About this replay · demo background
Same operating system as Cold case. This page replays a real day so the empty timers and people counts are not made up. The surface is the UI we would build. This fold is for the presenter.
State capture: 109,950 samples · Tue 11:32 AM → Sat 7:48 AM, zero gaps · one camera · eight drawn facings · location_id=3 · edge-camera-001 · this page replays Tue–Wed · people tracking runs 24h
States are frame-confirmed ground truth
(labels_hotcase_3.json).
Empty timers run on the event clock — from the cleanout or take that
emptied the facing. Blocked never counts as empty or stocked.
No claim means no confirmed ground truth. Low is not a confirmed empty.
Overnight grill empty is the nightly cleanout — it stays on the case,
not in the queue. Walk now = confirmed empty during selling hours
(headline + queue). Soon = running low during selling hours (warn queue,
not in the headline). No walk = scheduled cleanout, overnight empty,
blocked, or no claim. The foil tray at 6:23 PM is the walk-now empty
during selling hours. Warmer top cleared at 7:20 PM; the camera cannot
prove waste — do not say “thrown away.”
People are entity_visits
on zone hotcase_conversion
(zone 9), the Foodservice area’s conversion role. The per-moment window is
that day so far — store-local midnight to the selected moment. All
people windows are calendar days, midnight to midnight, America/Denver
(24h store, no open clip). Bands match this location’s Foodservice
area_config:
detection 2s, presence 5s, conversion 60s, long 120s.
Flicker is under 2s and ignored; Passing is 2–5s; Walked away is 5–60s;
Converted is 60s+; Long is 120s+. Walked away is Joey’s phrase for the
Abandoned dwell band. Walked away and Converted are configured
dwell proxies, not proof that a person did or did not take a product.
Product-in-hand VLM verification is not built. Staff are not filtered.
A door in that corner adds walk-bys.
Tracking zones are larger than the drawn facings, so people and facing
state are two measurements.
People tracking runs around the clock — beyond the labeled state
capture. The people panel and the time-to-walk-away histogram are
day so far: store-local midnight to the selected moment, every
number a real entity_visits query per moment. In production this
is the live panel, and it reads as the full day in history. The
by-hour chart dims the hours after the moment. Full-day totals
for reference: Tue 68 walked away / 21 Converted (7 while a
facing sat empty), Wed 83 / 21 (16).
State-label coverage: the grills and the foil tray have ground truth
for both full days (the multi-day label extension); warmer labels end
Wed 7:58 AM, and a blank ribbon means unlabeled, not stocked.
The Wed non-grill states on the afternoon moments are read from the
saved 10-minute frames: condiments stocked all day, and the foil tray
emptied between 1:15 and 1:25 PM — a sell-out, hours before the
cleanout — then was reloaded by 5:46 PM. That is about 4h 20m empty
during selling hours, the largest walk-now miss in the replay. The
warmer stays no claim: wrapped items on that shelf are below what
this camera can confirm at frame resolution. Outside this replay,
the labels also confirm a 20h foil sell-out the next day (Thu
7:37 AM → Fri 3:50 AM — Thursday's entire selling day).
“Visited while a facing was empty” is walked-away visits (5–60s)
whose start falls inside a labeled confirmed-empty span, across all
labeled facings on both full days; it is a subset of Walked away,
not all tracked visits, and it is not limited to selling hours. It
says a facing was empty during the visit. It does not say the person
wanted that product; the customer reads the impact.
Staff windows (cleanouts, reloads) are unlabeled and do NOT count.
Basis correction (2026-08-25): earlier drafts cited Tue 8 · Wed 11 —
those used the chart's ribbon spans, which run to the 2:44 AM
reload, so Wed's 11 counted 8 reload-crew visits inside the
unlabeled 2:34–2:44 AM staff window. The strict label basis gives
Tue 7 · Wed 16 (Wed grew because the afternoon foil sell-out and
the evening post-cleanout grills are now labeled).
The cleanout → reload grill window dominates empty time, overnight
empty there is expected, and the 2 AM spike in the by-hour chart is
the reload crew. The cleanout time drifts day to day — Tue 5:11 PM,
Wed ~3:25 PM, and 2:40 PM–5:10 PM across the five observed days —
so never assume a schedule. Median dwell in the Walked away band:
Tue 16s, Wed 18s. Inside the original labeled window there are 164
tracked visits — 63 Walked away (10 while a facing sat empty on the
strict basis), 22 Converted, 22 Passing (2–5s), and 57 Flicker
(under 2s).
Build readiness.
Artifact is a POC. Item-lane counts are a shipped primitive. Facing
state, empty duration, reload, and the hourly overlay are a labeled
prototype. Dwell bands and visited-while-empty overlap are derivable.
Running totals up to this moment. Grill time after 5:11 PM is the nightly cleanout.
Real counts, midnight to this moment. In production this panel is the live day so far, and it reads as the full day when you look at history. “Visited while a facing was empty” counts only starts inside confirmed-empty spans — the reload crew and cleanout windows do not count. It says a facing was empty, not that the person wanted it.