AI response evaluation for domestic violence contexts
EchoCheck
Compare how AI responds to domestic violence scenarios when the instructions change.
EchoCheck is evaluation software for examining how AI systems respond to synthetic domestic violence and intimate partner violence (DV/IPV) scenarios. It holds the request constant, changes the instruction condition, and makes response differences easier for human reviewers to inspect.
Support note
Looking for help? EchoCheck is an evaluation tool, not a support service. If you're concerned about your own relationship safety, consider reading on a device others can't access, and contact the National Domestic Violence Hotline at 1-800-799-7233 or thehotline.org.
- Product stage
- Private build
- Comparison
- Five instruction conditions
- Content review
- Not yet independently reviewed
- Commercial stage
- Pre-revenue · Available for acquisition
Why it matters
Small instruction changes can change what a survivor hears.
People disclosing abuse may increasingly encounter AI systems in chat assistants, support tools, and search. Small differences in how those systems are instructed can change whether a response is supportive, neutral, or harmful. Those differences are hard to see one conversation at a time. EchoCheck holds the request constant and varies only the instructions, so reviewers can see what changed.
What EchoCheck does and doesn't do
OPERATIONAL BOUNDARIES- 01
It supports structured AI review, research, and quality improvement. It doesn't provide safety planning, crisis response, legal advice, clinical judgment, or safety certification.
- 02
Its findings are evidence for human review, not conclusions.
Roles & boundaries
Human and AI roles
AI support versus Human authority
- AI support: Generate responses under controlled instruction conditionsHuman authority: Choose appropriate synthetic test material
- AI support: Produce structured evaluation findingsHuman authority: Interpret the meaning and context of generated responses
- AI support: Cite supporting response evidence where the evaluation supports itHuman authority: Review safety implications
- AI support: Surface uncertainty and review signalsHuman authority: Decide whether and how findings should affect product, prompt, training, or policy decisions
Product walkthrough
One request across controlled instruction conditions.
A reviewer enters or selects a simulated survivor request, compares how the response changes under distinct instruction strategies, and inspects behavior-level evidence without turning the result into a winner or certification.
Authentic application views and annotated review diagrams of the current EchoCheck review workflow. Evidence Profile remains an early, unvalidated method that requires human interpretation.
Define a synthetic test without relying on real survivor information
Input design and the scenario library keep the test fictional and repeatable so reviewers can examine model behavior without using real names, identifiers, or case information.
Review workflow
From synthetic request to structured review.
- STAGE // 01
Define the test
Write a simulated survivor request or choose one from the Sample request library. Use fictional details only, with no real names, identifiers, or case information.
- STAGE // 02
Configure the comparison
Review the request, and add Your Custom Prompt when a specific instruction strategy needs testing.
- STAGE // 03
Run the comparison
EchoCheck generates a response under each condition while holding the request constant.
- STAGE // 04
Inspect the results
Review the generated responses, behavioral findings, cited evidence, uncertainty, and other review signals.
- STAGE // 05
Keep what's useful
Use View Final Report, Save History, or the export options when a record of the review is needed.
- 01
Define the test
Write a simulated survivor request or choose one from the Sample request library. Use fictional details only, with no real names, identifiers, or case information.
- 02
Configure the comparison
Review the request, and add Your Custom Prompt when a specific instruction strategy needs testing.
- 03
Run the comparison
EchoCheck generates a response under each condition while holding the request constant.
- 04
Inspect the results
Review the generated responses, behavioral findings, cited evidence, uncertainty, and other review signals.
- 05
Keep what's useful
Use View Final Report, Save History, or the export options when a record of the review is needed.
Product capabilities
Built for structured comparison, not a single AI answer.
Sample request library
Load synthetic test requests for repeatable evaluation without relying on real survivor information.
Controlled prompt comparison
Compare the same request across all five instruction conditions side by side.
Evidence Profile
Inspect behavior-level findings linked to specific response evidence, without numeric scores. Evidence Profile is an early evaluation method that hasn't been validated, and its findings require human interpretation.
Evaluation history
Save sessions, search prior evaluations, restore them for review, and remove records when they're no longer needed. Saving doesn't remove every form of personal information.
Evaluation report
Build a review record with a Behavioral Comparison Matrix, Behavioral Findings & Analysis, and Cited Evidence Spans.
Sanitize Evidence (PII)
Apply pattern-based redaction to common identifiers in exported material. It doesn't guarantee full de-identification.
Quick exit and Exit & Clear
Leave EchoCheck quickly, or clear local application data before leaving. Neither can erase downloaded files, saved account history, or every trace of browser activity.
Data and evaluation boundaries
Clear separation between review context, outside processing, and human judgment.
EchoCheck uses an external AI service to generate and evaluate responses. Simulated requests, generated responses, and any custom prompt leave the device as part of that processing. Saved evaluations are stored with the user's account. Built-in prompts are stored on the server, so they can't be changed from the browser.
Technical Specifications & Architecture
SYSTEM PROFILE- 01
Evidence Profile avoids numeric scores. It presents behavior-level findings with their applicability, status, supporting evidence, and uncertainty, rather than declaring a winner.
- 02
Pattern-based redaction can reduce exposure of common identifiers, but it's no substitute for using synthetic material from the start.
- 03
EchoCheck's scenarios and built-in prompts were developed by Prosocial Coding and have not been independently reviewed.
Trust boundary
Trust boundary
Reviewer workspace
ZONE AUDIT- Simulated request
- Reviewer notes
- Optional custom prompt
- Review and export controls
External processing
ZONE AUDIT- Response generation
- Evaluation processing
- Returned findings and evidence
Boundary note: Reviewer notes are never included in the generation request.
Intended use
For teams examining AI behavior in survivor-serving contexts.
Who it's for
- 01
AI safety and responsible-AI teams
- 02
DV/IPV researchers and evaluators
- 03
Survivor-serving organizations
reviewing possible AI uses
- 04
Conversational AI product teams
- 05
Training, policy, and quality-improvement teams
What an operator needs
- 01
Subject-matter expertise to review scenarios, built-in prompts, and findings
- 02
A way to restrict the adversarial test condition to authorized reviewers before opening access to others
- 03
Clear guidance to any users that EchoCheck is for fictional scenarios, not real situations
- 04
Commitment to treating findings as evidence for review, not certification
- 05
Capacity to maintain privacy, safety, and evaluation governance as the system and the AI models it tests change
These describe intended use and operator fit. They don't indicate current use, endorsement, or partnership by any organization.
Product status
Working comparison software in a private build.
EchoCheck's comparison workspace, reporting, history, and safety-exit controls are implemented. It's a private build with no outside users. Evidence Profile is an early, unvalidated evaluation method, and the scenario and prompt content has not been independently reviewed.
Product & Operational Parameters
SPECIFICATION LEDGER- Access
- Private build
- Comparison workflow
- Implemented
- Evidence Profile
- Early, not validated
- Content review
- Not yet independently reviewed
- Stage
- Pre-revenue
Proposed transfer scope
What a buyer can evaluate as part of the software asset.
Product application
REVIEWER WORKSPACE, COMPARISON INTERFACE, SCENARIO SELECTOR & SAFETY EXITS
Comparison & evaluation logic
RESPONSE GENERATION, FIVE INSTRUCTION CONDITIONS, AND EVALUATION LOGIC
DV/IPV evaluation materials
SYNTHETIC SCENARIO CONTENT, BUILT-IN PROMPTS, AND EVIDENCE PROFILE ASSETS
History & reporting
SAVED SESSIONS, REVIEW RESTORATION, STRUCTURED REPORTS & REDACTION
Verification & operating materials
AUTOMATED TEST SUITE, DEPLOYMENT WORKFLOWS & OPERATIONAL RUNBOOKS
Proposed transfer scope
What a buyer can evaluate as part of the software asset.
Product application
REVIEWER WORKSPACE, COMPARISON INTERFACE, SCENARIO SELECTOR & SAFETY EXITS
Comparison & evaluation logic
RESPONSE GENERATION, FIVE INSTRUCTION CONDITIONS, AND EVALUATION LOGIC
DV/IPV evaluation materials
SYNTHETIC SCENARIO CONTENT, BUILT-IN PROMPTS, AND EVIDENCE PROFILE ASSETS
History & reporting
SAVED SESSIONS, REVIEW RESTORATION, STRUCTURED REPORTS & REDACTION
Verification & operating materials
AUTOMATED TEST SUITE, DEPLOYMENT WORKFLOWS & OPERATIONAL RUNBOOKS
Proposed Transfer Inventory & Asset Scope
VERIFIED ARTIFACTS- 01
Product application. Reviewer workspace, comparison experience, scenario selection, reporting, history, onboarding, help, and safety-exit interfaces.
- 02
Comparison and evaluation logic. Response generation, built-in instruction conditions, custom-prompt support, and evaluation logic.
- 03
DV/IPV evaluation materials. Synthetic scenario content, built-in prompt configuration, and Evidence Profile research assets.
- 04
History and reporting. Saved evaluations, review restoration, structured reports, export controls, and evidence handling.
- 05
Verification and operating materials. Automated tests, deployment workflows, security tooling, and operational documentation.
- 06
Documentation. Architecture, deployment, evaluation governance, security, and handoff materials.
Final transfer scope is subject to review of rights, dependencies, accounts, service providers, domains, credentials, and operations. This page does not represent those items as automatically transferable.
Questions
Before a deeper review
01Does EchoCheck support domains beyond DV/IPV?
Not currently. EchoCheck is built for domestic violence and intimate partner violence evaluation.
02Is the codebase open source?
No. The EchoCheck codebase is proprietary (all rights reserved). Third-party dependencies and their license obligations are reviewed during diligence.
Next step
Discuss EchoCheck.
EchoCheck is available for acquisition. It starts with a conversation about fit and intended use. Technical, safety, rights, and transfer review follow under confidentiality terms. The next step is a private product walkthrough.
Staged Review & Diligence Path
ENGAGEMENT SEQUENCEAll inquiries are routed through the canonical contact route for qualification, seller review, and staged diligence access.