When something contested happens in public — a shooting, a raid, a protest — the record of it now assembles in real time from official statements, bystander phone footage, and an immediate flood of AI-generated fakes, doctored frames, and misattributed photos. The footage that could actually settle what happened surfaces in scattered pieces, and it routinely takes OSINT investigators days of manual work to sync those pieces into a coherent account. My final presentation opens with exactly such a case study: the days between the event and the reconstruction are where false narratives take hold.
The existing tools — manual forensics, AI-detection classifiers, provenance standards like C2PA — all work on the question "is this image real?" It's a question worth answering, but it isn't the one that resolves contested events. That question is "does other evidence of this moment exist?" Authentication has an emerging ecosystem; corroboration is still a slow, manual, heroic effort.
PIGEON makes corroboration searchable.
The device
PIGEON is a wrist-worn beacon built from three parts — a GNSS module, a microcontroller, and an e-ink display — for under £15, fully open source. The display shows a constantly-refreshing truchet pattern that encodes the device's ID, a per-device counter, a timestamp, and coordinates, authenticated with an HMAC and wrapped in Reed–Solomon error correction.
Anyone who films the wearer captures the pattern as a side effect. The encoding is designed to survive the hostile pipeline real footage actually goes through — odd angles, motion blur, JPEG compression, cropping, screenshots of screenshots — and still decode from a single photo. E-ink was chosen deliberately: it's sunlight-readable, holds an image at zero power, and is robust in the field. The form is deliberately not a QR code aesthetic — the decoding decals on the faceplate double as visual identity, and the exposed screws invite modification of the open hardware.
The log
Decoding a captured pattern does two things at once: it returns the time, place, and authenticity context to whoever submitted the image, and it appends a leaf to an append-only transparency log — the same lineage as certificate transparency. Because every capture of the same pattern points at the same moment, corroboration becomes a lookup instead of an investigation: submit one clip, find every other verified capture of that moment.
The log's architecture carries the trust model. It's append-only, so tampering is detectable by anyone. Image storage is kept separate from the record, so content can be redacted without breaking the log. And it's federated — any organization can run a node, and the pattern itself encodes which log it reports to, so there's no single gatekeeper. Wearer privacy is opt-in by architecture: the wearer chooses to broadcast, and patterns only enter a log when someone decodes them.
Validation
The problem framing was validated through expert interviews — with people from the Content Authenticity Initiative, WITNESS, the VII Foundation, and the Centre for Information Resilience — which confirmed the two core claims: corroboration is the unmet need, and citizen witnesses are the users.
The third claim — that the pattern survives the real world — was validated in the field. I took the working prototype around London and photographed it in real conditions; every capture decoded to the correct time and place, replayed patterns were rejected by the counter check, and each decode landed as a pin on the map in the evidence browser.
Where it's going
I would love to take this project further. The next steps are: hardening the pattern generation and getting it to work on a wider array of devices; finalizing the spec and distributing log nodes to OSINT organizations; surfacing corroboration in-feed through a browser extension and social bots; and investigate moving beyond the wristwatch (such as time-only stamps for live broadcasts, orientation sensing).
The full gateway presentation is at pigeon.bencooper.xyz. It's desktop-only — move through it with the arrow keys and open deep dives with spacebar.
