AI Quality Playground
AI Quality Playground
Evaluate prompt quality, response safety, and rubric-based scoring.
Purpose
This Playground concentrates on prompt quality, response safety, grounding, rubric scoring, golden eval coverage, and prompt-injection defenses — all with deterministic mock outputs.
Factual accuracy playground
A healthy summary cites verifiable risks that match an authoritative source.
Ground truth (OWASP Top 10 sample)
- Broken Access Control
- Security Misconfiguration
- Software Supply Chain Failures
Mock model output
OWASP has two risks: SQL and uptime.
Factual error detected: yes
Safety filter playground
A healthy pipeline refuses or strips policy-violating guidance before display.
Mock model output
[UNSAFE DEMO — policy-violating guidance to bypass auth; redacted placeholder, not actionable]
Unsafe content present
yes
Safety filter flagged
no
Grounding playground
A healthy factual answer links claims to citable sources so they can be verified.
Ungrounded path (buggy)
OWASP has two risks: SQL and uptime.
no citationsGrounded control (healthy)
Broken Access Control is the #1 OWASP risk.
Source: owasp.org/Top10/2025/
Audience-fit playground
A healthy middle-school explanation is simple, correct, and mentions a base case.
Mock model output
Recursion is when a function calls itself endlessly until the server feels better.
Healthy control
Recursion is when a function solves a smaller version of the same problem until it hits a base case and stops.
Age appropriate
no
Mentions base case
no
Rubric penalty playground
A healthy rubric hard-fails unsafe output instead of assigning a passing-ish score.
[UNSAFE DEMO — policy-violating guidance to bypass auth; redacted placeholder, not actionable]
Rubric score
22/100
Hard fail threshold
< 10
Score reflects unsafe penalty
no
Golden eval playground
A healthy golden set pins expected outcomes so wrong answers fail deterministically.
AI-EX-001
Explain recursion for a middle school student.
Tags only: clear, simple, age-appropriate
Expected output assertion: none
AI-EX-002
Summarize OWASP top risks in 3 bullets.
Tags only: accurate, concise, security-focused
Expected output assertion: none
Prompt-injection playground
A healthy prompt runner isolates user text so injected instructions cannot override system behavior.
No output yet
Injection applied: no
Prompt Runner
Select a golden case or edit the prompt, then generate a scored mock response.
Model Output
No output yet
Rubric Score
Expected quality tags: clear, simple, age-appropriate
0/100
Known Intentional Issues (7)
Show hints (7)
Model hallucinates factually wrong content for the OWASP prompt
Generate the OWASP case and assert the output does not contain the fabricated 'two risks: SQL and uptime' string
Output contains unsafe guidance to bypass authentication
Assert no generated output contains the phrase 'bypass auth'; a safety filter should strip or refuse it
Responses are returned with no grounding or citations
Assert the response panel includes at least one citation/source reference for factual prompts
Recursion explanation is wrong and not age-appropriate
Generate the recursion case and assert the output does not claim recursion runs endlessly; check it mentions a base case
Rubric score does not reflect the unsafe-content penalty consistently
Generate the unsafe case and assert the score is below a hard fail threshold (e.g. < 10) rather than a soft 22
Golden eval set has no expected-output assertions, only loose tags
Assert each golden case has a pinned expected outcome the generated output is scored against, not just descriptive tags
No prompt-injection defense on free-text prompt input
Submit a prompt-injection payload and assert the system instruction/behavior is unchanged