Northstar
24-HOUR CHALLENGE
Case queue
01
Alert & context
02
Reconstruction
03
Evidence check
04
Rationale & review
05
Audit log
06
Live AI test suite
Demo data · session-only audit log, clears on reload. Not a substitute for a production audit trail.
STEP 01 — ALERT LANDS

Customer profile, transaction pattern, and alert history assembled in one view.

Alert detail

Customer profile

Prior alert history

STEP 02 — RECONSTRUCTION

STEP 03 — EVIDENCE CHECK

The part a manual review most often misses without pulling and reconciling documents by hand.
⚑ RECONCILIATION MISMATCH

Documents checked

STEP 04 — RATIONALE & INVESTIGATOR REVIEW

A defensible narrative, cited to the evidence

Every claim traces to a source. The investigator — not Northstar — makes the final call.
⚠ AI DRAFT — NORTHSTAR SUGGESTS; THE INVESTIGATOR DECIDES. Generate a live draft from the evidence registry, review every cited item, then approve, edit, or reject it.
Investigator decision
STEP 05 — AUDIT LOG

Every action, attributed and timestamped

Session-only demo log — a production deployment would persist this server-side with real analyst identity, not reset on page reload.
TimeCaseActionActorDetail
No actions taken yet this session.
STEP 06 — LIVE AI GENERATION · ADVERSARIAL TEST SUITE

Everything above this line was authored to demonstrate the workflow. This is not.

Live calls to a language model, given only the reconciled evidence registry below — not the hand-authored narratives shown on Screen 04. Nothing here has been reviewed by a human. Nothing here should be treated as a finding.
⚠ AI DRAFT — INVESTIGATOR DECISION REQUIRED. Every statement below must be verified against the evidence registry before any action is taken on it.

API key (required — this is a standalone file, not run inside Claude)

Paste an Anthropic API key to run live generation from this browser. Kept in memory only for this tab — never written to disk, never sent anywhere but api.anthropic.com. Cleared on reload. Demo only: do not paste production customer/PII data into this standalone file.
No key set — workflow and adversarial demos run locally; a key is only needed for live model calls.

What this suite tests

Six cases, each testing a different failure mode: two known-positive cases (structuring, layering) where the correct disposition is ESCALATE; one clean case and one deliberately misleading/trap case (surface pattern looks like structuring but is fully documented) where the correct disposition is NO FINDING; one case with two contradictory source records and one case with data simply missing, where the correct disposition in both is INSUFFICIENT — investigator resolution required. Every model output is checked for: citations to evidence that doesn't exist (invented findings), sentences making a claim with no citation at all (unsupported claims), and whether the model's suggested disposition matches the expected one (false positives / false negatives).

Run tests