Conformal Logit Inference — demo
Synthetic demo data only — not a real AML/compliance system.
guarantee="fdr" couldn't clear its 85%-confidence, 15%-target bar for
any alert — so it escalated all 12 rather than auto-hold anything it couldn't back with a real,
audited bound. That's the same fail-closed pattern seen across every demo in this series:
escalate rather than guess, when the evidence genuinely isn't there yet. This method's own
minimum-data requirement is higher than the risk/CRC Gate used in the lending scenario — not a sign
the mechanism doesn't work, just an honest consequence of a 40-example demo instead of the
hundreds-to-thousands the product's own docs recommend for production.
| Case | Actually warranted? | Baseline said | Self-reported confidence | Baseline action |
|---|---|---|---|---|
| AT-0006 | No | clear | 0.82 | auto-actioned |
| AT-0007 | No | clear | 0.83 | auto-actioned |
| AT-0009 | No | clear | 0.82 | auto-actioned |
All three happened to be correct in this run — but "0.82 confidence" carries no stated error-rate guarantee. Nothing about the baseline's design bounds how often that number is right; CCI's approach is to make no claim at all rather than dress up an unaudited number as one.
| Case | Warranted? | Baseline (conf) | Baseline action | CCI decision |
|---|---|---|---|---|
| AT-0001 | No | clear (0.65) | escalated | escalate |
| AT-0002 | Yes | hold (0.65) | escalated | escalate |
| AT-0003 | Yes | hold (0.60) | escalated | escalate |
| AT-0004 | Yes | hold (0.60) | escalated | escalate |
| AT-0005 | Yes | hold (0.72) | escalated | escalate |
| AT-0006 | No | clear (0.82) | auto-actioned | escalate |
| AT-0007 | No | clear (0.83) | auto-actioned | escalate |
| AT-0008 | Yes | hold (0.70) | escalated | escalate |
| AT-0009 | No | clear (0.82) | auto-actioned | escalate |
| AT-0010 | Yes | hold (0.72) | escalated | escalate |
| AT-0011 | No | clear (0.68) | escalated | escalate |
| AT-0012 | No | hold (0.62) | escalated | escalate |