acceptodds
Under review as a conference paper at ICLR 2027

Namesakes on the List: Composition Amplification and Target-Decoy Calibration in Sanctions Name Screening

Abstract

Sanctions screening compares each customer's name with every entry on a watchlist of tens of thousands of listed persons and their aliases, and raises an alert when any entry scores above a similarity threshold. We show that this one-to-many design, more than the pairwise matcher, decides who bears the false alerts. Screening non-sanctioned politically exposed persons from 36 countries against a consolidated list of about 40,000 sanctioned persons, a standard fuzzy matcher at a common operating threshold alerts on people from the countries most represented on the list more than an order of magnitude more often than on others. We formalize this composition amplification: a customer's false-alert probability grows with the number of listed namesakes, so no single score threshold can equalize alert burden across populations whose names collide with the list at different rates. We then adapt target-decoy competition from proteomics: each customer is screened against the real list and against decoy lists of non-sanctioned persons drawn to match the list's composition, and is flagged only when the real list wins. When listed and decoy names are exchangeable, the resulting per-customer p-value bounds every customer's false-alert probability, and hence every group's, without using the customer's demographic attributes. On real lists the guarantee is only as good as the decoys: we show how composition misspecification shifts false alerts between groups rather than removing them, and quantify the recall the calibration costs for listed persons with common names. We evaluate string, embedding and large-language-model matchers and release the benchmark.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
sanctions screening, anti-money laundering, entity matching, algorithmic fairness, calibration, target-decoy competition, conformal inference, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VSZYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets at the demographic-unfair false alert problem introduced by one-to-many name screening, by adapting target-decoy competition from proteomics, i.e., comparing scores of the real list and the generated decoy lists. The idea is smart and straightforward. The paper also provide some theoretical guarantees. However, my biggest concern is that the proposed decoy constructions do not achieve the exchangeability assumption required by the theoretical results.

Critical strengths
  1. This paper provide a clear problem formulation to the name screening and unfair false alert problem. The authors also provide formal theoretical results under an explicit exchangeability assumption, covering composition amplification, performance guarantees, misspecified decoy consequences, etc.
  2. The experiments evaluate several matcher families, multiple decoy sources and generation strategies, matched alert volumes, detection recall, and multiple matcher families, providing a useful view of the trade-offs involved.
Critical weaknesses

CW1. The theoretical guarantees rely on the exchangeability assumption that practical decoys are hard to satisfy. The empirical results show substantial deviations from nominal false-alert levels for several practical decoy constructions, roughly between 2.3α and 11.3α. Thus, the formal guarantees do not directly apply to the proposed decoy constructions. CW2. The per-country threshold baselines simply assign a threshold for each country. They are not sufficiently strong for evaluating the value of demographic-blind calibration. Stronger country-aware baselines shoule be evaluated. CW3. The demographic-blind claim needs to be further clarified. During decoy constructions, using the country information helps to meet the exchangeability assumption but violates the demographic-blind cliam; not using the country information remain demographic-blind but can lead to severe violation of the exchangeability assumption.

Questions for authors
  1. Can the authors provide stronger evidence that the proposed decoy generator/strategy satisfies the exchangeability assumption?
  2. How does TDC compare to stronger country-aware baselines?
  3. Can the authors clarify CW3?
Reviewer_wk4hOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies unequal false-alert burden in one-to-many sanctions name screening. It relates alert probability to expected namesake counts under a random-list model. It then compares each customer against the real list and decoy lists to obtain a rank p-value. Validity requires exchangeability. Experiments use 40,755 non-listed persons from 36 countries and a list of 39,678 sanctioned persons. They show large country disparities and reduced disparity under some budget-matched TDC rules. Practical decoys exceed the nominal false-alert level. The paper also studies recall loss for common names.

Critical strengths
  1. The empirical burden audit is useful. At threshold 85, country alert rates range from 38.8% to 0.13%. The composition predictor reaches Spearman correlation 0.88 (Sec. 5.1). Appendix H reports bootstrap intervals and an identity-leakage check. Together, these results show a substantial disparity in the studied population and motivate evaluating the full screening list.

  2. The rank argument gives a clear conditional guarantee under exchangeability. The practical experiments also show a useful trade-off without establishing nominal validity. At roughly 12% alert volume, list-derived TDC reduces Gini from 0.503 to 0.351, with recall changing from 76.3% to 75.4% (Table 2). This supports further work on decoy construction.

Critical weaknesses
  1. Some theoretical claims exceed their proofs. Appendix A.3 compares each group with its exchangeable reference. Corollary 4 instead compares the group rate with the overall budget B. These are different baselines. Two groups can both be under-covered and have different alert rates, so one lies below their overall average. Proposition 5 has another gap: its proof conditions on S0 = sigma, but the statement defines sigma as the person’s own-entry score. Equation 1 takes the maximum over all entries. The binomial argument also concerns this TDC rule. It does not establish the Introduction’s impossibility claim for any calibrated name-only rule.

  2. The calibration experiments are not reported consistently enough to check. Section 5.2 describes disjoint random blocks, while Appendix E describes independent inclusion of sampling units. Section 5.2 reports 1.17 alpha for the worst country under exchangeable decoys; Table 1 gives 0.9 alpha. Pool-synthetic decoys give 3.21 alpha in Sec. 5.2 and 2.9 alpha in Appendix H at K = 199. These may be different runs, but the mapping is missing. The Abstract also claims LLM evaluation, yet no LLM result or model identity is reported. These gaps limit confidence in the validation and matcher-generality claims.

Questions for authors
  1. How does Corollary 4 handle this analytical counterexample? Set K = 199 and alpha = 0.05. Decoy scores are uniform on [0,1]. Two equally weighted groups have real-score CDFs t^2 and t^3. Both dominate the decoys. Equation 4 gives alert rates 9.73% and 14.20%, with overall B = 11.96%. The first under-covered group is below B. Scores can be rescaled above the evidence floor.

  2. Does Proposition 5 require the own-entry score to equal the whole-list maximum? Please state the needed conditions and distinguish the TDC-specific result from the claimed limit for all calibrated name-only rules.

  3. Please provide a run-to-result mapping for Sec. 5.2, Table 1, and Appendices G–H. Include list construction, list size, B, K, seeds, and partition averaging. This is needed for verfication of the calibration results.

  4. Which LLM was evaluated, and where are its results? Please provide its operating point and the same recall and disparity metrics used in Table 3 for a direct comparision.

Reviewer_NyVoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies disparities in false positive rates in sanctions name screening, showing that list composition can amplify errors across demographic groups even when using the same matching algorithm and threshold. Experiments on 40,755 non-sanctioned politically exposed persons across 36 countries reveal substantial disparities, with false positive rates ranging from 0.13% to 38.8%. The authors propose Target-Decoy Competition (TDC), which compares customer names against both real sanctions lists and decoy lists, generating alerts only when the real list wins. Under an exchangeability assumption, TDC provides false-alert calibration without explicitly using demographic attributes. The paper further analyzes robustness to imperfect decoy construction and the trade-off between false-alert reduction and recall, evaluating multiple matching methods and decoy-generation strategies.

Critical strengths

s1: Addresses an important real-world fairness problem, demonstrating substantial disparities in false positive rates across nationalities.

s2: Provides theoretical guarantees for TDC under exchangeability and systematically analyzes robustness and recall trade-offs.

Critical weaknesses

According to ICLR guidelines, the main paper is limited to 9 pages. Since this paper violates the page limit, it should be rejected.

Questions for authors

N.A.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.