Nexiphor Verifier

Evidence state: Early

Trust & certainty

Three different questions — we never mix the answers:

  1. Did it deliver? Send-safety among emails we called Valid (Tier A). Still collecting.
  2. Did we return a status? Coverage on a real production batch — not bounce proof.
  3. Did we match a lab answer key? Curated test set only — not a bounce rate. Details collapsed below.

What each status means

  • Valid — mailbox accepted by SMTP (or an authoritative provider path) without catch-all ambiguity. Protocol evidence — not a delivery receipt.
  • Invalid — syntax, disposable, no MX, or a definitive mailbox reject.
  • Risky — catch-all, role/toxic signals, or a non-authoritative accept. Still actionable.
  • Unknown — timeout, greylist, or probe blocked. No charge for Unknown — one credit is refunded automatically. See the refund policy.

Primary claim · send-safety

Did our Valid emails actually deliver?

Deliverability Valid precision: Delivered ÷ (Delivered + HardBounce) among emails we predicted Valid. Soft bounces excluded. This is the only number that answers send-safety.

Production Valid precision is unmeasured — no matched Tier-A delivery outcomes yet.

Did we return a status? (coverage)

How often a production run returns an actionable status — not whether that status matches delivery ground truth.

No production resolution run has been published yet.

Controlled mailbox benchmark

Tier B — mailboxes Nexiphor owns, so existence is known by construction. Strong evidence, intentionally small until we grow the set.

n=—

Target ≥100 Valid/Invalid/Risky cases with verified_at and decay policy. New cases only land in new immutable dataset versions.

Supporting detail · not a bounce rate

Answer-key check (lab only)

Exact-match vs a curated labeled answer key. This is laboratory classification agreement — not observed bounce or inbox delivery. Useful for engine regression. Never the primary send-safety claim.

No laboratory classification figure is attached to the public snapshot yet.

Methodology

Four packs, ranked for readers. Primary: deliverability Valid precision (Tier A). Then coverage (not Unknown). Lab exact-match / P/R/F1 is supporting only — never a bounce rate. False-Valid among predicted Valid is tracked separately. Soft bounce is never HardBounce in the deliverability pack.

Publication gates. Public deliverability % requires ≥1,000 outcome-observed Valids plus Wilson CI. Below 100: Insufficient sample. Laboratory claims need ≥1,000 labeled rows, Valid share, live probe share, and provider diversity.

How verification works. Syntax and disposable checks, MX, SMTP RCPT from our probe IPs, optional Microsoft 365 GetCredentialType, then status mapping. Unknown is preferred over guessing.

Limitations

  • Production Valid precision stays unmeasured until Tier-A delivery feedback accumulates.
  • Catch-all domains accept everything by design — we grade risk, we do not invent mailbox existence.
  • Laboratory labels can disagree with delivery ontology (e.g. Risky vs Valid). That is classification agreement, not bounce rate.
  • A mailbox can be deleted after we check it. Every figure is a point in time.

Verify these numbers

This page renders whatever the public API returns. Call it yourself — no account required:

GET /api/v1/public/trust-metrics

GET /api/v1/public/trust-metrics/history