Sanctions screening for company and vessel names against OFAC, CSL, UN and EU lists, with published, blind-tested accuracy.
See the codeSee what people are saying
README
Cascade · Screening
Which name matcher suffices, at which threshold, measured on your own alert history. Nothing of yours goes up: the lists come down, your alerts stay on your machine.
It also screens a list of counterparties (companies, vessels, IMO numbers) against five public
sanctions lists, on your machine: see npm run cribler below.
Name screening (sanctions, PEP, internal lists) raises alerts; most of them are false, and nobody can say why the threshold sits at 85 rather than 90 except "that is the vendor's setting". This tool measures it: several matchers, from exact to phonetic to a local embedding model, are run over the alerts your analysts already dispositioned, and each matcher × threshold cell is read for recall on confirmed matches, false-alert rate and alerts per thousand screenings, with its count and its interval, or not at all.
It is the second tool of the Cascade suite. The first, cascade-routing, measures which extraction tier suffices per field. Same method, same seal, same key.
| Command | What it does, in the order that makes sense |
|---|---|
npm ci --ignore-scripts | install exactly the versions the lockfile pins, and run no install script from any dependency; the only command besides listes -- --fetch that needs the network |
npm run listes [-- --fetch] | the five public sanctions lists: OFAC SDN and OFAC consolidated (non-SDN), the US Consolidated Screening List (its Commerce and State lists; its Treasury rows come from the two OFAC files), UN consolidated, EU consolidated (FSF), downloaded into data/ with a committed manifest (source, date, sha256, entry count); without the flag it reports what is on disk and touches nothing. The lists download to your machine, and nothing of yours is sent |
npm run poids [-- --fetch] | the embed tier's weights: four files pinned by bytes and sha256, fetched only by this command (never during install or tests, refused offline) into data/models/: absent weights make an absent tier, named, not a surprise download |
npm run test | types, the README blocks, the licence inventory, and the suite. Start here; it runs with the network cut |
npm run measure [-- --yes-overwrite] | the public measure: every tier at every threshold on pairs we authored (hard negatives included) plus declared synthetic variants, sealed into releve-public.json and readable in RELEVE-PUBLIC.md: the record the catalogue requires, and it refuses to overwrite a sealed one without the flag |
| `npm run measure:yours -- --alerts= [--screened= | --volume=N]` |
npm run cribler -- --names=<csv> [--client=<name>] [--previous=<record>] | screen a list of counterparties (companies, vessels, IMO numbers) against every list on disk: each name gets its candidates, their list and score, at two levels. STRONG is measured: the lowest threshold that keeps false alerts under 5 % on the authored training pairs. POSSIBLE is structural: the method saw a precise reason to doubt (a name found inside a longer one, a legal form of another country, a distinctive word on one side only, a short word one letter apart, a vessel number on one side) and the candidate is shown for review. Rates at both levels come from a held-out set written by another hand, and are quoted in the record. With --previous, only what changed since the last record. A sealed record and a spreadsheet beside your file, no network; an English word list (web2, public domain) ships with the tool to tell a typo from another word |
npm run mesure-entites [-- --detail] | the company and vessel matcher measured on the training pair sets: found names and false alerts at the strong and possible levels, the whole curve, and with --detail every pair missed or wrongly alerted; the held-out verdict set is never detailed |
npm run releve-entites [-- --check] | the company and vessel matcher's public record, releve-entites.json: the training cells measured here, the realistic verdict every client report cites, the blind verdicts of the rounds and the two thousand-name books, each read from its source and frozen with the date and commit, then sealed with npm run sceller; --check refuses a record its sources no longer match |
npm run verdict [-- --version=vN] | the single verdict on the held-out pair set: overlap with the training sets counted first, code hashes frozen, both levels with their intervals; no pair is read (see verification/JUGE.md) |
npm run temoin-index [-- --exhaustif] | the screening index against the exhaustive comparison on the real lists (needs data/): same candidates, same scores, and the time per name |
npm run optimise -- --from=<record> --recall=<min> | the best trade-off: fewest alerts with the recall lower bound held, or --alert-budget=<N> for the highest bounded recall under a monthly alert budget |
npm run sceller -- <record.json> | seal a record: the content hash that makes a silently edited measurement fail loudly; the same content hash as cascade-routing |
npm run verify -- <report> | check that a report was issued by the holder of the suite's public key, cle-publique.pem, without asking us |
npm run licences | regenerate LICENCES.md, the licence of every shipped package; --check fails the suite when the table drifts |
Requirements
Node 24 or newer, on macOS, Linux or Windows: the whole test suite runs on all three at every push to main, on GitHub's runners (.github/workflows/tests.yml). Nobody has yet run the tool on a client's Windows machine, and this page does not claim it.
What leaves your machine
Nothing, except two explicit downloads: npm run listes -- --fetch pulls the five public lists
named above, and npm run poids -- --fetch pulls the embed tier's weights. Every other command
runs with the network cut:
CASCADE_OFFLINE=1 is honoured by the one module allowed to touch it, and a test reads every
source so that no second one appears (src/frontiere.test.ts).
What is measured, assumed, synthetic
Every rate in a report carries its n and its 95 % Wilson interval. Analyst minutes per
alert and analyst cost are assumed and declared. Perturbed variants of list entries are
synthetic, measured apart, and never merged into a measured recall. Below five confirmed
matches, no recall is quoted, and the report states why.
Trust package
For a procurement or compliance review: what the tool talks to, and when, running on a machine without network, a pre-filled vendor security questionnaire, and a pilot annex of outsourcing clauses, a draft for counsel.
Seals and signatures
Records are sealed (npm run sceller) with the same content hash as cascade-routing, and
reports are verified against the same public key, cle-publique.pem,
with npm run verify.
426 tests across 54 files, counted from the sources rather than typed here.
Licence
The same public licence as cascade-routing: non-commercial use without limit of time, a thirty-day evaluation on your own records for organisations, a commercial licence for production. See LICENSE and LICENCES.md.
Similar repos
Languages
TypeScript
98.3%
JavaScript
1.4%
Sanctions screening for company and vessel names against OFAC, CSL, UN and EU lists, with published, blind-tested accuracy.
See the codeSee what people are saying
README
Cascade · Screening
Which name matcher suffices, at which threshold, measured on your own alert history. Nothing of yours goes up: the lists come down, your alerts stay on your machine.
It also screens a list of counterparties (companies, vessels, IMO numbers) against five public
sanctions lists, on your machine: see npm run cribler below.
Name screening (sanctions, PEP, internal lists) raises alerts; most of them are false, and nobody can say why the threshold sits at 85 rather than 90 except "that is the vendor's setting". This tool measures it: several matchers, from exact to phonetic to a local embedding model, are run over the alerts your analysts already dispositioned, and each matcher × threshold cell is read for recall on confirmed matches, false-alert rate and alerts per thousand screenings, with its count and its interval, or not at all.
It is the second tool of the Cascade suite. The first, cascade-routing, measures which extraction tier suffices per field. Same method, same seal, same key.
| Command | What it does, in the order that makes sense |
|---|---|
npm ci --ignore-scripts | install exactly the versions the lockfile pins, and run no install script from any dependency; the only command besides listes -- --fetch that needs the network |
npm run listes [-- --fetch] | the five public sanctions lists: OFAC SDN and OFAC consolidated (non-SDN), the US Consolidated Screening List (its Commerce and State lists; its Treasury rows come from the two OFAC files), UN consolidated, EU consolidated (FSF), downloaded into data/ with a committed manifest (source, date, sha256, entry count); without the flag it reports what is on disk and touches nothing. The lists download to your machine, and nothing of yours is sent |
npm run poids [-- --fetch] | the embed tier's weights: four files pinned by bytes and sha256, fetched only by this command (never during install or tests, refused offline) into data/models/: absent weights make an absent tier, named, not a surprise download |
npm run test | types, the README blocks, the licence inventory, and the suite. Start here; it runs with the network cut |
npm run measure [-- --yes-overwrite] | the public measure: every tier at every threshold on pairs we authored (hard negatives included) plus declared synthetic variants, sealed into releve-public.json and readable in RELEVE-PUBLIC.md: the record the catalogue requires, and it refuses to overwrite a sealed one without the flag |
| `npm run measure:yours -- --alerts= [--screened= | --volume=N]` |
npm run cribler -- --names=<csv> [--client=<name>] [--previous=<record>] | screen a list of counterparties (companies, vessels, IMO numbers) against every list on disk: each name gets its candidates, their list and score, at two levels. STRONG is measured: the lowest threshold that keeps false alerts under 5 % on the authored training pairs. POSSIBLE is structural: the method saw a precise reason to doubt (a name found inside a longer one, a legal form of another country, a distinctive word on one side only, a short word one letter apart, a vessel number on one side) and the candidate is shown for review. Rates at both levels come from a held-out set written by another hand, and are quoted in the record. With --previous, only what changed since the last record. A sealed record and a spreadsheet beside your file, no network; an English word list (web2, public domain) ships with the tool to tell a typo from another word |
npm run mesure-entites [-- --detail] | the company and vessel matcher measured on the training pair sets: found names and false alerts at the strong and possible levels, the whole curve, and with --detail every pair missed or wrongly alerted; the held-out verdict set is never detailed |
npm run releve-entites [-- --check] | the company and vessel matcher's public record, releve-entites.json: the training cells measured here, the realistic verdict every client report cites, the blind verdicts of the rounds and the two thousand-name books, each read from its source and frozen with the date and commit, then sealed with npm run sceller; --check refuses a record its sources no longer match |
npm run verdict [-- --version=vN] | the single verdict on the held-out pair set: overlap with the training sets counted first, code hashes frozen, both levels with their intervals; no pair is read (see verification/JUGE.md) |
npm run temoin-index [-- --exhaustif] | the screening index against the exhaustive comparison on the real lists (needs data/): same candidates, same scores, and the time per name |
npm run optimise -- --from=<record> --recall=<min> | the best trade-off: fewest alerts with the recall lower bound held, or --alert-budget=<N> for the highest bounded recall under a monthly alert budget |
npm run sceller -- <record.json> | seal a record: the content hash that makes a silently edited measurement fail loudly; the same content hash as cascade-routing |
npm run verify -- <report> | check that a report was issued by the holder of the suite's public key, cle-publique.pem, without asking us |
npm run licences | regenerate LICENCES.md, the licence of every shipped package; --check fails the suite when the table drifts |
Requirements
Node 24 or newer, on macOS, Linux or Windows: the whole test suite runs on all three at every push to main, on GitHub's runners (.github/workflows/tests.yml). Nobody has yet run the tool on a client's Windows machine, and this page does not claim it.
What leaves your machine
Nothing, except two explicit downloads: npm run listes -- --fetch pulls the five public lists
named above, and npm run poids -- --fetch pulls the embed tier's weights. Every other command
runs with the network cut:
CASCADE_OFFLINE=1 is honoured by the one module allowed to touch it, and a test reads every
source so that no second one appears (src/frontiere.test.ts).
What is measured, assumed, synthetic
Every rate in a report carries its n and its 95 % Wilson interval. Analyst minutes per
alert and analyst cost are assumed and declared. Perturbed variants of list entries are
synthetic, measured apart, and never merged into a measured recall. Below five confirmed
matches, no recall is quoted, and the report states why.
Trust package
For a procurement or compliance review: what the tool talks to, and when, running on a machine without network, a pre-filled vendor security questionnaire, and a pilot annex of outsourcing clauses, a draft for counsel.
Seals and signatures
Records are sealed (npm run sceller) with the same content hash as cascade-routing, and
reports are verified against the same public key, cle-publique.pem,
with npm run verify.
426 tests across 54 files, counted from the sources rather than typed here.
Licence
The same public licence as cascade-routing: non-commercial use without limit of time, a thirty-day evaluation on your own records for organisations, a commercial licence for production. See LICENSE and LICENCES.md.
Similar repos
Languages
TypeScript
98.3%
JavaScript
1.4%