Methodology
One standard, four kinds of media.
Every upload starts the same way: the original file's SHA-256 digest is computed locally and logged to the audit trail before any analysis runs.
01 · Authenticity & manipulation
Image
Method
Image-level manipulation detection plus localization models described exactly as they were trained and evaluated. Frequency-domain (2D FFT) inspection is shown as a separate signal, never folded into a verdict.
What it cannot conclude
A heatmap shows where a model responded — not proof that a specific pixel or face was altered.
02 · Edited-text inspection
Document
Method
Coordinate-aware layout analysis and aligned difference maps against a supplied original. Personal identifiers are masked by default in previews, logs and reports.
What it cannot conclude
A difference map proves two aligned files differ there; it does not prove fraud or reveal who changed them. Not an identity-verification service.
03 · Temporal deepfake evidence
Video
Method
Frames are sampled and scored per chunk so every finding maps to a timestamp range. Optional audio-visual checks only when a validated model is configured.
What it cannot conclude
Low resolution, heavy compression or few faces lower input quality — and PiperTrace says so instead of guessing.
04 · Synthetic-voice & signal
Audio
Method
If a permitted synthetic-audio checkpoint is installed, its output is shown with its documented semantics. Otherwise PiperTrace presents signal inspection, clearly labelled as not voice-clone detection.
What it cannot conclude
No checkpoint means “model not configured” — never a fabricated score.