COMPACTED COMPANY NAMES AGAINST A FEDERAL SCREENING LIST
Aervik Labs — measured 2026-09-19
https://aerviklabs.com/research/compacted-names-federal-screening-2026/

WHAT THIS IS
For 97 real entity names from the U.S. Consolidated Screening List, the name-similarity
score of six spellings of each name against the listed record: the whole name, its first
two words and its first three words, each with spaces and each with the spaces removed.

THE QUESTION
When a company name is typed without spaces — "ArdAlkhair", "XinjiangProduction" — how
often does a name-similarity screen miss the record it should have found, and how often
does it miss it below the floor where anything is reported at all?

THE RESULT
Compacting a WHOLE name is harmless: an edit-distance score barely notices, and all
97 survived. Compacting a SHORT name is not: the first two words of a listed
name, spaces removed, found the record 0 times in 53. With spaces, 52 of 53.
Of 80 missed compacted short-name queries, 38 scored below the 0.5
near-miss floor and would be reported by nothing.

FILES
  results.csv    one row per query: listed_name, listed_tokens, query, query_shape,
                 variant, score, outcome
  summary.json   the aggregates the published page prints, including the matcher's
                 thresholds and the method note
  README.txt     this file

COLUMNS
  query_shape    spaced | compacted
  variant        whole_spaced | whole_compacted | head2_spaced | head2_compacted |
                 head3_spaced | head3_compacted. head-k is generated only where the name
                 has MORE than k words, so a short name's "first two words" is never just
                 the whole name counted twice.
  outcome        match (>= 0.92) | possible_match (>= 0.72) |
                 near_miss_disclosed (>= 0.5) | silent (< 0.5)

METHOD
Every score comes from a production name-similarity matcher, bundled from its own source
rather than reimplemented for this study, at its own thresholds. No model is involved.
Harness: scripts/research/compacted-name-study.mjs in the Aervik repository.

SCOPE AND LIMITS
97 names, one list, one matcher. Entity records only — person-shaped names were excluded
before measurement. The six spellings are plausible, generated forms, not observed user
input; we have not measured what real users type, and that distribution is what turns
this rate into a risk.

SOURCE AND LICENCE
Names from the Consolidated Screening List, International Trade Administration,
https://www.trade.gov/consolidated-screening-list — a work of the U.S. Government in the
public domain. ITA is cited as the source and does not endorse this analysis.
This dataset is published under CC BY 4.0. Attribution: Aervik Labs.

CORRECTIONS
Errors are corrected on the published page with the date and what changed, and noted in
the changelog at https://aerviklabs.com/changelog/.