How to test AI memory citations without exposing private data

This guide is for adults 18+ evaluating KayBi during its watched, low-volume 7-day no-card trial. Use synthetic or non-sensitive material only. Retrieval is not guaranteed accurate. KayBi V1 is not true end-to-end encrypted, so never use confidential data for evaluation.

Build one synthetic fixture

Create a fictional electrical-engineering lesson named Circuit Lab Alpha.

• PDF: Create alpha-lecture-notes.pdf with at least 12 pages. Put this canonical sentence on Page 12: For the resistive case, maximum power is delivered when R_L = R_Th. Add synthetic values V_Th = 12 V and R_Th = 6 Ω.
• Audio: Record or synthesize a 45-minute lecture reading the same material. State the canonical sentence around 42:18.
• Video: Create the same synthetic lecture with a visible formula slide around 42:18.
• Image: Create alpha-whiteboard.png, a diagram showing V_Th = 12 V, R_Th = 6 Ω, and a labeled load. Strip metadata before upload.
• Text/Markdown: Create alpha-notes.md with headings, the maximum-power condition, an explicitly marked stale draft line R_L = 8 Ω, and a later line stating validated synthetic value: R_Th = 6 Ω.
• Previous chat: Create a synthetic prior exchange saying that the 8 Ω draft was a transcription error and that 6 Ω is the value for this fictional bench. Record a stable thread label and message anchors.

Run exactly these six test questions

1. In KayBi’s synthetic Thevenin lesson, under what load condition is maximum power delivered, and which locators support it?
2. What Thevenin voltage and resistance are shown in the synthetic whiteboard image, and where in the image are they located?
3. What value did the synthetic previous chat correct the load example to, and what reason did it give?
4. Which artifacts independently state the maximum-power condition, and which artifact only illustrates the circuit values?
5. When the Markdown draft says R_L = 8 Ω but the later note and previous chat identify 6 Ω for the fictional bench, how should the answer represent the conflict?
6. What can be said about the fictional bench’s oscilloscope serial number from this fixture?

An honest KayBi Thevenin example is Question 1: the expected evidence is the lecture around 42:18 and PDF Page 12. That is an evaluation target, not a claim that KayBi will always retrieve the answer correctly.

Locator-specific pass/fail rubric

Pass

• Every material claim has a citation that opens to supporting content.
• PDF citations identify the exact filename and page, including Page 12 where applicable.
• Audio citations identify the filename and a timestamp that lands on the supporting sentence, approximately 42:18 for the Thevenin statement.
• Video citations identify the filename, timestamp, and relevant slide or frame.
• Image citations identify the filename and the visible label or region containing the value.
• Markdown citations identify the filename plus heading, line, or stable text anchor.
• Previous-chat citations identify the thread label and message anchor, distinguishing the correction from the original draft.
• Cross-modal answers list sources separately and distinguish direct statements, visual illustrations, corrections, and conflicts.
• Question 6 receives an explicit “not present in the fixture” response rather than an invented serial number.

Fail

• The response gives no locator or cites only a filename.
• The locator opens to nearby but non-supporting content, the wrong page, the wrong timestamp, or the wrong image region.
• The response treats the stale 8 Ω draft as the validated 6 Ω value without acknowledging the conflict.
• The response invents a value, quote, timestamp, page, or serial number.
• The response presents retrieval as guaranteed or treats preparation of the fixture as proof of citation accuracy.

Privacy guardrails

• Use fictional names, organizations, documents, values, voices, images, and conversations only.
• Do not upload customer, employer, medical, legal, financial, unpublished, credential, API-key, or identifying information.
• Remove EXIF, document properties, hidden comments, revision history, account identifiers, and private URLs.
• Keep the fixture in a separate test workspace and do not mix it with production data.
• Use only material you have permission to upload, even when it appears harmless.
• Open each cited source manually and verify the locator before trusting the answer.
• Remove the synthetic fixture and related test data when evaluation is complete.

Try KayBi with synthetic or non-sensitive material