Research evidence

Method: 15-minute calls; the participant searched their own phone live while the moderator watched; then the Retrace idea was described. Counts only (n = 6). Quotes are as reported in the moderator's notes (paraphrased unless marked). P4 is an older adult and a low-confidence searcher: the edge of the target segment, kept in and flagged.

	P1	P2	P3	P4	P5	P6	Count
Photos / years	~10k / 5+	~20k / 6-7	~1k / 2 (Android)	n.n.	15-20k / ~10	3-4k / 4-5
Usual way to find photos	scroll	scroll	search	scroll	Google Memories & albums	scroll	scroll first 4 of 6
Placed the photo by an event, trip, activity or period	Shimla trip	COVID period; Ladakh trip	Kerala trip	a 2017 wedding	paragliding 2016; Dubai	n/a (objects, people)	5 of 5 who described a moment
Unsure of month or year	yes	yes	n.n.	n.n.	n.n.	n.n.	2 noted
Raised wrong dates from their own library		WhatsApp photos	phone change	phone changed many times			3 of 6
A search needed rewording or failed first	"suit/blazer/coat" failed					"check" failed, "bank check" worked	2 of 6
Place/people + thing did not narrow	"cafe in Shimla"	"museum in Ladakh"		"wedding" spread across years	skywalk mostly ok	"kids with grandmother"	4 of 6
Object/topic search worked	café (noisy)	prescription	medicinal plants		paragliding	bank cheque (2nd try)	5 of 6
Interested in Retrace	yes	yes	yes, excited	yes	yes	yes	6 of 6
Privacy	worried ("on the internet")	"almost all apps do that"	none voiced	not worried	fine with Google; worried about non-Google AI	yes, "needs some guardrails"	conditions or worry 3 of 6

Findings for the deck

1. People hold the moment, not the date. 5 of 5 who described a moment placed it by a trip, event, activity or period; 2 were unsure of the month or year.
2. Wrong dates are real and people notice them. 3 of 6 raised wrong dates from their own library without being asked (WhatsApp saves, phone changes). P4: "if I don't find it, I probably don't have it" (paraphrase): a failed scroll reads as a deleted photo, invisible to the app.
3. Search works when you know the word; it breaks on wording and on combinations. Object searches worked 5 of 6; but 2 of 6 had to reword, and 4 of 6 saw place/people + thing return a crowd instead of the photo.
4. Most people scroll (4 of 6), one navigates through Google's Memories and albums, one searches first.
5. Everyone wanted it (6 of 6); half attached privacy conditions (control, local storage, guardrails, non-Google AI).

Linked evidence

• Prototype tests (5 first-run testers): research/mvp_tests/README.md.
• Held-out query test (60 queries): research/mvp_tests/engine_eval.md.
• Survey (n = 30) and engine (25,065 posts): HANDOVER.md section 3.

Date of call: 4 Oct 2026. Format: not noted. Phone: Motorola, Android. Main photo app: not noted (gallery search used). Years of photos: at least 5. About 10,000 photos; storage almost full. Age band / city: not noted.

Method note: the moderator asked P1 to use search rather than scroll (P1 normally scrolls). P1 asked what to type; the moderator answered "Whatever you feel like you can write."

Search 1: a café on a Shimla trip

• The photo: a café visited about 4 years ago on a trip to Shimla with five friends.
• Kind: moment (trip, with friends).
• How they placed it in time: by the trip (life event) plus a month they were unsure of: "December 2021 or January 2022". They had two trips around then and were confused between the two months. Sure of the event, unsure of the month.
• Typed: "cafe in Shimla" (1 search).
• Outcome: found. The results showed the wanted photo mixed with cafés from every place they had visited, not only Shimla. Time and tries: 1 search; time not noted.
• Date in the app vs memory: matched (December 2021).

Search 2: trying on a suit in a store

• The photo: P1 trying on a suit in a store; wanted to find that suit again.
• Kind: moment / something to remember (a product to buy).
• How they placed it in time: remembered only the year; not the month.
• Typed: "suit", "blazer", "coat" (3 searches).
• Outcome: not found by search. Not pursued further: P1 searched for it only because the moderator asked for an example, not from a real need, so "gave up" would overstate it. Count as: search failed, not a real retrieval attempt.
• What they would do next (workaround): "scroll through that year and get that photo out" (paraphrase).
• What they thought was wrong: not sure how they could find it.

Reaction to Retrace

• On the 4 Oct call: the moderator described the Retrace idea (no prototype existed yet). P1 found it very interesting and wanted it to work.
• Later, after the prototype was built: P1 tried the Retrace prototype (the demo library). They said searching in their own words was much easier than searching in the gallery (paraphrase). Date of this try and whether the 4 test tasks were used: not noted.
• Privacy worry: if AI goes through all their photos, will the photos "be on the internet all the time"? Will it create a privacy or data-security issue? (paraphrase)

Quotes (paraphrased from notes, not verbatim)

• Never used search before: scrolling the timeline seems easier, they just scroll back to the event they want.
• After Search 1: "This is also an amazing feature", they hope to use it more now.

Observations for analysis

• Placed photo 1 by a life event (the Shimla trip) and was unsure of the month: event-first, date-second.
• Default behaviour is scrolling, not search (matches survey: 17 of 30 did not use search).
• Search ignored the place in "cafe in Shimla" and returned cafés everywhere: results not narrowed by place.
• Object words for an uncommon item ("suit", "blazer", "coat") failed; the fallback plan was scrolling a whole year.
• Privacy concern raised unprompted when AI over all photos was described.

Date of call: 4 Oct 2026. Format: not noted. Phone: Android, current phone about 2 years. Apps: Google Photos and the phone gallery. Photos: about 20,000; phone almost full. Years of photos: 6 to 7. Age band / city: not noted.

Search 1: a prescription from the COVID period

• The photo: a doctor's prescription from around 2020 (COVID period).
• Kind: something to remember (document / medicine).
• Why: P2 chose it to check whether search works.
• How they placed it in time: "the COVID period", around 2020 (a life period), not sure of the month.
• Typed: "prescription in 2020".
• Outcome: found on the first try. Results showed 5 to 6 prescriptions; one was the right one.
• Date in the app vs memory: fine (2020); month was never certain.

Search 2: a museum on a Ladakh trip (to recall the place's name)

• The photo: a museum they visited in Ladakh; they wanted the place's name.
• Kind: moment (trip).
• How they placed it in time: by the trip (Ladakh); the photos were from 2022.
• Typed: a search for Ladakh photos → many Ladakh photos; then narrowed to a museum in Ladakh (exact wording of each not noted).
• Outcome: found by scrolling through the Ladakh photos. Narrowing to the museum did not work: search kept showing all Ladakh photos.
• Date in the app vs memory: matched (2022).

Usual behaviour

• Has used search, but not much. "Usually, scrolling is the first thing that the mind automatically sets to" (paraphrase from notes): scrolls first.

Wrong dates (unprompted observation)

• P2 had noticed that photos saved from WhatsApp and some third-party apps carry wrong dates, while photos carried over from one gallery to another kept the right dates.

Retrace idea

• Moderator described Retrace: connect Calendar, Gmail and Maps to the gallery so it knows what happened when, even if dates went wrong while transferring.
• Reaction: not much concern: "Okay, yeah, almost all the apps do that, so that's not a problem." (as reported by the moderator)
• In the /test closing question (prototype): "Yes, if I can choose which apps it uses", adding "only if my data is stored locally and not online".

Observations for analysis

• Placed photos by a life period ("COVID period") and a trip (Ladakh); unsure of the month in both.
• Scrolls first (second participant out of 2 to say so; survey: 17 of 30 did not search).
• Object search for a document worked first try ("prescription" returned 5 to 6, one right).
• Place search worked, but place + object ("museum in Ladakh") did not narrow; fell back to scrolling a trip.
• Independently confirmed the root cause: WhatsApp and third-party-app photos had wrong dates.

Date of call: 4 Oct 2026. Format: not noted. A quick call. Phones: iPhone (Apple Photos) and Android (Google Photos). The Android phone has about 2 years of photos, about 1,000. Age band / city: not noted. Self-described as tech-savvy.

Usual behaviour (different from P1 and P2)

• Searches first, not scrolls. "it's very easy to search for it. At least it narrows down a bit so that you can scroll only in the narrowed version." Called scrolling "an old-school thing" (paraphrase of notes).
• Asked if search always gives the right result: "Not always, but it narrows down a lot, so it's easier to scroll then."

Search: a medicinal plant on a Kerala trip

• The photo: a particular medicinal plant photographed on a trip to Kerala.
• Kind: moment (trip) / something to remember (a plant).
• How they placed it in time: by the trip (Kerala). Year or month: not noted.
• Typed: "medicinal plants" (exact wording as reported).
• Outcome: found. Results showed 10 to 12 medicinal-plant photos from across the phone; the wanted one was among them. Tries: 1. Time: not noted.
• Date in the app vs memory: not noted.

Wrong dates (unprompted, in their own explanation of when search and scrolling fail)

• When they know an event happened in 2019 but the phone was changed and photos come up with a different date, "that is when the search and scrolling become useless" (paraphrase). Described as the reason they rely on search.

Retrace idea

• Moderator described Retrace (connect Calendar, Gmail and Maps so photos with wrong dates are put back on the right dates).
• Reaction: quite excited; said it would make even scrolling easier.
• Privacy: no concern raised at the time; moderator unsure whether they would worry. Count as "no concern voiced", not "no concern".
• /test: not noted.

Observations for analysis

• Places the photo by a trip (Kerala), like P1 (Shimla) and P2 (Ladakh): 3 of 3 placed a photo by a trip or period.
• Searches first (1 of 3; P1 and P2 scroll first). Uses search to narrow, then scrolls.
• Object/topic search worked ("medicinal plants"), returning many candidates to scroll.
• Independently described the root cause: a phone change gives photos a different date, making both search and scrolling useless (2 of 3 participants raised wrong dates unprompted: P2 WhatsApp, P3 phone change).

Date of call: 5 Oct 2026. Format: not noted. Older adult ("elderly"). Phone, app, years and number of photos: not noted. Phone changed many times. Segment note: older and a low-confidence searcher, so an edge of the target segment (long-time users with years of photos). Report separately where it matters, never hide.

Usual behaviour

• "I open the gallery and I scroll back and try to look for the photo." If they don't find it, they probably think they don't have it (paraphrase from notes). Scrolls only; a miss is read as the photo not existing.
• On search: unsure search could find what they mean. They can't say "a photo with my son" and expect the phone to know who the son is (moderator's account).

Search (prompted: "a place, a trip, a moment, or an event")

• The photo: photos from a wedding in 2017.
• Kind: moment (event).
• How they placed it in time: by the event (a wedding) and a year (2017).
• Typed: "wedding" (looking for the 2017 wedding).
• Outcome: wedding photos came up, but spread across many years, not only 2017 (moderator's account). Whether the specific photo was then picked out: not noted.
• Dates: the phone's timeline is distorted: photos from one year appear under other years because the phone has been changed many times (moderator's observation of their library).

Retrace idea

• Moderator described Retrace correcting the timeline and guiding them to the photo. Reaction: quite interested.
• Calendar/Gmail and privacy: not worried; "didn't care much about it" (moderator's account). /test: not done.

Observations for analysis

• Scrolls first (3 of 4: P1, P2, P4); reads a failed scroll as "I don't have it": the costliest failure, invisible to the app, and the business metric's exact loss.
• Placed the photo by an event and a year.
• Wrong dates from phone changes seen in their own library (3 of 4 now: P2 WhatsApp, P3 phone change, P4 phone changes).
• Naming relationships ("my son") is a barrier to people search for this user.

Date of call: 5 Oct 2026. Format: not noted. Phone: Android, Google Photos with cloud backup to their Google account (which they don't like). Photos: 15,000 to 20,000; about 10 years. Age band / city: not noted.

Usual behaviour

• Likes Google Photos' Memories and events / highlights: "very exciting" (moderator's account).
• Finds photos through albums and Memories that Google creates, then scrolls inside those, so they don't scroll the whole gallery. Has used search for objects: a chat, a prescription, a receipt.

Search 1: paragliding, 2016

• Typed: paragliding (exact words not noted; reported as a paragliding search).
• How they placed it in time: the activity and a year (2016).
• Outcome: found; search showed exactly the paragliding photos.

Search 2: skywalk with a view of the Burj Khalifa (Dubai)

• Typed: a skywalk / Burj Khalifa view search (exact words not noted).
• How they placed it: by place and scene (Dubai trip).
• Outcome: found. Results included Dubai photos with the Burj Khalifa view plus the skywalk photos, narrowed to about 5 to 6; one or two exact shots were not among them.

Retrace idea

• Would value describing a photo naturally: "Maybe 2018, when I saw the doctor, and he wrote me a prescription, and I took its photo" (as reported). "If it works, that would be very good."
• Calendar / Gmail: already uses Google's apps, "so it wouldn't matter much".
• Privacy: main worry is data going online if an AI other than Google is used; asks for privacy settings to protect their data. Also dislikes cloud backup of their photos.
• /test: not noted.

Observations for analysis

• Does not scroll the whole gallery: navigates via Google-made Memories and albums, then scrolls inside them (a third pattern, after scroll-first and search-first).
• Search worked for both an activity ("paragliding") and a place + scene (Burj Khalifa skywalk), narrowed to a handful.
• Placed photos by activity / trip and a year (2016, 2018).
• Their example of a natural query mixes a fuzzy year with a life moment ("maybe 2018, when I saw the doctor"): the exact shape Retrace parses.
• Privacy: comfortable with Google, concerned about third-party AI and cloud storage; wants privacy settings.

Date of call: 5 Oct 2026. Format: not noted. Phone: Android; Google Photos and the phone gallery. Photos: 3,000 to 4,000; 4 to 5 years. Age band / city: not noted.

Usual behaviour

• Not into searching; prefers scrolling. Searched when asked.

Search 1: a cancelled bank cheque

• The photo: a cancelled cheque from the bank.
• Kind: something to remember (document).
• Typed: "check" → nothing relevant ("random things"), then "bank check" → found. Two cheques in the results; the cancelled one was the right one. Tries: 2.
• Outcome: found on the second search; the first search failed.

Search 2: kids with their grandmother

• The photo: two kids with their grandmother.
• Kind: moment (people).
• Typed: "kids with grandmother" (reported also as "photo of kids with grandmother").
• Outcome: many photos: different kids, some with the grandmother, some kids only, some grandmother only. The wanted photo was among them, with many extras. Counted by the moderator as "quite a fail" for precision, though the photo was in the results.

Retrace idea

• "If retrace can solve this issue about when I write things naturally and it will give them to me, I guess that is something worth having in our phones." (as reported)
• Privacy: "Now almost all the things are available online, and our phones are also not that secure anymore. Even having an AI help us wouldn't be a bad thing, but yes, of course, it needs some guardrails to protect our data." (as reported) → yes, with guardrails.
• Dates: not discussed. /test: not noted.

Observations for analysis

• Scrolls first (4 of 6).
• Search needed the right wording: "check" failed, "bank check" worked (rephrasing, like P5 and P1 in the tests).
• Relationship / people query ("kids with grandmother") returned a crowd of partial matches: right photo present, low precision.
• Wants natural-language finding; accepts AI with guardrails.

One file per participant: P<n>.tsv is the "Copy results" output of /test, pasted as is.

Changes to the test during the study (state these in the deck)

When	Change	Sessions run before it
5 Oct, after the first dry run	Preview photo blurred and self-hiding (text on the prescription was being copied into search); T3/T4 stories no longer contain the label word; log records Retrace's own result	dry run (excluded)
5 Oct, after P1	Normal search made fairer: ignores words it doesn't know ("three"), ignores filler words ("when", "was"), shows best partial matches when no photo has every word ("dengue prescription"). Years and months stay hard date filters. Retrace shows a rescued photo first when it is the one searched for.	P1 run 1

Participants

• P1 run 1 (5 Oct): ran under the stricter normal search. Their rephrasing under normal search ("party with three friends" x3, "dengue prescription") was partly caused by that strictness, so do not count their normal-search searches/time as a fair baseline. Retrace results and the consent/trust answers stand.

• P1 run 2 (5 Oct, P1_run2.tsv, entered as participant 2 so modes swapped): a repeat; P1 had seen all four photos. All 4 found in 1 search each, ease 5 throughout, normal 8 s and 14 s vs Retrace 11 s and 16 s. Expected: once you know the photo and its words, both modes work. Useful only as a check that the fairer baseline behaves: normal search found the mandap ("wedding mandap") and the café ("beach side cafe"). One real signal: for "party with friends" plain search returned 6 group selfies but not the dancing photo; Retrace placed the 12 misfiled party photos (0 matches, 12 moved). Consent unchanged ("Yes, if I can choose which apps it uses"); reasons made sense 5/5.

• P2 (5 Oct, P2.tsv, entered as participant 3): first run, fair baseline. All 4 found with 1 search each, ease 5 throughout. Normal: "club party" (25 results, 16 s), "wedding mandap" (10 s). Retrace: "doctor's prescription" (19 s, 1 photo rescued from the wrong date), "beach cafe" (12 s). No difference between modes for P2: they left the year out of the party search and knew the word "mandap", so normal search did not hit the wrong-date trap. Consent: "Yes, if I can choose which apps it uses" + "only if my data is stored locally and not online"; reasons made sense 4/5.

• P3 (6 Oct, P3.tsv, entered as participant 4): first run, fair baseline. All 4 found, 1 search each, ease 5. Both Retrace tasks were real wrong-date cases: "party in 2018 with friends" (plain search: 1 result, a chat screenshot; target missed) and "wedding jaipur" (plain: 16 results without the mandap, which has no location and is filed under Feb 2024). Retrace found both (19 s, 13 s). Consent: "Yes" ("As long as search becomes easier"); reasons 4/5.

• P5 (6 Oct, P5.tsv, entered as participant 5): first run, fair baseline. All 4 found. Normal T1 took 4 searches and 36 s: "party with 3 friends" and "party with friends" (twice) returned only group selfies, "It earlier gave me only selfie photos not the dance photos", until plain "party" (25 results). Normal T2 "Doctor's prescription" 12 s. Retrace: "riya wedding" 12 s (plain search with the same words misses the mandap), "fries and cold coffee cafe" 11 s. Consent: "Yes, if I can choose which apps it uses"; reasons 5/5.

• P6 (6 Oct, P6.tsv, entered as participant 6): first run, fair baseline. All 4 found, 1 search each, ease 5. Retrace: "fresher's party" 12 s (plain search with the same words: 1 result, a chat; target missed), "doctor's prescription" 19 s. Normal: "riya's wedding" 10 s (26 results, mandap included), "sunset beach cafe goa" 12 s. Consent: "Yes"; reasons 5/5.

Rescue check: would the same words have worked in plain search?

For every query typed in a Retrace task, the same words were re-run through the plain search (script check, 6 Oct).

Session	Typed in a Retrace task	Plain search with the same words	Retrace
P1 run 1 T3	wedding Mandap	target included (6 results)	found
P1 run 1 T4	cafe at the beach	target included (7)	found
P2 T2	doctor's prescription	target included (4)	found
P2 T4	beach cafe	target included (7)	found
P3 T1	party in 2018 with friends	missed (1 result, a chat)	found
P3 T3	wedding jaipur	missed (16 results, no mandap)	found
P5 T3	riya wedding	missed (14 results, no mandap)	found
P5 T4	fries and cold coffee cafe	target included (#13 of 19)	found (in the top group)
P6 T1	fresher's party	missed (1 result, a chat)	found
P6 T2	doctor's prescription	target included (#3 of 4)	found
P1 run 2 T1 (repeat)	party with friends	missed (6 group selfies)	found
P1 run 2 T2 (repeat)	medicine prescription	target included (4)	found

Rescue (first runs only): 4 of 4 searches where plain search with the same words would have missed the photo, Retrace found it. In the other 6 first-run Retrace searches plain search would also have worked; Retrace did no worse.

Summary of first runs (P1 run 1, P2, P3, P5, P6; 20 tasks, 5 people)

	Normal search (10 tasks)	Retrace (10 tasks)
Found	10 of 10	10 of 10
Tasks needing more than 1 search	3 of 10 (P1 T1: 4, P1 T2: 2, P5 T1: 4)	0 of 10
Searches typed, total	17	10
Median time	13 s (range 10 to 36)	13 s (range 11 to 51)
Ease, median (1 to 5)	5 (range 2 to 5)	5 (range 3 to 5)
Rescued (plain search with the same words misses it)	n/a	4 of 4

Fair-baseline only (P2, P3, P5, P6; P1 ran under the stricter search): normal needed more than one search in 1 of 8 tasks (P5 T1: 4 searches, 36 s); Retrace in 0 of 8.

Reading. Everyone found every photo, and median time is the same (13 s), so the difference is effort and rescue, not speed: under normal search people reworded when their words did not match the photo's labels ("party with friends" brought selfies, not the dance photo); with Retrace nobody did, and every time the words would have missed the photo in plain search (4 of 4: "party in 2018 with friends", "wedding jaipur", "riya wedding", "fresher's party"), Retrace found it. Different tasks per mode and n = 5: report counts, not percentages or significance.

Consent and trust so far (first runs)

	Calendar / Gmail	Reasons made sense	Words
P1	Yes, if I can choose which apps it uses	4/5	call: worried photos would be "on the internet all the time"
P2	Yes, if I can choose which apps it uses	4/5	"only if my data is stored locally and not online"
P3	Yes	4/5	"As long as search becomes easier"
P5	Yes, if I can choose which apps it uses	5/5	call: fine with Google; worried about non-Google AI and cloud backup
P6	Yes	5/5	call: "it needs some guardrails to protect our data"

Totals: Calendar/Gmail yes from 5 of 5, 3 of them conditional on choosing which apps; reasons made sense 4, 4, 4, 5, 5.

Question: Retrace must work on how people actually phrase a half-remembered photo, not on sentences it was tuned on.

Method. Two query sets, each written by a separate AI agent that never saw the code, labels or library data; it got only a plain description of each target photo and was told to write queries the way different real people type them: very short, natural sentences, a wrong or fuzzy year, Hinglish or typos, and who/what/relative time. Each query is run through (a) the keyword-search stand-in and (b) Retrace as the person sees it (the grouped results page). Rank is among distinct photos (near-duplicates collapsed). Script: mvp/app/scripts/heldout.ts.

• Set A (80 queries, 10 photos) was used to find general weaknesses, so after the fixes it is no longer a clean test.
• Set B (60 queries, 10 different photos) was run once, after the fixes, and never used for tuning. Report set B.

Results

	Keyword search	Retrace
Set B: target found at all	34 / 60 (57%)	54 / 60 (90%)
Set B: target in top 5	32 / 60 (53%)	49 / 60 (82%)
Set B: target is the top result	18 / 60 (30%)	28 / 60 (47%)
Set A before fixes: found / top 5 / top 1	50 / 47 / 33 of 80	70 / 66 / 51 of 80
Set A after fixes (no longer clean)	50 / 47 / 33 of 80	78 / 73 / 61 of 80

General fixes made from set A (no phrase-specific rules)

1. A guessed moment is a strong hint, not a wall: photos outside it that fit the clues clearly better are still shown.
2. A year typed on its own also searches a year either side, with a note (years are often remembered a year off).
3. The fallback never jumps to a whole life period.
4. The model maps unknown words to the nearest known ones (khichdi → cooking, food) and only anchors a moment when the words point to one.
5. Data: the momos stall is also labelled "street food".

Found after set B (fixed, but set B numbers above are from before the fix)

• The IndiGo ticket screenshot's recognised text left out its header ("IndiGo"), so "indigo" failed in both modes.

Limits

• One fictional library; image labels are pre-written, standing in for real image understanding.
• Queries are AI-written, not from real users; Monday's tests give the human version.
• The model's reading of a query varies a little between runs; results are from one run each.

Each file is the "Copy results" export, pasted unchanged. Participant numbers in the files are /test numbers (P1 = 1, P2 = 3, P3 = 4, P5 = 5, P6 = 6; 2 = P1's repeat run). P4 did not take the test.

P1.tsv

participant	task	mode	found	seconds	searches	moment_taps	refines	fixes	opened	wrong_picks	ease	queries	comment
1	T1	normal	1	22	4	0	0	0	1	0	3	party with three friends (0) | party with three friends (0) | party with three friends (0) | party (25)	I searched party with three friends it didn't show me anything, then I searched party then it gave me bunch of photos
1	T2	normal	1	18	2	0	0	0	1	0	2	dengue prescription (0) | prescription (4)	When I searched for dengue prescription, it gave no results but when I searched for prescription it then gave me a result
1	T3	retrace	1	51	1	0	0	0	1	0	3	wedding Mandap (6 plain) | [Retrace: 1 matches, 9 moved, top moment Riya's wedding (Jaipur)]	found it
1	T4	retrace	1	13	1	0	0	0	1	0	5	cafe at the beach (7 plain) | [Retrace: 7 matches, 0 moved, top moment Goa trip]	
participant 1	consent: Yes, if I can choose which apps it uses	trust_reasons: 4	why:

P2.tsv

# P2 (call notes in research/sessions/P2.md). First run. Entered as participant 3 in /test (1 and 2 were P1's runs).
participant	task	mode	found	seconds	searches	moment_taps	refines	fixes	opened	wrong_picks	ease	queries	comment
3	T1	normal	1	16	1	0	0	0	1	0	5	club party (25)	easy to find
3	T2	retrace	1	19	1	0	0	0	1	0	5	doctor's prescription (4 plain) | [Retrace: 3 matches, 1 moved, top moment When you were unwell (Aug 2022)]	
3	T3	normal	1	10	1	0	0	0	1	0	5	wedding mandap (6)	
3	T4	retrace	1	12	1	0	0	0	1	0	5	beach cafe (7 plain) | [Retrace: 7 matches, 0 moved, top moment Goa trip]	
participant 3	consent: Yes, if I can choose which apps it uses	trust_reasons: 4	why: only if my data is stored locally and not online

P3.tsv

# P3 (call notes in research/sessions/P3.md). First run. Entered as participant 4 in /test.
participant	task	mode	found	seconds	searches	moment_taps	refines	fixes	opened	wrong_picks	ease	queries	comment
4	T1	retrace	1	19	1	0	0	0	1	0	5	party in 2018 with friends (1 plain) | [Retrace: 8 matches, 12 moved, top moment Freshers' Night]	
4	T2	normal	1	14	1	0	0	0	1	0	5	prescription (4)	
4	T3	retrace	1	13	1	0	0	0	1	0	5	wedding jaipur (16 plain) | [Retrace: 16 matches, 9 moved, top moment Riya's wedding (Jaipur)]	
4	T4	normal	1	11	1	0	0	0	1	0	5	goa sunset cafe (7)	
participant 4	consent: Yes	trust_reasons: 4	why: As long as search becomes easier

P5.tsv

# P5 (call notes in research/sessions/P5.md). First run. Entered as participant 5 in /test.
participant	task	mode	found	seconds	searches	moment_taps	refines	fixes	opened	wrong_picks	ease	queries	comment
5	T1	normal	1	36	4	0	0	0	1	0	4	party with 3 friends (6) | party with friends (6) | party with friends (6) | party (25)	It earlier gave me only selfie photos not the dance photos
5	T2	normal	1	12	1	0	0	0	1	0	5	Doctor's prescription (4)	
5	T3	retrace	1	12	1	0	0	0	1	0	5	riya wedding (14 plain) | [Retrace: 18 matches, 9 moved, top moment Riya's wedding (Jaipur)]	
5	T4	retrace	1	11	1	0	0	0	1	0	5	fries and cold coffee cafe (19 plain) | [Retrace: 19 matches, 0 moved, top moment Goa trip]	
participant 5	consent: Yes, if I can choose which apps it uses	trust_reasons: 5	why:

P6.tsv

# P6 (call notes in research/sessions/P6.md). First run. Entered as participant 6 in /test.
participant	task	mode	found	seconds	searches	moment_taps	refines	fixes	opened	wrong_picks	ease	queries	comment
6	T1	retrace	1	12	1	0	0	0	1	0	5	fresher's party (1 plain) | [Retrace: 1 matches, 12 moved, top moment Freshers' Night]	
6	T2	retrace	1	19	1	0	0	0	1	0	5	doctor's prescription (4 plain) | [Retrace: 3 matches, 1 moved, top moment When you were unwell (Aug 2022)]	
6	T3	normal	1	10	1	0	0	0	1	0	5	riya's wedding (26)	
6	T4	normal	1	12	1	0	0	0	1	0	5	sunset beach cafe goa (7)	
participant 6	consent: Yes	trust_reasons: 5	why:

P1_run2.tsv

# REPEAT RUN: P1 had already seen all four target photos in run 1. Not fresh data; report separately, never pooled.
# Entered as participant 2 so modes were swapped. Ran after the fairer normal search (5 Oct).
participant	task	mode	found	seconds	searches	moment_taps	refines	fixes	opened	wrong_picks	ease	queries	comment
2	T1	retrace	1	11	1	0	0	0	1	0	5	party with friends (6 plain) | [Retrace: 0 matches, 12 moved, top moment Freshers' Night]	
2	T2	retrace	1	16	1	0	0	0	1	0	5	medicine prescription (4 plain) | [Retrace: 3 matches, 1 moved, top moment When you were unwell (Aug 2022)]	
2	T3	normal	1	8	1	0	0	0	1	0	5	wedding mandap (6)	
2	T4	normal	1	14	1	0	0	0	2	0	5	beach side cafe (7)	
participant 2	consent: Yes, if I can choose which apps it uses	trust_reasons: 5	why:

Everyday photos without people in the prototype library are free stock images under a CC0 (public domain) licence, which needs no attribution; sources are listed for transparency. 133 images are used (near-duplicate copies made from some of them are not listed separately).

The fictional people were created from character reference portraits licensed from Adobe Stock; the scenes of their fictional life were generated with ChatGPT image generation. Documents, tickets, receipts and chat screenshots were drawn in code.

Image	Licence	Source
food_001	CC0	www.rawpixel.com
food_002	CC0	www.rawpixel.com
food_003	CC0	commons.wikimedia.org
food_008	CC0	commons.wikimedia.org
food_010	CC0	www.flickr.com
food_014	CC0	wordpress.org
food_015	CC0	www.flickr.com
food_016	CC0	www.flickr.com
food_017	CC0	www.flickr.com
food_018	CC0	wordpress.org
food_019	CC0	wordpress.org
food_020	CC0	wordpress.org
food_021	CC0	www.rawpixel.com
food_022	CC0	www.flickr.com
food_023	CC0	www.flickr.com
food_024	CC0	www.flickr.com
food_025	CC0	www.flickr.com
food_026	CC0	www.flickr.com
food_027	CC0	www.flickr.com
food_028	CC0	www.flickr.com
food_029	CC0	www.flickr.com
food_030	CC0	www.flickr.com
food_031	CC0	www.flickr.com
food_032	CC0	www.flickr.com
food_034	CC0	www.rawpixel.com
food_035	CC0	wordpress.org
food_036	CC0	www.rawpixel.com
food_037	CC0	wordpress.org
food_038	CC0	wordpress.org
food_039	CC0	wordpress.org
food_040	CC0	wordpress.org
sky_001	CC0	www.flickr.com
sky_002	CC0	stocksnap.io
sky_003	CC0	www.flickr.com
sky_004	CC0	stocksnap.io
sky_006	CC0	stocksnap.io
sky_007	CC0	stocksnap.io
sky_008	CC0	stocksnap.io
sky_009	CC0	stocksnap.io
sky_010	CC0	commons.wikimedia.org
sky_011	CC0	www.flickr.com
sky_012	CC0	stocksnap.io
sky_013	CC0	stocksnap.io
sky_014	CC0	stocksnap.io
sky_015	CC0	www.flickr.com
sky_016	CC0	stocksnap.io
sky_017	CC0	stocksnap.io
sky_018	CC0	stocksnap.io
sky_019	CC0	www.flickr.com
sky_020	CC0	www.flickr.com
sky_021	CC0	www.flickr.com
sky_022	CC0	www.flickr.com
sky_023	CC0	www.flickr.com
sky_024	CC0	www.flickr.com
sky_025	CC0	www.flickr.com
streets_003	CC0	www.flickr.com
streets_006	CC0	www.flickr.com
streets_012	CC0	www.rawpixel.com
streets_013	CC0	www.rawpixel.com
streets_015	CC0	www.flickr.com
streets_016	CC0	www.flickr.com
streets_017	CC0	www.flickr.com
streets_019	CC0	www.flickr.com
streets_020	CC0	www.flickr.com
streets_021	CC0	www.flickr.com
streets_022	CC0	www.flickr.com
streets_023	CC0	www.flickr.com
streets_024	CC0	www.flickr.com
streets_025	CC0	www.flickr.com
nature_001	CC0	www.flickr.com
nature_002	CC0	www.flickr.com
nature_003	CC0	www.flickr.com
nature_004	CC0	www.flickr.com
nature_005	CC0	www.flickr.com
nature_006	CC0	www.flickr.com
nature_007	CC0	www.flickr.com
nature_008	CC0	www.flickr.com
nature_009	CC0	www.flickr.com
nature_010	CC0	www.flickr.com
nature_011	CC0	www.flickr.com
nature_012	CC0	www.flickr.com
nature_013	CC0	www.flickr.com
nature_014	CC0	www.flickr.com
nature_015	CC0	www.flickr.com
nature_016	CC0	www.flickr.com
nature_017	CC0	www.flickr.com
nature_018	CC0	www.flickr.com
nature_019	CC0	www.flickr.com
nature_020	CC0	www.flickr.com
things_002	CC0	www.flickr.com
things_003	CC0	www.flickr.com
things_004	CC0	stocksnap.io
things_005	CC0	www.flickr.com
things_006	CC0	stocksnap.io
things_007	CC0	stocksnap.io
things_008	CC0	stocksnap.io
things_009	CC0	www.flickr.com
things_010	CC0	www.flickr.com
things_011	CC0	www.rawpixel.com
things_013	CC0	www.rawpixel.com
things_014	CC0	www.rawpixel.com
things_015	CC0	www.rawpixel.com
things_021	CC0	www.flickr.com
things_023	CC0	www.flickr.com
things_024	CC0	www.flickr.com
things_025	CC0	stocksnap.io
things_026	CC0	stocksnap.io
things_027	CC0	stocksnap.io
things_028	CC0	stocksnap.io
things_029	CC0	stocksnap.io
things_030	CC0	stocksnap.io
notes_001	CC0	www.flickr.com
notes_002	CC0	stocksnap.io
notes_003	CC0	stocksnap.io
notes_004	CC0	www.flickr.com
notes_006	CC0	www.flickr.com
notes_007	CC0	www.flickr.com
notes_014	CC0	wordpress.org
random_001	CC0	stocksnap.io
random_002	CC0	www.flickr.com
random_003	CC0	stocksnap.io
random_004	CC0	stocksnap.io
random_005	CC0	stocksnap.io
random_006	CC0	www.flickr.com
random_007	CC0	www.flickr.com
random_008	CC0	www.flickr.com
random_009	CC0	www.flickr.com
random_010	CC0	stocksnap.io
random_011	CC0	www.rawpixel.com
random_012	CC0	www.rawpixel.com
random_013	CC0	www.rawpixel.com
random_014	CC0	wordpress.org
random_015	CC0	www.flickr.com