FACIAL-RECOGNITION CLAIM REVIEW
Yenra | https://yenra.com/face-recognition-technology-aclu-fights/
Updated October 9, 2026

Purpose: Record the evidence behind a proposed deployment or accuracy claim.
Use a separate copy for each system/version and operating purpose.
Do not enter personal photographs, biometric templates or identifying records.
This is an evaluation worksheet, not a legal determination or product test.

1. THE PROPOSAL
Claim in plain language:
Who made it, when, and where (source URL/document/page):
Intended task and users:
Decision made after a result:
1:1 verification / 1:N identification / other (define):

2. THE MEASUREMENT
Algorithm and version:
Capture equipment and image conditions:
Gallery size and inclusion rules (for 1:N):
Threshold and score definition:
Test date, population and sample sizes:
Same-person/mated error metric and value:
Different-person/non-mated error metric and value:
Units and denominator for EACH rate:
How poor images and repeated attempts were handled:
Subgroup results and uncertainty:
Does the test represent local conditions? Evidence:

3. WORKFLOW AND GOVERNANCE
Purpose and applicable authority to verify:
Who adds, corrects and removes gallery records?
Retention for images, templates and unsuccessful searches:
Who has access, and what access is logged?
Who reviews candidates, using what independent evidence?
How can an affected person challenge an error?
Local acceptance criteria and stop/review conditions:

4. ILLUSTRATIVE QUEUE CALCULATION (FICTIONAL, NOT A FORECAST)
100,000 single searches; 10 enrolled targets.
Assumed correct-alert rate among targets: 90% (0.90).
Assumed false-alert rate among 99,990 non-target searches: 0.1% (0.001).
Expected correct alerts = 10 * 0.90 = 9.
Expected missed targets = 10 * 0.10 = 1.
Expected false alerts = 99,990 * 0.001 = 99.99 (about 100).
Expected total alerts = 9 + 99.99 = 108.99 (about 109).
Expected correct share = 9 / 108.99 * 100 = 8.26%.
Expected counts can be fractional; daily counts are whole numbers.
Assumes one alert outcome per search; omits wrong additional candidates in
searches for enrolled people. Uses a SEARCH-LEVEL false-alert rate, not a
pairwise false match rate. It does not model staffing time or local performance.

5. FOLLOW-UP
Unanswered question:
Evidence needed:
Responsible person/role:
Review date or trigger (version/camera/gallery/threshold/purpose change):
Current conclusion and its limits:

Definitions and current test results:
https://pages.nist.gov/frvt/html/frvt11.html
https://pages.nist.gov/frvt/html/frvt1N.html
Article above supplies historical context and further sources.