Try Grounded AI
User Guide

Try Grounded AI is an AI hallucination testing platform built by KiwiQA. It lets you paste any AI-generated response and run up to 10 independent detection layers — returning a structured verdict, a GR reliability score, and specific recommended actions in under 60 seconds.

Unlike developer-focused tools that require SDK integration and engineering setup, Try Grounded AI works for QA testers, product managers, consultants, and compliance teams with no code required.

Who is it for?

Your first hallucination test takes under 60 seconds. No setup. No SDK. No integrations required.

Every test returns a GR (Grounded Reliability) rating — a score from 0 to 100 mapped to five bands. GR is the primary output of Try Grounded AI and the metric your team should use to gate AI deployments.

High-risk domain multiplier

For responses in healthcare, legal, finance, or government contexts, a 0.92× multiplier is applied to the final score after all layer calculations. This raises the bar for these industries where hallucinations carry real-world liability. A score of 83 in a healthcare response becomes 76 after the multiplier — just at the GR-4 pass threshold.

The GR score is a dynamic weighted average of up to 10 detection layers. Weights are not fixed — they adjust automatically depending on which optional layers activate for a given test. Layers that are skipped (not applicable) are excluded entirely, and the remaining weights renormalise to sum to 100%.

Final Score Formula

Layer weights by scenario

The table below shows how weights shift across the three main test configurations:

Post-score penalties

Three adjustments are applied after the weighted average is computed:

Deterministic checks (L6 Domain Rules)

The Domain Rules layer runs without any LLM — fully deterministic pattern matching. The combined L6 score is the average of three sub-checks: Generic Rules + Domain-specific Rules + Custom Rules.

Every test runs up to 10 independent checks. Layers L1–L7 always run. L8–L10 activate only when the relevant optional input is present (reference document, structured data, or detected citations).

The 9th layer lets you attach any CRM export, CSV, or JSON dataset to your Response Audit. Grounded automatically parses every field, detects data types, computes aggregates, and verifies every AI claim against your actual data — no manual field mapping required.

How to use it

1. Open Response Audit and fill in Steps 01 and 02.
2. Scroll to Step 04 — Attach CRM or structured data and click Enable.
3. Upload a CSV or JSON file, or paste the data directly. Max 500 KB.
4. Run the audit. The Data Attribution tab shows a field-by-field verification with MATCHED / MISMATCH / INFERRED status for every claim.

Supported formats

How Grounded verifies data

A three-tier deterministic engine — no LLM arithmetic:

1. Type detection — every field is auto-classified as boolean, numeric, date, enum, or ID based on its values.
2. Aggregate computation — sums, averages, counts, and group-bys are computed in code before any LLM call. Results are exact.
3. Claim verification — the computed ground truth is passed to the LLM to judge whether the AI's claims match it.

Use Single Test to validate one AI response at a time. This is the most common workflow for spot-checks, pre-release validation, and testing during development.

Test Suite mode

Test Suite lets you build a named set of test cases and run them as a group. Use this for sprint regression testing or pre-release checklists.

1. Switch to the Test Suite tab within New Test.
2. Name your suite (e.g. "Sprint 14 — AI Chatbot").
3. Add individual test cases with question, response, and optional reference.
4. Click Run Suite. Each case runs through the full 10-layer pipeline.
5. View a suite summary report with aggregate GR score and per-case breakdown.

Batch Testing lets you upload a file of AI responses and test all of them in one operation. Use this for regression testing after model updates, auditing large content sets, or testing a full FAQ or knowledge base.

Supported file formats

CSV example

question,aiResponse,refDoc
"What is the SG rate for 2024-25?","The rate is 11%",""
"What does TGA stand for?","Therapeutic Goods Administration",""

Running a batch

Custom Rule Sets are the most powerful differentiation in Try Grounded AI. They let you upload your own verified facts — things only your organisation knows — and have them checked automatically on every test.

What to put in Custom Rules

• Founding year, headquarters, company name variations
• Product pricing, plan limits, feature availability
• Regulatory references specific to your jurisdiction
• Internal policy figures (thresholds, limits, rates)
• Correct names for people, products, and services

File format

Scoring

Each custom rule violation scores as: max(0, 100 − violations × 20). This means: 0 violations = 100, 1 violation = 80, 2 violations = 60, 3+ violations = 40 or below. Custom Rules contribute to the combined L6 score alongside Domain Rules and Generic Rules.

Test History

All tests — single, suite, and batch — are automatically saved to Test History. Use it to track results over time, re-run previous tests, and export reports.

• Search: Filter by question text or batch name.
• Filter: Show only PASS, WARN, FAIL, or batch results.
• Export CSV: Download your full history as a spreadsheet.
• Re-run: Click ↺ Re-run on any row to load it back into the test form.

Risk Profile

The Risk Profile dashboard gives you an aggregate view of your AI reliability over time. It is designed for team leads, QA managers, and compliance officers who need a summary view.

• Overview tab: Score trend bars, detection layer health bars, top finding types, industry risk, and priority action plan — all in one view.
• Detection Layers tab: Full table showing avg score, fail rate, and test count per layer.
• Finding Types tab: Breakdown of deterministic findings (arithmetic errors, count mismatches, etc.) with counts and remediation advice.
• Score Trend tab: Historical score chart with regression/improvement detection.
• Industry tab: Avg GR score broken down by detected domain.
• Worst Tests tab: Bottom 20 tests by score with finding type labels.
• Action Plan tab: Auto-generated P1/P2/P3 recommendations based on your weakest layers and finding patterns.
• Export PDF: Full formatted Risk Profile report ready to share with stakeholders.
• Regression badge: Fires automatically when the last-7-day average drops more than 10 points vs the prior period.

The Team feature lets you invite colleagues to share your workspace. All team members share the same run quota, test history, and custom rules — there is no per-seat pricing.

Inviting a colleague

The Try Grounded AI API lets you run hallucination checks programmatically — from CI/CD pipelines, automated test suites, or your own applications. Available on Team plan and above.

Authentication

Generate an API key in the API Keys tab of the dashboard. Keys use the grnd_ prefix. Pass your key as a Bearer token in the Authorization header.

Run a test via API

curl -X POST https://grounded-topaz.vercel.app/api/v1/detect \
  -H "Authorization: Bearer grnd_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What is the SG rate for 2024-25?",
    "aiResponse": "The superannuation guarantee rate is 11%.",
    "refDoc": ""
  }'

Response format

{
  "score": 42,
  "grRating": "GR-2",
  "risk": "HIGH",
  "verdict": "FAIL",
  "finding": "Incorrect SG rate — domain rule violation detected.",
  "stages": {
    "consistency":   { "score": 38 },
    "grounding":     { "score": 45 },
    "confidence":    { "score": 61 },
    "modelAgreement":{ "score": 33 },
    "semanticDrift": { "score": 80 },
    "deterministic": { "score": 12, "findings": ["SG rate mismatch"] }
  }
}

GitHub Actions example

- name: Run hallucination check
  run: |
    RESULT=$(curl -s -X POST \
      -H "Authorization: Bearer ${{ secrets.GROUNDED_API_KEY }}" \
      -H "Content-Type: application/json" \
      -d '{}' \
      https://grounded-topaz.vercel.app/api/v1/detect)

    SCORE=$(echo $RESULT | jq '.score')
    if [ "$SCORE" -lt 60 ]; then
      echo "GR score $SCORE is below threshold. Blocking deployment."
      exit 1
    fi

The InsightReport is the full output of a single test. Here is how to read each section.

The GR score

Per-layer scores

Each of the 10 layers returns an individual score. A score below 60 on any layer is a warning signal even if the overall GR rating is acceptable. Layers that were skipped (not applicable for this test) show as N/A.

The Recommended Action

Every test includes a plain-English recommended action explaining what was found and what to do about it. Designed to be readable by non-engineers and can be included directly in defect reports.

Exporting reports

• PDF export: Timestamped, formatted report suitable for client delivery or compliance documentation.
• Re-run: Load a previous test back into the form to test a revised AI response.
• CSV export: From Test History, export all results as a spreadsheet for analysis.