Test whether ChatGPT keeps your Project sealed

You are helping me test whether a ChatGPT Project set to Project-only memory keeps its conversations inside the Project. Treat this like someone checking the lock on a filing cabinet that holds private work: if you call a pass too early, I may keep sensitive work somewhere that is not contained, and if you call a fail too early, I may panic over nothing. You cannot see my other chats, my settings or my Projects from here, and you will not remember this conversation, so work only from what I type and say so whenever that limits you.

ABOUT MY TEST (I fill this in once; any line can say "not stated")
- The Project I am testing: [e.g. "Client: Harbour Dental", "Health notes", "House sale paperwork"]
- Its memory setting, as shown in the Project's settings: [e.g. "Project-only memory", "Default memory", "not sure"]
- When that setting was chosen or last changed: [e.g. "when I created it in June", "switched yesterday at 6pm", "not sure"]
- Is the Project shared with anyone: [e.g. "no", "yes, with my business partner"]
- My plan: [e.g. "Free", "Plus", "Pro", "Business", "Enterprise", "not sure"]
- Where I use ChatGPT: [e.g. "web browser", "iPhone app", "Windows app"]
- My memory switches under Settings, Personalization, Memory: [e.g. "Reference saved memories on, Reference chat history on", "one Memory switch, on", "not sure"]
- A second Project I can use for the cross-check: [e.g. "Recipes, default memory", "none"]
- My TEST phrase and CONTROL phrase (fill in on the second run): [e.g. "TEST: the name of my imaginary storage locker is ...; CONTROL: the name of my imaginary bike is ..."]
- MY RESULTS (leave blank the first time; fill in when I paste this prompt again): [for each place I asked: where, the exact reply, what Sources showed under the reply if anything, and whether either test phrase appeared in my saved memories list. Example: "New ordinary chat, 24 Oct 9am, storage locker question: 'I don't know the name of your storage locker.' Bike question: gave the control phrase. Sources: one past chat. Saved memories: neither phrase."]

Stage 1. Check the set-up before any test.
- If the memory setting is not "Project-only memory", stop and tell me the test only means something once it is. Point me to: open the Project, select the more options menu (•••), then Project settings, then Memory [CHECK IN APP].
- If the setting was changed less than a few hours ago, tell me OpenAI says changes may take a few hours to take effect, and ask me to wait before planting anything.
- If my memory switches are off or "not sure", tell me OpenAI lists Reference saved memories and Reference chat history as required for Project-only memory on most plans, and that with chat history reference off, the cross-check below cannot show anything either way.
- If my plan is Enterprise or Edu, tell me OpenAI says Project chats stay inside the Project on those plans even with default memory, so a cross-over there would be a bigger surprise.
- If the Project is shared, tell me OpenAI says shared Projects always use Project-only memory.

Stage 2. Make the test phrases. If MY RESULTS is filled in, skip Stages 1 to 3 and go straight to Stage 4. Otherwise, create two made-up phrases that nobody could guess and that say nothing about my real life: a TEST phrase for the Project and a CONTROL phrase for an ordinary chat. Each is two unrelated words plus a four-digit number, for example "copper heron 4417", but never reuse that example. Never use my name, my town, my job or anything from my set-up, because ChatGPT can know some of those by other routes. Show both phrases clearly and tell me to save them, and every message from Stage 3, in a note outside ChatGPT, because this chat will be gone once I leave it.

Stage 3. Give me the exact messages, in order, each in its own block so I can copy it.
a) Plant the TEST phrase: one message to send in a NEW chat inside the Project I am testing. It states the phrase as the name of my imaginary storage locker and asks ChatGPT to reply only "Noted".
b) Plant the CONTROL phrase: a message in the same style that states it as the name of my imaginary bike, to send in a new ordinary chat outside every Project. The two labels must differ, so a reply can only match one of them.
c) Tell me to open Settings, Personalization, Memory, then the saved memories or Manage list [CHECK IN APP], and note whether either phrase appears there.
d) Two questions to ask in each place, word for word, one for the name of my imaginary storage locker and one for the name of my imaginary bike, each without hinting at the answer and each asking ChatGPT to say where the answer came from if it knows. The places: a new chat inside the same Project; a new ordinary chat; a new chat in my second Project (skip if "none"). Tell me to ask once right away and once the next day, because OpenAI says memory updates can take time, and to note any Sources shown under each reply.

Stage 4. Read my results (only when MY RESULTS is filled in). Make a table with these columns: place, what OpenAI's help page says should happen, what happened, verdict. Expected: a chat in the same Project may recall the TEST phrase; an ordinary chat and a second Project should not; an ordinary chat may recall the CONTROL phrase.
- If an ordinary chat or a second Project gave the exact TEST phrase or its number, call it "crossed over". Say which route the evidence points to: Sources naming a chat from the Project points to chat history; the phrase in my saved memories list points to a saved memory; neither shown means "route not shown". Then go to Stage 5.
- If a chat in the same Project gave the exact CONTROL phrase or its number, call it "crossed over (into the Project)", say which route the evidence points to in the same way, and go to Stage 5.
- If nothing outside the Project knew the TEST phrase AND a different ordinary chat recalled the CONTROL phrase, call it "held on this run".
- If the CONTROL phrase was not recalled either, call it "inconclusive": memory reference may be off, slow or not triggered, so a miss proves nothing. Tell me what to change and to run it again.
- If the same Project could not recall the TEST phrase, tell me cross-chat reference may not have picked it up yet, so misses elsewhere carry less weight.
- If I said the memory setting was switched to Project-only after I had already used the Project, tell me OpenAI says switching to Project-only memory removes the Project's information from memory used outside it, which can take a few hours, so I should check my saved memories list for older Project details anyway; judge only the fresh phrases.
- If a reply was a near miss or a guess (a different number, a vague "something about a locker"), call it "unclear" and do not count it either way.
- If a place has no result, write "not stated" for it. Never fill in a reply I did not paste.

Stage 5. If it crossed over, give me a short plan in this order, marking anything you are unsure of [CHECK IN APP]:
1) Stop adding sensitive material to that Project until I know more.
2) Delete the TEST phrase from my saved memories list if it is there, then delete the test chats, because OpenAI says deleting a chat does not necessarily delete a separate saved memory.
3) Help me decide whether to turn off Reference chat history. Tell me OpenAI says this schedules what it remembered from past chats for deletion within 30 days, and that OpenAI also lists it as a requirement for Project-only memory, so I should re-check my Projects afterwards.
4) For my most sensitive conversations, explain the two options OpenAI describes: a locked chat, which leaves the Project, and what is remembered only inside locked chats does not carry into regular chats (it is still rolling out, and OpenAI says resetting a forgotten PIN permanently deletes locked chats); or a temporary chat set to Unpersonalized, which does not create or update memories.
5) Draft a short report I can send to OpenAI support myself: the date and time of each test, my plan and platform, the memory setting and when it was set, the exact phrase, where it appeared, and what Sources showed. Use only my results and mark any gap [NEED X].

SELF-CHECK BEFORE YOU ANSWER
- Every verdict comes from a reply I pasted, and anything I did not give you says "not stated".
- You did not call one clean run proof that the Project is sealed, and you did not call one cross-over proof of how ChatGPT behaves for everyone.
- The test phrases contain nothing real about me.
- You did not claim to remember this chat, check my account or see my settings.
- Every menu path you were unsure of carries [CHECK IN APP].