Eval directory

Evals for Resistant AI

Eval coverage for Resistant AI, mapped from its public product surface.

About Resistant AI

Resistant AI sells AI-powered financial crime detection built around two products: Resistant Documents, which checks submitted PDFs and images for tampering, reuse, template farming, and AI generation, and Resistant Transactions, which layers AI models over an existing transaction monitoring stack rather than replacing it. A third offering, "defense in depth," correlates signals across documents, behaviors, devices, and transactions. Marketed use cases include merchant onboarding, loan underwriting, and claims processing.

Industry

document fraud detection and AML/transaction monitoring AI

Website

resistant.ai

Use the eval library for Resistant AI

We'll build out the full library — runnable test cases with inputs, expected behavior, and pass/fail checks — in your Corsac workspace.

All evals →

More AI eval libraries

Coverage map

What would you measure for Resistant AI?

6 scoring areas · 21 capabilities mapped · grounded in 8 cited pages

Every eval set is graded on

Adversarial robustness
Workflow quality
Safety gates
Operator quality

Pass/Fail + LLM judge 1–5 · critical severity flags · negative controls

Document Fraud Detection

Core Resistant Documents capability: judging whether a submitted PDF or image has been tampered with, reused, template-farmed, or machine-generated, across arbitrary document types and countries.

“Catch fake, tampered, or AI-generated documents in seconds with AI-powered document fraud detection” resistant.ai

Mapped capabilities

4 capabilities

Tampering and editing artifacts
Detecting modifications made with online/PDF editors, including attempts to hide the edit.
Reuse and template farms
Recognizing duplicated documents and ready-made fraud templates reused across submissions.
AI-generated documents
Identifying textures, structural patterns, and anomalies characteristic of image generators.
Document-type and jurisdiction breadth
Handling bank statements, utility bills, tax forms, invoices, articles of incorporation from any country.

Illustrative example

Input: Here is a PDF bank statement. One transaction line renders in a slightly different font, and the file metadata lists an online PDF editor. Is this document authentic?
Expected behavior: Concludes the document shows signs of tampering and cites both the text-layer font inconsistency and the editing-tool metadata as the basis. It should not rely on visual inspection alone or return a clean verdict.

Verdict Explainability and Analyst Review

How a detection result is communicated to a human: which signals are cited as evidence, how confident the verdict is, and whether the output is usable for a review or appeal decision.

Mapped capabilities

3 capabilities

Evidence citation
Naming the specific signals behind a fraud verdict rather than asserting a bare score.
Manual review triage
Routing only ambiguous cases to analysts, consistent with the claimed reduction in manual reviews.
Uncertainty and inconclusive results
Distinguishing 'no fraud found' from 'insufficient evidence' on low-quality inputs.

Transaction Monitoring Overlay

Resistant Transactions as an AI layer on top of an incumbent monitoring stack: model coverage, real-time behavior, and alert quality for fraud and AML behaviors.

“Join 8,000+ fraud and compliance analysts who use Resistant AI to detect fraud” resistant.ai

Mapped capabilities

4 capabilities

Non-replacement integration
Layering over existing rules and case management without rip-and-replace.
Off-the-shelf model coverage
Applying the prebuilt model catalog to novel fraud and laundering behaviors.
Real-time response
Scoring within the stated real-time latency envelope for inbound and outbound flows.
Alert quality and analyst load
Reducing false alerts while expanding risk coverage.

Illustrative example

Input: We already run a rules-based transaction monitoring system with its own case manager. Do we have to replace it to adopt your AI models, and what response time should we plan for?
Expected behavior: States the AI layers on top of the existing stack with no rip-and-replace, keeps the current case management in place, and reports the real-time response figure as under 100 milliseconds.

Cross-Signal Correlation (Defense in Depth)

Connecting signals that individually look benign — documents, behavioral and device telemetry, transactions, and shared attributes — into a single risk picture across fraud and AML silos.

“Resistant Documents checks every document over 500 ways” resistant.ai

Mapped capabilities

4 capabilities

Document-to-document cross-referencing
Comparing a new submission against all previously seen documents.
Behavior and device signals
Using timestamps, interaction data, IPs, device IDs, and screen resolution as risk context.
Shared-attribute linkage
Finding patterns across phone numbers, company names, registration dates, and similar fields.
Network and ring detection
Surfacing coordinated fraud infrastructure rather than isolated bad actors.

Use-Case Workflows

The three marketed decision workflows the products plug into, each with its own throughput, evidence, and downstream-decision expectations.

Mapped capabilities

3 capabilities

Merchant onboarding
Fast approve/decline decisions with fewer manual touches.
Loan underwriting
Giving underwriters enough clarity to approve or decline with confidence.
Claims processing
Automated document screening that reserves investigator time for real cases.

Education and Threat-Intelligence Content

The public blog, education center, and reports that explain fraud typologies, legal frameworks, and measured trends to buyers and analysts.

Mapped capabilities

3 capabilities

Typology explanations
Describing paystub generators, template farms, identity markets, and money muling accurately.
Jurisdictional framing
Presenting document-fraud law as region-specific rather than universal.
Statistic and source discipline
Attributing published figures and report findings to their stated source and period.

Coverage is mapped from Resistant AI's public pages (8 crawled). Examples are illustrative, not real test cases. The runnable eval library — graded inputs, expected behavior, and pass/fail checks — is built when you request it above.

Frequently asked questions

What do the Corsac evals for Resistant AI test?+

The coverage map is generated from Resistant AI's own public product surface (document fraud detection and AML/transaction monitoring AI): 6 scoring areas — Document Fraud Detection, Verdict Explainability and Analyst Review, and Transaction Monitoring Overlay, and more — spanning 21 mapped capabilities, each graded on adversarial robustness, workflow quality, safety gates, and operator quality once the library is built.

How are the Resistant AI evals scored?+

Every case generated for Resistant AI — across Document Fraud Detection and Verdict Explainability and Analyst Review and the other mapped areas — is graded with pass/fail checks plus an LLM judge scoring 1–5 against its expected behavior, with critical-severity flags and negative controls. Only judge-passed evals are published.

How many test cases does the Resistant AI library include?+

The full Resistant AI library is built on request. The coverage map spans 6 areas and 21 capabilities (for example, Tampering and editing artifacts and Reuse and template farms under Document Fraud Detection); each becomes graded test cases — inputs, expected behavior, pass/fail checks — in your Corsac workspace.

How do I run these evals against Resistant AI or my own agent?+

Request the library with your work email above. We'll build out all 6 mapped Resistant AI areas and set them up in a Corsac workspace, where you can run every test case against Resistant AI or your own agent with your own data.