Mean tracked detections; smoothed one-second peak 8.8. Missed detections remain unmeasured.
OPERATIONS / OBSERVED EVIDENCE
See the pressure.
Inspect the evidence.
Find an interval worth reviewing. Understand what a zone measurement says, where tracking becomes uncertain and what evidence to collect before acting.
Licensed public footage · Existing 72-second analysis · No live camera, uploads or customer data
01 / INSPECT THE SCENE
A crossing, a zone,
a review window.
Observations come from the original licensed clip. This display is a smaller, globally blurred derivative; the analysis was not rerun on it.
ProtoplasmaKid / Wikimedia Commons / CC BY-SA 4.0. Changes: 72-second excerpt, resize, global blur, audio removed and H.264 encoding. This derivative and poster retain CC BY-SA 4.0. No endorsement implied.
EXISTING SENSORIUM → TYPED WORLD
Loading the aggregate event and its existing World phase…
Rule-based phase label from detector observations. This is not validated congestion or a forecast.Loading the published aggregate measurements…
0s36s72s
02 / KEEP THE CONTRACTS SEPARATE
Measurements with
their practical limits.
Configured directions IN / OUT. Accuracy requires independent crossing annotations.
Distinct track IDs, not distinct people. Association quality remains unvalidated.
No independent identity validation; the value is unavailable, not zero.
Read censored closures carefully.
143 observed entries; 37 geometric exits, 102 closures from a lost track and 4 from the end of the video. 74.13% of closures are censored. A lost track is not evidence that a person left the zone.
75 tracks span at most one second, including 16 single-observation tracks. These are review signals, not measured ID switches.
Forecast gate stays closed.
This 72-second clip cannot meet the current evaluation gate: at least 600 seconds of holdout and 120 origins per cell, with uncertainty against persistence. No forecast is promoted.
No validated queue time, congestion intervention, unique visitor count or business outcome is claimed.
03 / PROPOSED EVALUATION — NOT A COMPLETED CUSTOMER PILOT
Start with a question.
Agree how to measure it.
Proposed scope: one authorized stationary scene, two zones and one directional line, up to three 15-minute clips. Review a selected interval and compare results with independent annotations.
Occupancy and congestion review
Ask when a zone is busier and which interval deserves manual inspection. Congestion, queue time and staffing benefit remain hypotheses until separately measured.
Evaluation
Freeze zones and thresholds before heldout clips. Independently annotate at least 300 sampled frames and all crossings in chosen intervals. Measure zone count MAE, missed/extra detections and crossing precision/recall separately. Report uncertainty and each failure scene.
Acceptance thresholds require agreement before evaluation. No accuracy percentage is available from this demo.
Observed compute — this run only
- Engine wall time
- 25.87s
- CPU time
- 31.91s
- Peak process memory
- 648.8 MB
- Processed frames
- 360 at target 5fps
- Monetary cost
- UNKNOWN
DirectML GPU on the existing local machine. Hardware, energy, hosting, annotation and support prices are unmeasured; these timings do not establish a production cost per camera.
Pricing hypotheses
€250 for a bounded evidence evaluation; €49 per camera/month after readiness gates. Unapproved planning hypotheses, with tax, hosting and support scope unresolved. No checkout or service activation.
Privacy and retention proposal
Process only footage the customer is authorized to provide. Use zone aggregates and transient tracking; no face recognition, biometric embeddings or persistent person IDs.
Proposed evaluation retention: raw clips up to 7 days and agreed aggregate evidence up to 30 days, then verified deletion. This is a pilot policy to implement and test; it does not certify a production retention service.
The demo browser code has no upload, analytics beacon, cookies, local storage or private backend request. Hosting logs follow the site policy.
Support and launch gate
Agree an operator, contact route and review cadence before taking footage. Export evidence and record failures; no live safety decision or SLA is available.
TECHNICAL_MVP: yes
PILOT_READY: no
PRODUCTION_READY: no
COMMERCIAL_VALIDATED: no
Missing seams: independently annotated second scene, durable private video hosting and verified privacy/deletion boundary, approved price/terms and a supported customer workflow.
04 / SOURCE, MODEL AND VERSION
Know what produced
the observation.
Existing executed chain
- Licensed crosswalk excerpt, exact analyzed source hash.
- RT-DETR v2 R18 detector and ByteTrack association.
- OLYX / Sensorium v1 observations.
- Existing AVARY World zone and line summaries.
- SuperVision v2 evidence and retrospective review plan.
God's Eye View — design input only
Canonical code reference: commit 591f299d11f38a612629a274463196d57ae3862e, MIT code. Its source model informs provenance fields; its upstream runtime and layers are not connected to this demo.
No aircraft, weather, news or third-party world layer is loaded here. Code licensing does not license its data providers.
View the aggregate input model projection
Published source model is available in provenance.json.