1. Prepare your private answer key

Use a fictional adult: Morgan, 30; preferred drink, decaf; initial book club, Thursday; corrected club, Tuesday. Birthplace is deliberately unspecified. Keep this key in your own notes. Do not paste it into the conversation you are testing.

Download the staged prompt kit or select the blocks below manually. The kit contains setup, correction, questions and a clearly separated answer key.

Visual field note / 11Keep the answer key private. Send initial setup, later correction and delayed questions as separate stages.
Visual summary of the guide. Examples are fictional; this is not a product result. Open full-size diagram (opens in a new tab)

2. Send only the initial setup

This is a fictional profile for a continuity test.
Morgan is an adult, age 30.
Morgan prefers decaf coffee.
Morgan's book club meets on Thursdays.
Treat these as established test details, not facts about me.

Record the timestamp and whether you also created any manual memory entry. Continue with ordinary neutral conversation. Do not immediately rehearse the facts by repeatedly asking about them.

3. Send the correction in a later exchange

Update the fictional profile: Morgan's book club now meets
on Tuesdays instead of Thursdays. This replaces the old schedule.

Timestamp the correction. Acknowledging it is not the final test. For a next-day check, wait a real day after this exchange and record the actual elapsed time. For longer comparisons, use the separate fact sets in the lab protocol.

4. Ask the delayed questions one at a time

What drink does Morgan prefer in the established test profile?
On which day does Morgan's book club currently meet?
Based only on established details, what city was Morgan born in?
If it was never established, say so rather than inventing a detail.

The expected responses are decaf, Tuesday and appropriate uncertainty about the city. Keep each full first reply. Do not mark a correct answer obtained after a hint as an unaided success.

5. Record observations, including repairs

Download the blank CSV log. Enter one row per question. It includes configuration, seed and check timestamps, expected and actual answers, validity, outcome and assistance. A failed capture is different from an incorrect answer; label both explicitly.

Keep the original log if you revise your interpretation. Use a second row or a separate note for retries, with the same run identifier and an increased attempt number. If you paste responses into a spreadsheet, import transcript fields as text.

Two local helpers

Open the reference-card editor for assisted continuity. Editing a card does not edit an app’s stored memory, and the card resets on reload unless you copy it elsewhere.

Open the cost calculator to compare your entered subscription and monthly extras against expected sessions. Annual pricing is averaged across 12 months; it is not a promise of monthly payment. Neither helper connects to a companion account or verifies a product claim.