open-standards/openpay/crates/op-screening/src/screener.rs
Transaction Science 2d7de0396c
All checks were successful
standards workspaces build + test / which workspaces this change touches (pull_request) Successful in 10s
standards workspaces build + test / macOS energy backends (pull_request) Has been skipped
repository checks / lockfiles, tests in CI, workflow hygiene, spec and section references, public text (pull_request) Successful in 1m4s
standards workspaces build + test / openpay (openpay, openpay, linux-arm64) (pull_request) Successful in 6m13s
openpay: a PEP match is recorded as a PEP match, not an OFAC sanctions hit
op-connect's PEP index tagged every entry `SanctionsList::OfacConsolidated`,
because the enum had no place for a PEP list. Its comment said the operator
surface "should not surface this". But a PEP index is a `Screener`, and every
screen appends `source_list.label()` to the screener's signed, hash-chained
audit log. So each PEP match was written into the tamper-evident compliance
record as "OFAC Consolidated": a sanctions designation (block, report) where
there was a due-diligence flag. An adverse-media index had the same problem:
it too could only tag its entries with a sanctions authority.

- `SanctionsList` is now `WatchList`, the term `ScreenerConfig::watchlist_filter`
  already used. It gains `Pep` ("PEP list") and `AdverseMedia` ("Adverse
  media"), and `is_sanctions()`, which is false for exactly those two. Its
  docs say why a match must never be recorded under a list it did not come
  from. A deprecated `SanctionsList` alias keeps the old name compiling.
- `build_pep_index_from_names` tags its entries `WatchList::Pep`.
- Serialized forms are unchanged: the enum's variant names stay the same, so
  stored indexes still load. No published vector or doc carries the label.

Tests:
- `a_pep_match_is_audited_as_a_pep_match_not_a_sanctions_match` screens a PEP
  fixture and checks that the audit chain verifies and that its row names
  "PEP list".
- `only_the_authorities_are_sanctions_lists` and `every_list_has_its_own_label`
  pin the kinds and labels. An exhaustive match makes a new list fail to
  compile until it is classified.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-15 20:47:42 -04:00

271 lines
8.8 KiB
Rust

//! Top-level screening orchestration.
//!
//! [`Screener`] is the type operators actually wire into their payment
//! flow. It owns the index, the config (threshold, max results, list
//! filter, optional DOB / address gating), and an [`crate::audit::AuditLog`]
//! that every screen call appends to. The result classification
//! ([`ScreenDecision::Clear`] / [`Hit`](ScreenDecision::Hit) /
//! [`AmbiguousNeedsReview`](ScreenDecision::AmbiguousNeedsReview))
//! drives the caller's downstream action.
use std::collections::HashSet;
use std::sync::Mutex;
use chrono::NaiveDate;
use serde::{Deserialize, Serialize};
use crate::audit::AuditLog;
use crate::error::Result;
use crate::lists::{Address, Identification, WatchList};
use crate::matching::{MatchScore, screen};
use crate::storage::SanctionsIndex;
/// Tunable knobs for [`Screener`].
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ScreenerConfig {
/// Combined-score threshold below which a candidate is dropped.
/// Default in production deployments: `0.85`.
pub threshold: f32,
/// Hard cap on the number of hits returned to the caller.
pub max_results: usize,
/// If true and the request carries a DOB, hits whose entity DOB
/// disagrees by more than ±1 year are demoted to
/// [`ScreenDecision::AmbiguousNeedsReview`].
pub screen_dob: bool,
/// If true, address country mismatch demotes the result similarly.
pub screen_address: bool,
/// If non-empty, only consider hits whose `source_list` is in the set.
pub watchlist_filter: HashSet<WatchList>,
/// Above-`threshold` but below `hit_threshold` lands in
/// `AmbiguousNeedsReview`. Defaults to `0.95`.
pub hit_threshold: f32,
}
impl Default for ScreenerConfig {
fn default() -> Self {
Self {
threshold: 0.85,
max_results: 25,
screen_dob: false,
screen_address: false,
watchlist_filter: HashSet::new(),
hit_threshold: 0.95,
}
}
}
/// What the operator wants to screen.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ScreenRequest {
/// Name to screen. Required.
pub name: String,
/// Date of birth, if known. Used when `screen_dob` is on.
pub dob: Option<NaiveDate>,
/// Address, if known. Used when `screen_address` is on.
pub address: Option<Address>,
/// Other identifications to cross-screen (passport, national ID).
pub additional_ids: Vec<Identification>,
}
/// Three-way result.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
pub enum ScreenDecision {
/// No hits at or above threshold.
Clear,
/// At least one hit at or above `hit_threshold`.
Hit,
/// Hits between `threshold` and `hit_threshold`. Human review required.
AmbiguousNeedsReview,
}
/// The output of [`Screener::screen`].
#[derive(Debug, Clone)]
pub struct ScreenResult {
/// All hits that survived filtering, sorted desc by score.
pub hits: Vec<MatchScore>,
/// Top-line decision.
pub decision: ScreenDecision,
/// Audit-log row id that recorded this call.
pub audit_id: String,
}
/// Stateful screening engine.
///
/// Wrap the audit log in a `Mutex` because every screen call is a
/// write into a chained, signed log; we need exclusive access for
/// the duration of the append. The index itself is read-only.
pub struct Screener {
index: SanctionsIndex,
config: ScreenerConfig,
audit_log: Mutex<AuditLog>,
}
impl Screener {
/// Build a new screener.
#[must_use]
pub fn new(index: SanctionsIndex, config: ScreenerConfig, audit_log: AuditLog) -> Self {
Self {
index,
config,
audit_log: Mutex::new(audit_log),
}
}
/// Borrow the audit log (e.g. for verification).
pub fn with_audit_log<R>(&self, f: impl FnOnce(&AuditLog) -> R) -> R {
let log = self.audit_log.lock().expect("audit-log mutex poisoned");
f(&log)
}
/// Active configuration.
#[must_use]
pub fn config(&self) -> &ScreenerConfig {
&self.config
}
/// How many entities the index covers.
#[must_use]
pub fn index_len(&self) -> usize {
self.index.len()
}
/// Screen a single request.
///
/// Always appends to the audit log, even on `Clear` — operators
/// rely on the negative-evidence trail when regulators inspect.
pub async fn screen(&self, request: &ScreenRequest) -> Result<ScreenResult> {
let raw_hits = screen(&request.name, &self.index, self.config.threshold);
// Optional list filter.
let filtered: Vec<MatchScore> = if self.config.watchlist_filter.is_empty() {
raw_hits
} else {
raw_hits
.into_iter()
.filter(|h| {
self.config
.watchlist_filter
.contains(&h.entity.source_list)
})
.collect()
};
// Optional DOB / address discrimination. We don't drop hits —
// a mismatched DOB on a perfect-name match is still ambiguous;
// we just demote the per-hit confidence so the overall decision
// gates into AmbiguousNeedsReview.
let mut hits = filtered;
if self.config.screen_dob {
if let Some(req_dob) = request.dob {
for h in &mut hits {
if let Some(ent_dob) = h.entity.dob {
let delta = (ent_dob - req_dob).num_days().abs();
if delta > 366 {
// ±1 year. Scale the score down by 20%.
h.score *= 0.8;
}
}
}
}
}
if self.config.screen_address {
if let Some(addr) = request.address.as_ref() {
for h in &mut hits {
let any_country_match = h
.entity
.addresses
.iter()
.any(|a| a.country == addr.country);
if !any_country_match && !h.entity.addresses.is_empty() {
h.score *= 0.9;
}
}
}
}
hits.sort_by(|a, b| b.score.partial_cmp(&a.score).unwrap_or(std::cmp::Ordering::Equal));
hits.truncate(self.config.max_results);
let decision = if hits.is_empty() {
ScreenDecision::Clear
} else if hits[0].score >= self.config.hit_threshold {
ScreenDecision::Hit
} else {
ScreenDecision::AmbiguousNeedsReview
};
let audit_id = {
let mut log = self.audit_log.lock().expect("audit mutex poisoned");
log.record(
&request.name,
&hits,
match decision {
ScreenDecision::Clear => "clear",
ScreenDecision::Hit => "hit",
ScreenDecision::AmbiguousNeedsReview => "review",
},
)?
};
Ok(ScreenResult { hits, decision, audit_id })
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::audit::AuditLog;
use crate::lists::{EntityType, SanctionedEntity};
use chrono::Utc;
fn ent(id: &str, name: &str) -> SanctionedEntity {
SanctionedEntity {
id: id.to_string(),
name: name.to_string(),
name_aliases: vec![],
entity_type: EntityType::Individual,
dob: None,
place_of_birth: None,
addresses: vec![],
nationalities: vec![],
identifications: vec![],
programs: vec![],
last_updated: Utc::now(),
source_list: WatchList::OfacSdn,
}
}
#[tokio::test]
async fn clear_when_no_match() {
let idx = SanctionsIndex::build(vec![ent("1", "John Smith")]);
let s = Screener::new(idx, ScreenerConfig::default(), AuditLog::new());
let r = s
.screen(&ScreenRequest {
name: "Totally Unrelated Person".into(),
dob: None,
address: None,
additional_ids: vec![],
})
.await
.expect("ok");
assert_eq!(r.decision, ScreenDecision::Clear);
assert!(r.hits.is_empty());
}
#[tokio::test]
async fn exact_match_is_hit() {
let idx = SanctionsIndex::build(vec![ent("1", "John Smith")]);
let s = Screener::new(idx, ScreenerConfig::default(), AuditLog::new());
let r = s
.screen(&ScreenRequest {
name: "John Smith".into(),
dob: None,
address: None,
additional_ids: vec![],
})
.await
.expect("ok");
assert_eq!(r.decision, ScreenDecision::Hit);
assert_eq!(r.hits.len(), 1);
}
}