Video integrity
& frame analysis.

A video file.

• Inventory the available container, codec and stream fields.
• Inspect available timestamps and software traces.
• Use the technical record to guide deeper content-level examination.

Metadata is editable. A clean result does not establish an unaltered video, and an anomaly does not itself prove tampering.

A video supported by the selected container parser.

• Read the top-level layout of eligible MP4 or MOV files.
• Compare component order, counts and declared sizes.
• Flag unusual layout for deeper structural examination.

This quick check does not inspect component contents or stream data. Re-authoring can preserve normal structure. Confirm parser coverage for the source container.

A supported video file, with its claimed source where known.

• Read container fields, encoded-stream parameters, metadata and encoder strings across the file.
• Compare features with available device, recorder, editor, sharing-platform and encoder references.
• Assess whether the structure is consistent with the device claimed in metadata.

Structure indicates encoding or packaging families, not a unique device or proven origin. Only families represented in the library can be recognised. A closest reference is a consistency observation.

A video retaining suitable encoded-stream information.

• Inspect periodic frame-encoding differences that may preserve an earlier encoding’s structure.
• Compare coded size with picture change as an additional, weaker signal.
• Test the strongest periodicity against shuffled data and discount a period matching the file’s own encoding cycle.

Strongest when an earlier encoding was coarser than a later one. Fine earlier encodings may leave little trace; support and signal strength depend on the codec. An encoding-history lead is not proof of malicious editing.

A video file; examiner-entered clock anchors are needed for estimated wall-clock mapping.

• Record each decoded frame’s timestamp, picture type, packet size and pixel/visual fingerprints.
• Inspect variable frame intervals, non-advancing timestamps, gaps, frozen stretches and returns to earlier content.
• Review encoded frame-group structure and interlacing.
• Map frames to estimated clock time only with stated anchors, uncertainty and recorder drift.

Ordinary recorders can drop, repeat or re-time frames. File timestamps are not independently proven capture times, and estimated missing-frame counts or clock mappings require context.

A video file.

• Decode and fingerprint frames, with caps for long recordings.
• Compare each frame with its immediate predecessor.
• Inspect container-timestamp gaps alongside the decoded frames.

A later reinsertion of an earlier frame may be missed, and re-encoding can affect recognition. The workbench’s broader authentication workflow also checks cloned sequences; report the actual coverage used.

Camera footage; optional examiner-defined regions and exclusions.

• Measure frame-to-frame scene change and group it into review segments.
• Separate sudden lighting or infrared-switching changes rather than count them as activity.
• Optionally add screening labels for people, bicycles, motorcycles, cars, buses or trucks, marked as machine-learning output.
• Retain a reviewed/not-yet-reviewed record for each segment.

Segments are leads to watch, not conclusions about events or identity. Activity scores and optional labels require review; lighting, camera movement and occlusion affect what is visible.

Several camera recordings with suitable shared-frame, visible-clock or audio-event anchors.

• Use shared events, examiner-transcribed on-screen clocks or sufficiently strong audio correspondences.
• Estimate offsets and drift when anchors are far enough apart to support them.
• Flag disagreeing methods rather than silently average incompatible estimates.
• Keep unanchored cameras unsynchronised and fingerprint the history of anchor changes.

A common timeline is an estimate with uncertainty, not an independently established capture clock. Weak correspondences are discarded; missing anchors are not guessed.

A video file.

• Sample up to 12 evenly spaced frames in the quick workflow.
• Assess each sampled frame independently for a consistent duplicated-region signal.
• Use the workbench authentication workflow for keyframe and shot-start coverage across the recording.

Quick sampling leaves intervening frames unchecked and supplies no cross-frame temporal comparison. Workbench coverage across a recording is still defined by the frames it actually examines.

A compatible encoded video stream; coverage depends on the codec.

• Read block-level decisions from the encoded stream and show sampled motion fields.
• Display block-type, zero-motion, compression-setting and motion-discontinuity maps.
• Flag a textured region that remains motionless while the surrounding camera view moves, only where the required frame type and camera motion are present.

Other maps are displayed for examination without scores. Encoding choices, scene structure and camera motion affect interpretation. Coverage includes supported AVC and older MPEG streams; modern high-efficiency exports are not supported by this workflow.

Video with detectable faces.

• Locate faces in sampled frames and prepare face-image crops.
• Score those face images for patterns represented in the detector’s training material.
• Associate each output with the sampled frame and timestamp.

This workflow classifies sampled face images. It does not establish full-sequence temporal analysis, lip-sync checking or comprehensive synthetic-video detection.

A video file; confirm accepted formats before submission.

• Extract visual feature descriptions from selected video imagery.
• Apply a trained probe to the extracted visual features.
• Retain the sampling context and detector configuration with the output.

Temporal processing, thresholds, training coverage and generation-specific performance require confirmation. The component’s presence alone does not establish those capabilities.

Questioned footage plus reference videos from the candidate camera at the same resolution and recording mode.

• Build a reference fingerprint at native resolution, using suitable keyframes first.
• Compare the questioned recording overall and segment by segment.
• Inspect segments where agreement with the candidate fingerprint changes.

Diagnostic only: calibration uses synthetic patterns and real-device validation is pending, so no match verdict is supplied. Stabilisation and strong compression weaken the signal; missing agreement is not itself proof of an inserted segment.

A supported recorder export, raw video stream or multi-camera source; raw streams need an examiner-stated frame rate.

• Copy the encoded streams into a standard container and split eligible multi-camera sources into separate camera copies.
• Check every decoded output frame is pixel-identical to its source counterpart.
• Refuse registration if that verification finds a difference.
• Record raw-stream frame-rate assumptions and custody relationships on both sides.

Compatibility depends on the source streams and decoder. A playable derivative does not independently validate recorder timing, and an assumed frame rate must remain visible.

A video with suitable neighbouring frames.

• Align a short burst of consecutive frames around each anchor.
• Combine captured pixel information to reduce sensor and compression noise.
• Apply local contrast adjustments without generative reconstruction or upscaling.

A viewing aid alongside the original, not replacement evidence. Alignment and available captured information limit the result; this does not recover detail beyond what the recording contains.

A verified working source and selected frame ranges.

• Navigate using the decoder’s frame index, including frame type, timestamp, packet size and pixel fingerprint.
• Select frame ranges and optionally combine redundant frames in static scenes.
• Render a lossless Matroska output, preserve original timestamps and optionally create a separately labelled lossy presentation copy.
• Replay the processing project and verify every output frame.

No generative enhancement. Retain the untouched original alongside every derived copy, disclose processing choices and distinguish the verified lossless output from presentation compression.

A verified frame plus suitable scale/calibration; a stand-in requires the same ground position and an unchanged camera.

• Use a known reference, calibrated view or suitable ground-control geometry on the selected frame.
• Where appropriate, compare a known-height stand-in filmed at the same spot by the same unmoved camera.
• Report worst-case bounds and a probability-based uncertainty interval.
• Reviewed exports use two-person sign-off.

Perspective, point choice, reference suitability and camera movement determine validity. A stand-in comparison fails if the camera moved; report assumptions and uncertainty rather than a bare number.

Native footage with defensible timing and suitable geometry, road markings or examiner-confirmed device telemetry.

• Check native timing and re-encoding before deciding whether a speed estimate can be reported; refuse processed derivatives.
• Use suitable positions between two marked frames, verified road-marking observations or confirmed device telemetry.
• Cross-check independent approaches where the available evidence supports them.
• Report worst-case bounds, probability-based intervals and the two-person sign-off state.

Some recordings cannot support a speed figure. Geometry, timing, reference assumptions and telemetry provenance must be defensible. Indicative estimates are not equivalent to a fully supported measurement.