Which scenario suffices, at which threshold?

Seven scenarios, from a bare amount to the deviation from a peer profile, swept across fifty‑one thresholds.
Recall and false alerts carry their intervals, and every figure can be verified by you.

git clone https://github.com/ArslaneSempai-ui/cascade-monitoring npm ci --ignore-scripts npm run measure:yours -- --alerts=your-alerts.csv --transactions=your-transactions.csv Your dispositioned alerts, measured on your machine. Nothing of yours goes up.

Cascade · three instruments, one method Routing · extraction

Where should the next dollar go?Seven tiers, from a regular expression to a human, measured on sealed records.Open Routing Screening · name matching

Which matcher suffices, at which threshold?Seven matchers, from strict equality to a multilingual embedding, measured on your own alert history.Open Screening

Monitoring · transactions

Which scenario suffices, at which threshold?Seven scenarios, from a bare amount to a peer profile, measured on your own dispositioned alerts.You are here · Monitoring

The settling basins, state 01: The naked amount catches nearly everything.

one number, one blade: its bar stops under the brass lineseven amber diodes, all the same explained purchase

finding 01one twin in six alarms

Score nothing but the biggest amount in the window and almost every archetype trips it. So does the family purchase the history explains: every one of its false alerts is the same benign twin, the one-off purchase.

95.2%recall of the amount scenario at 0.50, interval [84–99], 42 written suspicious cases

16.7%false alerts on the 42 benign twins, interval [8–31]: all of them one-off purchases

The settling basins, state 02: A busy account is not a signal.

the counting blade: most archetypes pass under it, unlithalf the amber row burns: payday, season, savings

finding 02half the twins alarm

Count transactions and payroll looks like laundering: the twins velocity trips on are the salary account, the market season and the monthly savings, while most true archetypes pass under it quietly.

28.6%recall of the velocity scenario at 0.50, interval [17–44]: most archetypes are not busy

50%false alerts at the same threshold, interval [36–64]: payroll, season and savings, measured

The settling basins, state 03: No single scenario holds the floor.

the brass line at ninety: no bar reaches it alonethe tallest bar is amount's, and it stops short

finding 03the floor stays unmet

The tool's rule asks the lower bound of measured recall to reach 90. On 42 written cases no cell of any scenario gets there alone: the best bound stops at 84, the next at 75. That is the honest headline, and the reason the sweep is worth running on your history.

84.2%the highest lower bound any single cell reaches: amount at 0.50, recall interval [84–99]

75%the next best bound: peer at 0.50, recall interval [75–95]

The settling basins, state 04: Written and generated, kept apart.

the same blade twice: written cases, then variantsonly the variants' bar crosses the line: kept apart

finding 04never blended

The same cell, two provenances. On generated variants the amount scenario scores a perfect recall; on the written cases it does not. Blended, the synthetic half would round the record's own doubt away. It is measured apart, labelled synthetic, and stays apart.

95.2%recall at amount 0.50 on the 42 written suspicious cases, interval [84–99]

100%the same cell on 126 generated variants, interval [97–100]: kept apart precisely because it flatters

The settling basins, state 05: Your history, your scenarios.

the eighth blade is yours: unlit until measuredan ivory card, two files, your own alert history

finding 05yours replace both

Every case behind these findings is ours: written twin by twin, or generated under a seed. Two files from your own systems, alerts and transactions, replay the whole grid on your machine. Nothing leaves it, and the report lands next to your files.

42written suspicious cases behind the measured half, seven per declared nature

126generated suspicious variants behind the synthetic half, natures preserved by construction

Pick any cell, read what your threshold costs.

Recall over false alerts of each scenario at each threshold, on the written cases
scenario	0.50	0.60	0.70	0.80	0.90	1.00
amount	95% 17% fa	86% 17% fa	67% 17% fa	50% 0% fa	40% 0% fa	0% 0% fa
velocity	29% 50% fa	21% 19% fa	0% 5% fa	0% 0% fa	0% 0% fa	0% 0% fa
structuring	17% 0% fa	17% 0% fa	17% 0% fa	17% 0% fa	17% 0% fa	0% 0% fa
round	45% 0% fa	0% 0% fa	0% 0% fa	0% 0% fa	0% 0% fa	0% 0% fa
zscore	83% 67% fa	83% 50% fa	83% 33% fa	83% 33% fa	83% 33% fa	0% 0% fa
passthrough	50% 17% fa	50% 14% fa	48% 2% fa	48% 0% fa	40% 0% fa	5% 0% fa
peer	88% 17% fa	79% 0% fa	76% 0% fa	76% 0% fa	76% 0% fa	0% 0% fa

Measured on the 42 written suspicious cases and 42 benign look‑alikes: recall on top, false alerts below. The marked cell, amount at 0.50, carries the strongest recall lower bound; under its default rule the tool retains NO cell at the recall floor on these cases: the record says so, and the instrument shows it live. The synthetic variants stay apart, on the instrument.

Cascade · MonitoringOpen the live instrument Every scenario at every threshold, recall and false alerts with their intervals, live from the sealed record, and the tool's own selection rule under your recall floor.

Cascade Monitoring, on film.

The five findings, explained

The lapis Cascade robot, palms up, projecting the two highest lower bounds of recall any single scenario reaches: amount on the left, peer on the right, both under the 0.90 floor.

hosted on YouTube · link pending upload

The appendices your reviewers will ask for.

Method & what is measured What the frontier reads, and what it refuses.

Security & data handling What is rebuilt, what is assumed, and what never leaves.

Terms of engagement What the grant allows, for how long, and what a client buys.

Privacy No data is collected. Written down, and verifiable.

Accessibility Usable by keyboard, by screen reader, and with motion turned off.

Pricing, in figures Contact The fine print The repository, public