The Skoor, published

A number from observed events. Never a default.

A Skoor is computed from what the platform has seen on your rails. It carries its inputs, its signals with their weights, its version, the number of observations behind it, and a confidence. Where the history is too thin, it is null, and null is written as the word.

The bands

Four bands, one scale, every Skoor.

BandSkoorWhat the policy does

Hold0–30Held first, decided second. Fail closed.

Review31–70A person decides, with the narrative and the evidence already drafted.

Clear71–100The platform may close it under policy. Logged, sampled, never silent.

UnscorednullToo little history to score. Treated as Review, never as Clear.

The null rule.

A Transaction Skoor is null when the program has fewer than 30 transfers of history and the transfer carries no hard signal. An Entity Skoor is null under 5 transfers; a Counterparty Risk Skoor under 3; a Program Skoor under three measured KRIs. A hard signal always produces a number and always lands in the Hold band, whatever the total.

Versions

A version changes only with a note and a rerun.

Every Skoor and the policy carry a version. A version changes only with a written note saying what moved and why, and a rerun against the labeled set showing what it did to precision and recall. The old version stays readable on every record it produced.

VersionWhat it names

trs-v2Transaction Skoor

ers-v2Entity Skoor

crs-v2Counterparty Skoor

ars-v2Agent Skoor

prs-v2Program Skoor

policy-v1The automated-or-reviewed line

Transaction Skoor · trs-v2

Every signal, its weight, and whether it is hard.

Points add and cap at 100. A hard signal (marked) forces the Hold band regardless of the total: held first, decided second.

Entity

SignalPointsHard

Verification denied+60hard

Unverified, pending, or in manual review+25—

Flagged high risk by screening+20—

PEP: yes+30hard

PEP: potential+15—

Requires additional screening+10—

Screening older than 90 days, or never+10—

Entity return rate above 0.5%+15—

Counterparty

SignalPointsHard

First time this program has paid it+10—

First time, and above the program's 95th-percentile amount+15—

Has drawn an unauthorized return before+40—

Has drawn any return before (when no unauthorized one)+15—

Amount and velocity

SignalPointsHard

Above 3× the program's median amount+10—

Above 10× the program's median amount (instead of 3×)+20—

Above the program's declared monthly volume+20—

More than 5 transfers from the entity in 24 hours+10—

Entity's 24-hour sum above 2× its daily average+10—

Structuring pattern: repeated sub-$10,000 debits in 24 hours+35hard

Status, rail, and geography

SignalPointsHard

Transfer placed in manual review by the bank+15—

Transfer on hold+5—

Card decline for suspected fraud+30—

Counterparty in a sanctioned country+50hard

Counterparty outside the program's declared countries+15—

Confidence.

Confidence is 0.5 × min(1, program history ÷ 200) + 0.3 × min(1, entity history ÷ 20) + 0.2 if the entity was screened within 90 days. It says how much the platform has seen, not how sure it is that the transfer is bad. Both numbers are shown together.

What proves the Skoor is right

Six metrics, each with a formula and its n.

Calibration

For each band, the share of transfers later labeled bad (returned, disputed, or escalated) among transfers with any label. Higher bands should go wrong more often, month over month.

Coverage

Transfers with a Skoor ÷ all transfers, and programs with a Program Skoor ÷ all programs. Unscored volume is the honest measure of missing history.

Precision and recall

Of the Hold band against labeled outcomes: precision = held and bad ÷ held; recall = held and bad ÷ all bad. Per policy version.

Auto-closed and overturned

Alerts closed under policy without a person ÷ all alerts; and decisions where a person disagreed with the draft ÷ all human decisions.

Time to decision

Median and 90th percentile hours from alert opened to decision, for the automated and the reviewed queues separately.

Approved without waiting

Count and sum of transfers in the Clear band that were never alerted in the window.

Every number carries its n. Where nothing has been measured, the platform says null. Nothing is invented.

Looking for the rubric for accounting firms? It lives here →

See the Skoor on your own events.

A walkthrough is forty minutes. A sandbox pilot scores your programs on your test environment, with every number carrying its n.

Become a design partner

Four bands, one scale, every Skoor.

BandSkoorWhat the policy does

Hold0–30Held first, decided second. Fail closed.

Review31–70A person decides, with the narrative and the evidence already drafted.

Clear71–100The platform may close it under policy. Logged, sampled, never silent.

UnscorednullToo little history to score. Treated as Review, never as Clear.

The null rule.

A version changes only with a note and a rerun.

VersionWhat it names

trs-v2Transaction Skoor

ers-v2Entity Skoor

crs-v2Counterparty Skoor

ars-v2Agent Skoor

prs-v2Program Skoor

policy-v1The automated-or-reviewed line

Every signal, its weight, and whether it is hard.

Points add and cap at 100. A hard signal (marked) forces the Hold band regardless of the total: held first, decided second.

Entity

SignalPointsHard

Verification denied+60hard

Unverified, pending, or in manual review+25—

Flagged high risk by screening+20—

PEP: yes+30hard

PEP: potential+15—

Requires additional screening+10—

Screening older than 90 days, or never+10—

Entity return rate above 0.5%+15—

Counterparty

SignalPointsHard

First time this program has paid it+10—

First time, and above the program's 95th-percentile amount+15—

Has drawn an unauthorized return before+40—

Has drawn any return before (when no unauthorized one)+15—

Amount and velocity

SignalPointsHard

Above 3× the program's median amount+10—

Above 10× the program's median amount (instead of 3×)+20—

Above the program's declared monthly volume+20—

More than 5 transfers from the entity in 24 hours+10—

Entity's 24-hour sum above 2× its daily average+10—

Structuring pattern: repeated sub-$10,000 debits in 24 hours+35hard

Status, rail, and geography

SignalPointsHard

Transfer placed in manual review by the bank+15—

Transfer on hold+5—

Card decline for suspected fraud+30—

Counterparty in a sanctioned country+50hard

Counterparty outside the program's declared countries+15—

Confidence.

Six metrics, each with a formula and its n.

Calibration

For each band, the share of transfers later labeled bad (returned, disputed, or escalated) among transfers with any label. Higher bands should go wrong more often, month over month.

Coverage

Transfers with a Skoor ÷ all transfers, and programs with a Program Skoor ÷ all programs. Unscored volume is the honest measure of missing history.

Precision and recall

Of the Hold band against labeled outcomes: precision = held and bad ÷ held; recall = held and bad ÷ all bad. Per policy version.

Auto-closed and overturned

Alerts closed under policy without a person ÷ all alerts; and decisions where a person disagreed with the draft ÷ all human decisions.

Time to decision

Median and 90th percentile hours from alert opened to decision, for the automated and the reviewed queues separately.

Approved without waiting

Count and sum of transfers in the Clear band that were never alerted in the window.

Every number carries its n. Where nothing has been measured, the platform says null. Nothing is invented.

Looking for the rubric for accounting firms? It lives here →