case studies

Simple, repeatable tests
for checking vision model failures.

Each protocol says what to test, what to record, and what the result can and cannot prove. Evidence comes from running the perception lab, not from prefilled analytics.

Low-light reliability window
setup
input: webcam or calibration sample
preset: low-light route
model: browser-side COCO-SSD
cadence: browser-throttled inference
observation
mean confidence only when detections exist
empty inference frames
confidence variance across adjacent frames
required record
observation window
cadence jitter
FRAME.INTEGRITY
exported frame timestamps
proof boundary

The run can show detector output instability for this input and browser session. It cannot generalize to all detectors or deployments.

temporal traceevidence JSONdetection history
run protocol
Partial-visibility continuity
setup
input: uploaded scene or webcam
preset: partial visibility
degradation: occlusion + crop instability
model: browser-side COCO-SSD
observation
lost classes between adjacent frames
drop and recovery transitions
baseline comparison when an upload is used
required record
TRACKING.PERSISTENCE
continuity transitions
drop events
class continuity breaks
proof boundary

Continuity markers are class-level observations from model outputs. They do not establish object identity or tracking persistence beyond emitted detections.

continuity markersdrop eventsbaseline readout
run protocol
Compression frame integrity
setup
input: local image or webcam
preset: compressed feed
degradation: noise and contrast pressure
model: browser-side COCO-SSD
observation
empty-frame rate
class instability
confidence range for detected frames
required record
empty frames
min confidence
max confidence
observed classes
proof boundary

The record exposes whether this run produced unusable inference frames. It does not infer why the detector failed beyond recorded degradation settings.

FRAME.INTEGRITYconfidence rangeevidence packet
run protocol
Motion consistency trace
setup
input: webcam movement or sample
preset: motion instability
degradation: blur + motion offset
model: browser-side COCO-SSD
observation
adjacent-frame confidence deltas
class continuity breaks
replay index movement
required record
TEMPORAL.CONSISTENCY
cadence seconds
cadence jitter
replay frame state
proof boundary

The trace documents emitted detection history. It does not synthesize video replay frames or estimate motion vectors.

temporal replaycadence readoutclass breaks
run protocol
case-study run sheet
checklist structure / no prefilled run result
01 / input selected

requires webcam, upload, or imported calibration sample

02 / degradation configured

preset and manual controls must be visible before inference

03 / model loaded

browser-side model state must be available or explicitly unavailable

04 / observation window held

frame history must exist before temporal claims

05 / packet exported

evidence JSON is created only after real output history exists

case-study evidence boundary
real inputs / real outputs / explicit unavailable states
source data

papers, releases, policy pages, ingested records

traceable
model behavior

browser-side detections, confidence, timing, empty frames

observed
evidence packet

timestamps, frame history, continuity markers, exported JSON

recorded
claim boundary

missing data stays unavailable; outcomes are not prewritten

enforced
real source

Claims need source records or ingestion events.

model output only

Confidence and detections come from inference.

unavailable is valid

Missing data stays visible instead of guessed.

no prefilled claims

Protocols do not ship with conclusions.

real-world visual context
illustrative photos / not model evidence

These images show the kinds of real-world conditions SIGNALWATCH is designed to explain: low light, camera deployment, motion blur, sensor boundaries, and human monitoring. They are source-attributed visual context, not detections or evaluation results.

visual role
operational context
source status
externally attributed
evidence boundary
not inference output
dataset sequence registry
protocols only until real frames are imported
low-light / capture required
12 frame minimum
Low-light hallway sequence

Low illumination is common in deployed indoor monitoring paths and can hide detector failure states.

observationWindow
frames
continuityTransitions
degradation
occlusion / capture required
12 frame minimum
Partial visibility and occlusion

Partial visibility can make single-frame detections look reliable while temporal continuity is unstable.

continuityTransitions
packet
frames
observationWindow
motion / capture required
16 frame minimum
Webcam motion instability

Motion pressure can expose unstable inference cadence and adjacent-frame detection volatility.

observationWindow
cadenceJitterSeconds
frames
temporalTrace
compression / capture required
12 frame minimum
Compression artifact sequence

Compressed feeds can make missed detections look like normal absence unless empty frames are counted.

packet
emptyFrames
minConfidence
maxConfidence
continuity / capture required
16 frame minimum
Overlapping object continuity

Object overlap is a common source of continuity confusion in operational perception surfaces.

continuityTransitions
frames
operationalObservations
observationWindow
evidence protocol
01select input
02apply degradation preset
03run browser-side inference
04hold observation window
05inspect continuity markers
06export evidence JSON
07compare records without generalizing beyond the run
proof boundary
can show

confidence instability, empty frames, dropped detections, and continuity breaks in this browser/model/input session

can show

how a selected degradation changes the emitted COCO-SSD output history

cannot prove

universal model failure across all detectors, datasets, environments, or deployments

cannot prove

overall safety of a deployed perception system without broader evaluation coverage

case-study export shape
schema preview / values require real run
schema
signalwatch.perception.evidence.v1
generatedAt
set when a real run is exported
model
browser-side COCO-SSD when loaded
inferenceBoundary
local browser inference only
observationWindow
derived from actual frame timestamps
frames
empty until frames are observed
detections
empty until model emits detections
continuityTransitions
computed from emitted class history
non-populated export shape
{
  "schema": "signalwatch.perception.evidence.v1",
  "generatedAt": "<real export timestamp>",
  "model": "<loaded browser model>",
  "inferenceBoundary": "real outputs only",
  "frames": [],
  "detections": [],
  "continuityTransitions": []
}
case-study unavailable states
absence is observable / absence is not guessed
model unavailable
perception lab

If browser-side inference cannot load, detections remain unavailable instead of simulated.

no detections emitted
model output

Empty frames are recorded as empty frames; no bounding boxes or confidence values are fabricated.

collector offline
source ingestion

Source health reports offline or delayed states directly through collector telemetry.

insufficient window
evaluation

Temporal claims wait until a real observation window has enough frames or source events.

packet not exported
evidence packet

The interface can show the export schema without pretending an evidence packet exists.

source missing
provenance

A claim without a usable source trace should stay unresolved, not become narrative filler.

reproducibility boundary

Case-study evidence should be captured from the operational evidence packet: frame count, empty frames, confidence history, detection drop events, class continuity breaks, and exported JSON records. If the model emits no detections, the record should state that plainly.

schema
generatedAt
model
inferenceBoundary
observationWindow
operationalObservations
continuityTransitions
frames