ScammerProbability of scamming you 7–10/10
SCAMDAR SI is confident the image shows a scammer. The strongest signals all point the same way.
SCAMDAR-VISION v2.1 · ONLINE
From pixels
to a verdict.
Here's how SCAMDAR SI reads a face, step by step, in the few seconds between upload and result.
00 / PIPELINE
Four stages. A few seconds.
Every image goes through the same pipeline. You can watch it happen live in the agent log on your result page.
01 / DETECT
Face detection & alignment
Metadata is stripped on arrival, then a multi-scale face detector locates the subject. A 5-point affine transform aligns eyes, nose and mouth to a canonical pose, correcting head roll, yaw and lens distortion before a single measurement is taken.
scamdar-agent · stage 01[00:00]Metadata stripped · EXIF, GPS, XMP[00:00]Face detected · confidence 0.993[00:01]Aligned to canonical pose · roll 1.1° yaw −4.2°02 / MESH
468-point landmark mesh
SCAMDAR SI fits a dense 3D landmark mesh to the face and reconstructs its geometry: canthal tilt, interpupillary distance, gonial angle, philtrum length and the facial width-to-height ratio (fWHR).
scamdar-agent · stage 02[00:01]Fitting 468-point landmark mesh…[00:02]Canthal tilt · +3.8° · IPD 62.4mm[00:02]fWHR · 1.94 · gonial angle 118°03 / READ
Physiognomic feature read
Each facial region is scored against 42 classical physiognomic indices, from brow ridge and glabella to nasolabial folds and lip symmetry. A FACS encoder reads 17 action units for micro-expressions, including the Duchenne markers (AU6 + AU12) that separate a genuine smile from a rehearsed one.
scamdar-agent · stage 03[00:03]Scoring 42 physiognomic indices…[00:03]FACS · AU6 0.12 · AU12 0.81 · AU14 0.44[00:04]Non-Duchenne smile detected04 / INFER
Trust-score inference
Landmark geometry and expression vectors are fused into a 512-dimension facial embedding and passed through an ensemble of five classifier heads. Their votes are temperature-calibrated into a single trust score, then mapped to one of three verdicts.
scamdar-agent · stage 04[00:04]Projecting 512-d facial embedding…[00:05]Ensemble vote · 5/5 heads agree[00:05]Calibrated trust score · 0.27
05 / VERDICTS
One image. One of three labels.
SCAMDAR SI always commits to a single verdict. There's no partial credit and no blank result.
Maybe ScammerProbability of scamming you 3–6/10
SCAMDAR SI found some scammer signals, but not enough to be sure. A borderline call.
SafeProbability of scamming you 1–2/10
Nothing was flagged. SCAMDAR SI found no meaningful scammer signals in the image.
06 / TIMING
Seconds, usually
Most images finish in a few seconds. Under heavy demand it can take longer. Your result link keeps working, and the verdict appears there as soon as it's ready.
07 / PRIVACY
Metadata stripped
Location, camera and author details are removed from every upload before analysis. Images that can't be processed are never shown publicly.
08 / GALLERY
Results are public
Every analyzed image appears in the public gallery and live feed with its verdict, and every result has a link you can share.