Compliance & Screening

Screening isn't a moment.
It's a subscription.

Seven deterministic, rule-table-driven engines — restricted/denied party, country embargo, UFLPA forced labor, end-use, end-user, anti-boycott, military end-use — run on every shipment and every party. Qubere keeps re-screening what it already cleared when the party or the watchlist changes, and every hit is immutable evidence.

7
concurrent screening engines · never an LLM
21+
government lists ingested on a governed schedule
0
hits ever silently rewritten

The problem

You screened once, at onboarding. The world moved.

What a trade-compliance team will tell you:

"We screen against OFAC at onboarding. We do not re-screen when OFAC updates the SDN list — which is basically weekly."

Stale clearances

"A name match in our old tool is treated as a legal-identity match. Analysts waste days clearing false positives with no way to record why."

False positives

"When we clear a party, that's a status field someone can overwrite. There's no immutable record of the hit that was there before."

No audit trail

How Qubere addresses it

Deterministic verdicts. Immutable results. Continuous re-screening.

  • Every verdict comes from a rule table and a scorer you can inspect — not a model that sounds confident.
  • Results are frozen. Reviewer judgment (False Positive / Blocked / Approved) is a separate 1:1 record — a clean run never erases an earlier hit.
  • RDPS re-screens cleared parties automatically when their data changes or the watchlist changes — with a drill-down to exactly which parties and why.
  • A tenant can layer its own embargo rules — which can only add a HIT, never manufacture a CLEAR.
  • License determination refuses to fabricate a control-list outcome it can't back with data.
  • Scheduled sweeps now watch AD/CVD case status and bond sufficiency continuously, and a new Post-Entry hub adds ISF 10+2, prior disclosure, CBP reconciliation, and filer-admin health — plus a per-shipment reasonable-care checklist.
  • A deterministic champion/challenger re-ranker and a free-text entity extractor screen unstructured content (document narratives, OCR'd text) with the same immutable pipeline — still no model hosting, still fully inspectable.
  • The list-ingestion pipeline watches its own health — a source going stale or failing opens a tracked incident (acknowledge / resolve / suppress), not a silently missed weekly OFAC refresh.

Restricted / denied-party screening

A name match is never legal-identity proof.

Every party is screened against SDN, Consolidated non-SDN, DPL, ISN, SSI, FSE, PLC, NS-MBS, DOW JONES — exact, raw-word, and a phonetic shortlist feeding a fuzzy scorer, plus an independent red-flag keyword check. Results are immutable; the disposition is separate.

  • Phonetic algorithm (Double Metaphone / Metaphone2) selectable per account
  • An earlier HIT + its "false positive" note stays on the record forever
  • Champion/challenger re-ranker replays every match with an alternate weighting — observed, never applied, until a governance decision promotes it
  • Screening-on-create is an account-level toggle (off by default); a periodic scheduled review still catches any party whose result has simply aged out
  • State/postal address fields now feed the score directly — STATE_MATCH / POSTAL_MATCH evidence, parsed out of Dow Jones's composite city field instead of discarded
  • A clean pass never reports CLEAR if the list it just checked against was itself stale or had failed its last refresh — falls back to REVIEW_REQUIRED with the reason, a real HIT still wins
  • The Screening queue is keyed by party — one row per party, with a Quick Review drawer showing the screened-as name, matches and earlier decisions; party decisions include Not a Party, Clear All for a name, and Proceed Anyway on shipment checks
  • The ad-hoc screening entry form now offers HERE-backed live address autocomplete — fewer typo'd addresses reaching the fuzzy matcher, fewer spurious STATE_MATCH / POSTAL_MATCH misses and false hits
How to demo
Comply → Screening & Due Diligence → Screening → Party Screening. Open a party with two runs — an earlier HIT with a disposition and a later CLEAR. Both visible. "A real hit can never be status-fielded away."
Comply/Party Screening
AC
Alice · Acme
Screening history — immutable
2026-06-02 · name+address passHIT · score 0.91  SDN "GUANGZHOU RUIXIN"
   ↳ dispositionFALSE POSITIVE — diff. registration #, analyst J.K.
2026-08-25 · ref-data republish rescreenCLEAR · score 0.44
Formal overridenone · create (compliance.override)
Double Metaphone shortlistscoreDpsMatchComplianceKeywordRulechampion/challenger re-rankSTATE_MATCH / POSTAL_MATCH

RDPS · continuous / reverse screening

The watchlist changed this week. Did your cleared parties get re-checked?

A delta-impact dispatcher reacts to a specific reference-data change within minutes; a full-population dispatcher re-screens the whole account on a schedule. Every outcome records whether the result worsened and a deterministic transition type.

  • NEW_HIT / ESCALATED / RISK_REDUCED / CLEARED / UNCHANGED
  • Preview Impact: what a change would match today — records nothing
  • Hourly sweep catches watchlist entities that expired while still active in every feed
How to demo
Comply → Controls & Monitoring → Monitoring. Open a reference-data change set → Impacted Parties: exactly which parties were re-screened because of it.
Comply/Monitoring
AC
Alice · Acme
Open RDPS alerts
3
Parties monitored
2,184
Re-screened (24h)
47
Last full sweep
04:00 UTC
Change set OFAC-SDN 2026-08-28 · 62 entities added → Impacted Parties (2)
PartyBeforeAfterTransition
Sever­stroy OOOCLEARHITNEW_HIT
Baltic Marine LtdPOTENTIALHITESCALATED
rdps-delta-impact-dispatch · every 10mreference-data-expiry-sweep · hourly

RPS · matcher quality

Not "trust the score" — prove it, from your own reviewed history.

Every reviewed HIT-tier disposition on file — confirmed match or false positive — feeds a precision / false-positive-rate readout, plus a threshold sweep showing what recall and precision would have been at every other nameThreshold cutoff. Same ground truth an analyst already produced by doing their job; nothing new to label.

  • Precision and false-positive rate, computed live, per account
  • Threshold sweep: surfaced / precision / recall at each cutoff, against real reviewed matches
  • Feeds the champion/challenger re-ranker governance decision — a low reviewed count just means "come back after more reviews," never a fabricated confidence number
How to demo
Comply → Controls & Monitoring → Screening Quality. Point at the threshold sweep: "this is what tightening your threshold would have actually cost you in recall, on your own data."
Comply/Screening Quality
AC
Alice · Acme
Reviewed matches
1,406
HIT-tier reviewed
312
Precision
88.1%
False-positive rate
11.9%
Threshold sweep
ThresholdSurfacedPrecisionRecall
0.7048881.2%97.4%
0.80 current31288.1%91.0%
0.9020194.5%76.6%
computeMatchQualityMetricscomputeThresholdCurvereviewed dispositions only

Community screening · batch

A 4,000-row acquired supplier list, screened before lunch.

Submit a batch (manual or file upload); Qubere runs the canonical RPS and embargo engines against every row — it orchestrates, it never re-implements matching. A denied-party match and a red-flag keyword hit stay independent findings.

  • Per-row category: CONFIRMED_MATCH / POTENTIAL / RED_FLAG_ONLY / PAL_SUPPRESSED
  • A valid pre-approval short-circuits — its own status, not an ordinary CLEAR
  • Export the whole run as CSV / XLSX
How to demo
Comply → Screening & Due Diligence → Community Screening → upload a CSV of ~20 parties (include one obvious SDN-style name). Open the run; walk the per-row labels; export.
Comply/RUN-2026-0814
AC
Alice · Acme
#PartyRPSRed flagCategory
0007Nanjing Optics Import ExportMATCH—CONFIRMED_MATCH
0114Dalian Hengli Group Co.0.78—POTENTIAL_DENIED_PARTY
0233Meridian Defense Systems LLCNO_MATCH"military end use"RED_FLAG_ONLY
0341Hansa Components GmbH——PRE_APPROVED_REUSE
uniqueness key: (runId, rowNumber)most-severe-wins collapse for legacy consumers

Network & ownership intelligence

A direct match is the first question. "Who is this entity related to?" is the second.

Qubere persists Dow Jones' relationship rows (director/officer/associate/family/beneficial-owner) as a bounded graph and evaluates two things off every direct match: an ownership-attribution rule (OFAC's 50% rule, up the chain, never fabricated when a source lacks a percentage) and a relationship review signal — a director/officer/associate of a sanctioned entity, worded and colored as a review flag, never a hit.

  • Priority-ordered triggers — sanctioned director/officer, conflicting ownership data, an unresolved network relationship, incomplete ownership data — auto-open an Enhanced Due Diligence case
  • Dedicated EDD workspace tab: assign, annotate, decide — not a side spreadsheet
  • Vessel identity matching (IMO-first, with historical alias support) feeds the same graph for maritime shipments
  • Policy config (traversal depth, ownership threshold, jurisdiction scope) goes DRAFT → IN_REVIEW → APPROVED → ACTIVE → RETIRED through a platform-admin UI, never a raw DB edit
  • Relationship-staleness threshold honors an account's own override everywhere it's checked — the live screening path and the dashboard health metric now read the same governed number
How to demo
Open a party with an OFAC hit that also has a related director/officer — show the Network Findings badges and the read-only relationship graph. Then Compliance → EDD Cases — show the case that opened itself from that finding.
Comply/EDD Cases
AC
Alice · Acme
Relationship graph (read-only)
Director/officer — Wen Li CaoSANCTIONED_DIRECTOR_OR_OFFICER
Beneficial owner chain — 3 hopsOWNERSHIP_DATA_INCOMPLETE  no % reported by source
EDD-2026-0091 — opened automatically
TriggerSANCTIONED_DIRECTOR_OR_OFFICER
AssignedD. Osei · status IN_REVIEW
Dow Jones relationship graphOFAC 50% ownership-attribution rulevessel IMO alias matchingNetworkScreeningPolicy · ACTIVE

License determination

It will not guess. That is the pitch.

A deterministic engine (never an LLM) runs classification-format validation, tri-state end-use / end-user handling, and license-exception evaluation as pure functions — now with real jurisdiction-specific evaluators (US country-group, EU, UK/E1, RPL, plus a generic rule-graph fallback covering every other country in the reference table). US control-list data, EU classification flags, and EU license-exception pairing are loaded and published today — real verdicts, not just the fail-safe. For any capability whose reference dataset still isn't published, it never fabricates a LICENSE_REQUIRED / NO_LICENSE_REQUIRED outcome.

  • Live Regulatory Dataset Health report across 8 jurisdiction capabilities — READY / MISSING / INCOMPLETE, named component
  • Key capabilities including US_CONTROL_DATA_READY, EU_CLASSIFICATION_DATA_READY, and EU_PAIR_EXCEPTION_DATA_READY report READY today; ENC 740.17(b) tables, product-PCA assignment, and generic LD tiers 2–4 still honestly report MISSING
  • Returns RULE_DATA_UNAVAILABLE / INCOMPLETE / REVIEW_REQUIRED with full evidence when a capability isn't ready
  • Any sensitive end-use / end-user flag hard-stops to REVIEW_REQUIRED
  • On a positive determination: full utilization ledger, allocations, expiry alerts
How to demo
Comply → Licenses & Export Controls → Run determination. Hit the dataset-health endpoint first to show which jurisdiction capabilities are READY, run a real determination against US control-list data to show a live verdict, then show the fail-safe outcome and evidence for a capability that isn't ready yet. "An engine that says 'here's what I couldn't verify' beats one that guesses — and here's proof it doesn't guess when it does have the data either."
Licenses & Export Controls/Determination
AC
Alice · Acme
⚑ REVIEW_REQUIRED
Reasonsensitive end-user flag · hard stop
Evidence — what was evaluated
Classification format (ECCN)valid
End-use conditions2 true · 1 unknown
License-exception claimnone claimed
Country control-rule datasetRULE_DATA_UNAVAILABLE — not ingested

No control-rule dataset → the engine will not output a licensing verdict. This is fail-safe by design.

Audit, notifications & the partner API

Every screen is provable, and every screen can be a service.

Compliance events queue durable notifications through an outbox/dispatcher (never sent inline, so a mail outage can't lose a hit alert). Every action writes an immutable AuditLog entry. External systems hit the same read/rescreen behavior by API key.

  • Overview / Screening / Review Queue / Audit History workspace
  • /api/v1/screening/* · scoped (read vs screen) · idempotent · source: "API"
  • Every screen is a metered, invoiceable event (see Billing deck)
  • Historical replay: reconstruct exactly what a past screening claimed, and diff it against today's classification rules and policy — a pure read, it never re-screens or re-opens a case
  • A rule-change to sanctions consequence classification that isn't paired with a version bump fails the test suite before it ships — the same discipline the network policy lifecycle already enforces
How to demo
Comply → Evidence & Reporting → Audit History — the timestamped event stream and export. Then Admin → Settings — the API key panel and scopes.
Comply/Audit History
AC
Alice · Acme
TimeActionEntityActorSrc
12:41:07RPS_RESCREENPTY-4471J. KesslerUI
12:39:55EMBARGO_SCREENSHP-2026-000188systemPIPE
11:22:03RESTRICTED_PARTY_SCREENPTY-8890api-key: acme-prodAPI
09:15:40COMPLIANCE_FORMAL_OVERRIDE_CREATERES-11902D. OseiUI
ComplianceNotification → Dispatcher (outbox)embargo.read / embargo.screen scopes

What's real today

The lists we actually screen against — live and ingesting.

✓ OFAC SDN + Consolidated non-SDN (~20k entries, streamed)
✓ BIS Consolidated Screening List (10 agency lists)
✓ DHS UFLPA Entity List + CBP WRO/Findings (independent forced-labor checks)
✓ CBP CROSS rulings
✓ Dow Jones full feed (provider lineage, multi-valued aliases)
✓ UN Security Council + EU + UK (OFSI) consolidated sanctions lists
✓ Switzerland (SECO), Australia (DFAT), Canada (Consolidated + Public Safety terrorist entities), Japan (METI), Singapore (MAS) national sanctions/designation lists
✓ World Bank debarred firms, EU Air Safety List, FDA Debarment List, FBI Wanted, SAM.gov Exclusions
✓ FinCEN Special Measures (Sec. 311/9714/9915) + FCC Covered List
✓ Network & ownership intelligence — relationship graph, ownership attribution, auto-opened EDD cases
✓ License determination reference data — Commerce Control List, USML/country export controls, generic control numbers, EU classification & license-exception pairing loaded and published; real READY verdicts, not just the fail-safe
○ PEP list / corporate registry — not yet; ownership % depends on a source that doesn't send one today
○ Vessel registry (owner/operator/ISM manager/flag) — matching engine is live, no vendor feed contracted yet
○ SEC FCPA enforcement actions — free-text filings, no fixed schema; tracked as a scoped gap, not guessed by an LLM
○ ENC 740.17(b) rule tables, product-PCA assignment, generic LD tiers 2–4 — license determination still honestly reports these MISSING

Sell what's there: the denied-party, embargo, forced-labor, and network/relationship coverage is real and continuous, spanning the US, UN, and fifteen additional national/multilateral sanctions and exclusion sources. Say plainly where it stops.

Objection handling

Every hard question, answered directly.

"Is this an AI making sanctions calls?"No. Every screening verdict is a rule table and a scorer you can inspect. The AI in the platform does analysis inside stages — it never decides a hit.
"Do you do PEP / beneficial ownership?"No standalone PEP list or corporate-registry lookup — still a known gap. But ownership-attribution off a direct hit is real: OFAC's 50% rule, evaluated up the chain, plus a director/officer review signal from the relationship graph. Honestly scoped — Dow Jones doesn't send an ownership percentage today, so most cases surface as "ownership data incomplete," not a computed number, and an incomplete/conflicted/unresolved finding auto-opens an EDD case for an analyst. Sell what's there: denied-party, embargo, forced-labor, and network/relationship coverage is real and continuous.
"License determination sounds like it does nothing."Not anymore — US control-list data and EU classification/license-exception data are loaded and published, so it returns real verdicts there today. For everything else it's still the honest thing: it refuses to output a fake control-list answer when the data to back it isn't loaded. "REVIEW_REQUIRED, here's what I checked" is worth more than a guess.

The close

"The watchlist changed this week. Ours re-screened your cleared parties — and can show you which ones, and why."

Immutable results, separate dispositions, continuous re-screening. Continuous assurance, not a onboarding checkbox.

Next step: send us an anonymized slice of your party master. We'll run it live — RPS, embargo, UFLPA — and walk the findings with your compliance lead.

Full feature list and talk track: docs/apps/customs/sales/compliance-and-screening.md