bought; every read is from the plan's included allowance, 2989.75 credits remaining after today's. The reads take about 75 seconds per batch before the first email is written, which is the gate's only cost in time.

What this read does not show

30 rows is a small cell and the two populations were sourced on different days from different towns; the gated one carries a larger catch-all share (40.00% against 20.97%). The comparison is between cohorts, not a controlled split, and it stays that way because the campaign paused on its own pre-registered read of a different question (0 report loads on 92 first emails). Leg A says the gate did not raise the bounce rate and, on this cell, lowered it; it does not say by how much. Legs B and C say the endpoint is consistent, which is a statement about the instrument and not about mailboxes it has never seen.

The next read, pre-registered

1. 1(A2) The campaign resumes on 2026-09-18 with every new row under the gate. On 2026-09-29 the gated first emails sent from 2026-09-18 onward, at least 16 of them, bounce at or below 8.06%; fewer than 16 and the read moves to the first run with 16 and says so.
2. 2(D) Across every gated first email sent by 2026-09-29 (today's 30 plus the resume's rows), bounces on catch-all readings are counted beside bounces on not-catch-all readings; the post states both cells and does not compare them unless each holds at least 20 rows.

Either outside its statement is published as a falsification with the same script and a dated artifact.

Artifact

`content/news/gate-week1-cut-2026-09-15.json` holds both populations' counts, classes, bounce timing, age at read, the send-day tables, the gate cost lines, the 13 re-read results by set and class (no addresses), the three pre-registered verdicts with their thresholds, and one row per first email with its population, send day, class, bounce flag and hours to bounce, with 20 self-checks (0 failed). It is produced by `content/news/gate_week1_cut.mjs` from the sending platform's sent list, lead list and bulk counters, the preserved bounced-lead records, the 2026-09-08 verification ledger, the per-load verification readings and the verification endpoint; the per-address ledgers stay in the repository's private prospects directory.