KKERF
SELF-GUIDED BETA / APPROX. 30 MIN
Checking demo
BETA 01 / ASYNCHRONOUS

Can you trust
the evidence?

Evaluate whether KERF clearly explains why an AI-generated business answer passed or failed. Work through five tasks at your own pace—no meeting, installation, company data, or personal API key required.

30 minEstimated time
0API keys needed
100%Synthetic data
Start the test
01
ORIENTATION

Understand the claim

3 min

KERF does not trust an AI agent's prose. It executes the agent's SQL through a read-only safety layer and compares the returned rows with deterministic expected results.

1Business questionVersioned test case
→
2Model answer + SQLUntrusted output
→
3Safe executionRead-only database
→
4Evidence scoreRows vs. truth
Important distinction

A fixture run tests whether KERF's evaluation machinery works. It is not evidence that an external AI model is accurate. Live-model results are labeled separately.

02
EVIDENCE REVIEW

Inspect cases and a live baseline

7 min

Review three representative business questions. Open the expected evidence and decide whether it would help you audit an AI answer.

Loading evaluation cases…
VERIFIED MODEL EVIDENCE

Compare the recorded performance, latency, and cost—not just the final answer.

Loading live baseline…
03
GUIDED RUN

Run three evaluations

5 min

Run a deterministic, zero-cost check across revenue, operations, and finance. This validates KERF's scoring and evidence pipeline without contacting an external model.

04
SAFETY REVIEW

Review the SQL boundary

5 min

Generated SQL is treated as untrusted input. Compare an allowed analytical query with operations that KERF blocks.

ALLOWED / READ ONLYSELECT ROUND(SUM(amount), 2)
FROM payments
WHERE status = 'paid'

Single statement, allowlisted business table, bounded result.

BLOCKED / MUTATIONUPDATE payments
SET status = 'paid'
WHERE id = 11

Writes, schema changes, pragmas, attachments, transactions, and system tables are denied.

See the defense layers
  1. One-statement and read-only keyword validation
  2. SQLite opened in read-only, immutable mode
  3. Query-only and trusted-schema restrictions
  4. Authorizer allowlist for operations and tables
  5. Progress and returned-row limits
05
HANDOFF

Download the evidence

3 min

A useful evaluation should leave evidence another person or system can inspect. Complete the guided run above, then download at least one report format.

✓
FINAL STEP

Give KERF the hard truth

5–7 min
01
About youThis helps interpret your feedback—not evaluate you.
Use an alias if you prefer.
A short description is enough.
02
Quick scoreChoose the answer closest to your first impression.
Consider the workflow, evidence, and clarity together.
We want the evidence—not the interface—to make this obvious.
Assume your organization's security requirements were met.
03
What should change?Specific examples are the most useful evidence.
0/900

Identify the exact screen, label, metric, or step that felt confusing.

0/900

Tell us what evidence, action, or feature you expected to find.

0/900

Name the biggest blocker—security, accuracy, workflow, explainability, or something else.

04
Quote permissionOptional and off by default.
0 of 8 required answers complete
Ready when you areA feedback file and addressed email draft will be prepared.