Methodology

Effective Nov 17, 2026, revised Nov 19, 2026.

Redaction rules v1 · Matching rules v1 · Skeleton grammar v1 · Area-code data nanpa-2026-11-17

Who the reports come from

Reports are voluntary submissions from people who chose to file them. They are not a sample of all spam, and counts do not measure how much spam is sent. Early data may come from very few reporters.

The geographic proxy

Every geographic view uses the sender’s displayed area code. Sender numbers can be spoofed, so this is the claimed origin, not a location. The area code’s numbering-plan reference region is the state or territory it is assigned to. It is kept separate from any city named in the message, which is never shown publicly.

Area-code data: Generated from the NANPA NPA database CSV (https://reports.nanpa.com/public/npa_report.csv, linked from https://www.nanpa.com/reports/npa-reports) by scripts/verify-area-codes.mjs. Scope: in-service US geographic codes (states, DC, territories) and in-service toll-free codes. Area-code assignment is a numbering-plan fact, not a sender location. Corrections welcome (see CONTRIBUTING.md).

Area codes outside the US (Canada and the Caribbean) are not accepted in this version. Toll-free codes and US territories are counted in the table but have no map tile.

How patterns are matched

Every report is reduced by fixed rules to a masked template: names, numbers, links, emails and places are replaced with labels. Matching runs only on that template, in three separate tiers. Each tier has its own label and they are never merged.

  1. identical message: the masked templates are the same after lowercasing and spacing are normalized.
  2. same message, lightly edited: the templates share most of their 3-word sequences (estimated with 64-way MinHash; similarity at least 0.6 to the group’s first report). Each report is compared to the group’s first report only, so a chain of small edits cannot pull unrelated texts together.
  3. shared script pattern: the templates follow the same script skeleton, described below. This tier alone never implies a shared sender.

3 eligible reports that match are a review trigger, not automatic publication (see Review before publication). A report is eligible while it exists and is not a short-window duplicate (the same sender hash and the same template within 24 hours, or the same sender hash, template and connection within 7 days, while the server has been running). Three reports do not establish three independent reporters, three senders, or common control. If deletions take a pattern below 3, it disappears from every public view at once. A group that matches the same reports as a higher tier is listed once, under the higher tier.

Review before publication

A group that reaches 3 eligible reports is held for review and is not shown anywhere, or counted, until a maintainer publishes it. The review checks that the match is real, that no personal detail survived scrubbing, and that the pattern does not read as identifying a person or a single business. The maintainer can publish it or reject it.

A decision covers the exact set of distinct masked templates that was reviewed. If a template joins or leaves the group, or deletions take it below 3 at any point, the decision lapses and the group goes back to review. More reports of templates already reviewed do not; the report count on a published pattern updates as reports are filed or deleted.

Review affects only whether a pattern is published. Every report still counts in the totals, the area-code table and the map from the moment it is filed.

Skeleton grammar v1

The skeleton is the order in which these slots first appear in the masked template:

SlotRule
GREETINGOpens with a greeting word or addresses the recipient by name.
PERSONAThe sender introduces themselves by first name.
LOCAL_CLAIMClaims to be local, or names a city or neighborhood.
TRADENames a home-service trade (roofing, lawn, tree, painting, and so on).
OFFEROffers a quote, estimate, discount, or something free.
CTAAsks a question or asks for a reply or call. Placed at the last such sentence.

Fewer than 3 slots: no skeleton. Skeletons with fewer than 4 slots, and these generic skeletons that ordinary texts also produce (appointment reminders, realtor check-ins, contractor quote follow-ups, opt-in local ads), are never shown as patterns:

  • GREETING>PERSONA>TRADE>CTA
  • GREETING>PERSONA>OFFER>CTA
  • GREETING>PERSONA>LOCAL_CLAIM>CTA
  • GREETING>OFFER>TRADE>CTA
  • PERSONA>TRADE>OFFER>CTA
  • GREETING>PERSONA>TRADE>OFFER>CTA
  • GREETING>TRADE>OFFER>CTA
  • GREETING>LOCAL_CLAIM>TRADE>OFFER>CTA

Pattern IDs are stable within a grammar version on this site. When the grammar changes, its version changes, and patterns may regroup.

Rules versions

The redaction rules (how a message becomes a masked template) and the matching rules (the tiers, the threshold and eligibility, and skeleton grammar v1) each carry a version: Redaction rules v1 · Matching rules v1. A version changes whenever that behavior changes. A change to the matching rules sends every published pattern back to review. The changelog is in the project repository (docs/rules-changelog.md).

Worked example (synthetic)

Message as filed
Hi Robin, Theo here! I do pressure washing for a few homes around Brookfield this month. Happy to give you a free estimate, text 555-201-0147. Want one?
City named in the message
Brookfield
Masked template (stored, never published)
Hi [recipient name removed], [sender name removed] here! I do pressure washing for a few homes around [city named in message] this month. Happy to give you a free estimate, text [phone number removed]. Want one?
Skeleton
GREETING>PERSONA>TRADE>LOCAL_CLAIM>OFFER>CTA
Published example for this skeleton
Hi [recipient name], [sender first name] here. Do you need any [trade] work done? I’m working near you in [city]. I can give you a [free quote / estimate]. [Question asking for a reply]?

Known limitations

  • Selection bias: people who report are not a random sample.
  • Spoofing: area codes can be faked.
  • Extraction errors: screenshot reading and the masking rules can miss or over-remove text.
  • Template splits and collisions: small edits can split one script into several groups, and short generic texts can collide.
  • Coordinated submissions: a few people can file matching reports on purpose. The threshold and review stop accidents and obvious cases, not determined abuse.
  • Whether a given script is distinctive cannot be known from a handful of reports.

Revision history

  • Nov 19, 2026: patterns are published only after review; 3 matching reports now trigger a review. Redaction rules v1 and matching rules v1 named as the rules already in use. The review step changes publication only, not redaction or matching.
  • Nov 17, 2026: grammar v1 and this page first published.