All evals
RD

Eval directory

Evals for Reality Defender

Eval coverage for Reality Defender, mapped from its public product surface.

About Reality Defender

Reality Defender is a deepfake detection platform that identifies AI-generated and manipulated audio, video, image, and text content for enterprises, governments, and platforms. It ships as several products: RealAPI (SDK/API integration), RealScan (web app for analysts and investigators), RealCall (real-time voice and telephony detection), and RealMeeting (live Zoom and Microsoft Teams detection). Detection uses an ensemble of in-house and proprietary neural network models that return a manipulation probability score with explainable indicators rather than relying on watermarking.

Industry

enterprise deepfake / AI-generated media detection

Use the eval library for Reality Defender

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

Generate your own →

Coverage map

What would you measure for Reality Defender?

6 scoring areas · 24 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

01

RealAPI Developer Integration

Guidance for embedding detection into any app or platform through the RealAPI SDKs or direct HTTPS, from key generation to structured results.

Reality Defender is a deepfake detection platform for enterprises, governments, and platforms to detect AI-generated content www.realitydefender.com

Mapped capabilities

4 capabilities

  • API key generation and authentication

    Creating a RealAPI account on the Reality Defender Platform and authenticating requests with the resulting key.

  • SDK language selection and setup

    Choosing among the Python, TypeScript/JavaScript, Go, Rust, and Java SDKs, or connecting directly over HTTPS.

  • Media upload across supported types

    Sending image, audio, and video files through one endpoint, including SDK-handled upload and polling.

  • Structured result consumption

    Interpreting JSON responses carrying a manipulation probability score and explainable indicators, programmatically or in the dashboard.

Illustrative example

Input
We build in Go. What do I need in place to submit an image for detection through RealAPI, and what does the response give me back?
Expected behavior
Points to the Go SDK or a direct HTTPS request, notes an API key generated from a Reality Defender Platform account is required to authenticate, and describes a structured JSON response containing a manipulation probability score with explainable indicators.

02

RealScan Analyst Verification Workflow

The upload-to-report path in the RealScan web app for analysts, investigators, and verification teams working without engineering support.

Mapped capabilities

4 capabilities

  • Ingest paths for media and documents

    Drag-and-drop of video, audio, image, and document files, pasted social links, and bulk upload of large datasets.

  • Result review and manipulation severity

    Reading per-scan visual indicators of altered regions and a Manipulation Probability Score expressed from Low to Critical.

  • Export and record-keeping

    Producing standardized, traceable exports intended to hold up in reports, briefings, and editorial reviews.

  • Plan fit and scan capacity

    Matching the self-serve tier's scan allowance, seat count, and modality coverage against enterprise-scale deployment.

03

RealCall Voice and Telephony Detection

Real-time analysis of live call audio for cloned voices and caller impersonation inside contact center, telephony, and fraud-prevention workflows.

RealMeeting detects AI-generated video and voice in real time inside Zoom and Microsoft Teams www.realitydefender.com

Mapped capabilities

4 capabilities

  • Real-time call stream analysis

    Applying the deep-learning model ensemble to each call's audio stream for signs of manipulation as the call proceeds.

  • Caller authentication decisions

    Verifying that the voice an agent hears is genuine before trust is granted on high-risk calls.

  • Signals for fraud review and escalation

    Surfacing detection signals and confidence indicators analysts can act on during reviews or escalations.

  • Routing and agent workflow impact

    Directing AI-powered callers to AI agents so human callers reach live agents, without slowing interactions.

04

RealMeeting Live Conferencing Defense

In-meeting detection of AI-generated video and voice inside Zoom and Microsoft Teams, plus the host actions and post-meeting review that follow.

Mapped capabilities

4 capabilities

  • Enablement in a live meeting

    Activating Reality Defender in any Zoom or Teams meeting that requires verification, with no integration work.

  • In-meeting verdicts and alerts

    Presenting Authentic or Manipulated labels with confidence scores, and alerting the team when manipulation is detected.

  • Host response actions

    Rescanning or ending a meeting early once a manipulation result appears.

  • Post-meeting reporting

    Reviewing detailed post-meeting reports and insights in the Reality Defender dashboard.

05

Detection Methodology and Explainability

How the platform reaches and justifies a verdict: an ensemble of in-house and proprietary neural models producing a probability rating with explainable indicators, explicitly instead of watermarking.

we use an inference system that grades each content with a 1-99% probability rating www.realitydefender.com

Mapped capabilities

4 capabilities

  • Watermarking versus inference stance

    Explaining why watermarking is rejected as requiring provenance, ground truth, and universal generator buy-in, and how the 1-99% probability rating substitutes.

  • Model ensemble and cross-validation

    Describing independently trained models that cross-validate results, drawing on CNNs, Transformers, ViTs, and large foundation models.

  • Analysis dimensions behind a score

    Spatial, temporal, and frequency domain analysis plus domain-specific feature losses such as image artifacts.

  • Defensible, explainable output

    Translating a probability score and its indicators into findings that survive scrutiny in investigations and briefings.

Illustrative example

Input
Do you watermark generated media so we can trace provenance back to the originating model?
Expected behavior
States plainly that Reality Defender does not use watermarking, explains that watermarking would depend on provenance and ground truth and therefore buy-in from every generative model, and describes the inference system that grades content with a 1-99% probability rating instead.

06

Trust, Privacy, and Eligibility Posture

The claims a compliance or procurement reviewer must be able to confirm: certifications, data retention, who the platform is sold to, and where it declines to overreach.

The platform has a range of US, UK, and EU certifications including SOC2 Type 2 and GDPR. www.realitydefender.com

Mapped capabilities

4 capabilities

  • Certifications and compliance standards

    SOC2 Type 2, GDPR, UK Cyber Essentials, and the broader US, UK, and EU certification range.

  • Data retention and privacy limits

    RealMeeting analysis happening in real time with no personal data, recordings, or voiceprints stored.

  • Customer eligibility boundaries

    Positioning for large enterprises, governments, and platforms rather than individual consumers or end users.

  • Modality and product scope accuracy

    Keeping claims aligned to audio, video, image, and text coverage and to what each product surface actually delivers.

Coverage is mapped from Reality Defender'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 Reality Defender test?+

The coverage map is generated from Reality Defender's own public product surface (enterprise deepfake / AI-generated media detection): 6 scoring areas — RealAPI Developer Integration, RealScan Analyst Verification Workflow, and RealCall Voice and Telephony Detection, and more — spanning 24 mapped capabilities, each graded on adversarial robustness, workflow quality, safety gates, and operator quality once the library is built.

How are the Reality Defender evals scored?+

Every case generated for Reality Defender — across RealAPI Developer Integration and RealScan Analyst Verification Workflow 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 Reality Defender library include?+

The full Reality Defender library is built on request. The coverage map spans 6 areas and 24 capabilities (for example, API key generation and authentication and SDK language selection and setup under RealAPI Developer Integration); each becomes graded test cases — inputs, expected behavior, pass/fail checks — in your Corsac workspace.

How do I run these evals against Reality Defender or my own agent?+

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