v4 · The receipt

v3 said the aha takes a minute.
Here is the minute.

Versions 1–3 argued about flow order, paywalls and personas. All of it rests on one claim: point the camera at a messy drawer and the app names what's inside, fast enough to feel like magic. On July 21 we ran that claim end-to-end — real shutter tap, real upload, real production AI, real database row. This page is the receipt.

Live capture

The uncut journey: shutter → upload → AI → Detected Items

capture receipt● verified

shutter	Jul 21, 2026 · 16:18:43 UTC
AI complete	16:19:16 UTC
wall clock	33seconds, snap → named
items found	8of 8 in frame
pipeline	production · gemini-2.5-flash-lite

row D22A58F3 · images table, production Postgres.
Not a comp. Not a prototype. A database row.

What the AI wrote down

Eight rows in the items table, named by the vision model with no human help. The names below are pasted from the database, not typed for this page.

01 keys on keychain 02 orange handled scissors 03 yellow tape measure 04 4 aaa batteries 05 usb-c charging cable 06 black electrical tape 07 blue lighter 08 red and grey marker

The clock, honestly

The onboarding thesis needs three numbers to be small. Two already are; one is the app's whole job and it held.

0:00 → 0:04

Open to shutter

Camera-first means the first screen is the lens. Nothing to read, one thing to press.

0:04 → 0:37

Snap to named

33 seconds of real upload + real AI. The badge walks Saving → Uploading → Analyzing → Done while the user does nothing.

any time later

Ask to found

Typed "scissors" into search — the drawer capture came back as orange handled scissors. That's the retention loop.

What this proves for the v3 flow

Each locked decision from v3, against what the live run actually demonstrated.

Camera-first default

The recording opens on the viewfinder, not a slide deck. The distance from launch to value is one tap — that is the entire argument for the Hybrid's default lane, running.

One shared aha

Both lanes of the Hybrid converge on the first snap. The receipt shows the aha is not aspirational: 8 items, named, in 33 seconds, on the production pipeline a user would hit.

Value before the ask

Everything in the video happens before any paywall would appear under the reverse-trial design — the free-through-first-find gate has real value to give away.

Search is the retention loop

"Where are the scissors?" is the question the app exists to answer. Asked it, got the answer, 54 minutes after the snap, from a cold search box.

What's real, what's staged

Recorded 21 July 2026, iPhone 17 Pro simulator, production backend.

Real: every pixel of app UI; the upload; the AI call (production process-image, gemini-2.5-flash-lite); the 8 item names; the sync badges; the search result; the timestamps. The pipeline had zero human assists — the same code path a TestFlight user exercises.

Staged: the drawer photo itself. The simulator has no physical camera, so the shutter "captures" a pre-shot photograph of a real junk drawer through a debug fixture — the exact bytes then travel the real pipeline. On hardware, the only difference is which camera took the picture.

Where this leaves the exploration

The direction is decided (v3), and the machine behind it is proven (v4). What remains is construction: the shipped app still opens with the old slide deck — the camera-first spine, the guided escape, segmentation and the reverse trial are all still to build. The next artifact should be the Hybrid running in the app itself, not another page about it.

Next → v5 · The ten-second story Back to tracker