Methodology & limitations

Metadata inspection

Camera files may contain EXIF fields describing the device, exposure, capture time and GPS position. Editing applications can add software tags. Parallax uses exifr to decode these fields entirely in your browser, alongside supported XMP and IPTC data.

Why it helps

Compare metadata against a claim: a date, camera model or coordinate may support a lead or expose a contradiction. The SHA-256 fingerprint lets you distinguish the exact original bytes from a recompressed copy.

Where it stops

Metadata can be changed or copied. Social platforms often remove it, and an empty metadata panel is not evidence of deception. A software tag does not prove misleading editing. Most EXIF timestamps have no time zone unless a separate offset is present.

Reverse image search

Search indexes compare visual patterns to find copies, crops and similar scenes. Use multiple providers because their coverage differs. An older appearance can show that an image predates the event attached to it.

A practical sequence

Open a provider, upload the original image, and inspect the oldest relevant results. Compare crops and captions. Record the original page URL, publication date and an archived copy in your report notes.

Where it stops

Parallax opens the search service; it cannot transfer a local file in a URL or read its results. Public-image URLs can be passed directly. Search rankings and publication dates are not capture dates. No matches does not mean an image is new or authentic.

Visual clue analysis

Language, road markings, architecture, plants, vehicles and weather may narrow a region. Look for independent clues that agree, and actively search for alternate explanations. Vehicles travel; architectural styles cross borders.

Inspect details and compare hypotheses

Each category has an Analyse further action and a separate conversation. Selecting a region crops the original image before resizing, which retains available detail that can disappear in a whole-scene overview. A crop cannot recover pixels missing from the original. Candidate identifications should cite visible features and alternatives; identifying a cultivated flower does not establish its native range as the location.

Location and capture-time candidates use qualitative labels: Likely means distinctive evidence favors the candidate; Plausible means compatible but not distinguishing; Unlikely means contradicted by a visible clue. Unresolved records missing evidence. These are model judgments, not verified matches or calibrated probabilities. The model has no browsing or map tools. Compare independent reference images before accepting a location, and avoid inferring a calendar date from season or daylight alone.

AI is optional

Configure Anthropic, OpenAI, Gemini, or a local/OpenAI-compatible vision model in Provider settings. Supply your own key and explicitly run analysis. Cloud requests pass through Parallax; local and compatible requests go directly from the browser to your endpoint. Demo scenes use hand-authored example observations. Keys can be saved in an unencrypted local config file or browser storage and are never included in reports or investigation history.

Where it stops

AI can misread text, invent observations or overstate a regional pattern. Confidence labels are qualitative, not calibrated probabilities of location. Plate observations are limited to region; identifying people and transcribing plate numbers are excluded.

Shadow geometry

For a candidate latitude, longitude and date, SunCalc 1.9 estimates the sun’s direction throughout the day. A vertical object’s ground shadow points opposite the sun’s horizontal bearing.

What Parallax calculates

Mark the object base, its top and the shadow tip. Calibrate north clockwise from image-up. Parallax uses the base-to-shadow-tip bearing, scans the candidate day at one-minute intervals, and returns daylight matches within your angular tolerance. The entered UTC offset defines that day. Sun altitude is predicted for the selected match, not measured from image lengths.

The essential limitation

A normal oblique photograph distorts ground angles. Use a level ground plane rectified to a top-down view with independently known north. The object top is recorded only for context; it does not correct perspective. Sloped ground, leaning objects, artificial lights and uncertain shadow endpoints can invalidate the estimate. Earliest and latest matches need not form a continuous interval. A match is a hypothesis, not a timestamp.