The Black Box for Edge AI: Introducing Field Recorder

Your CI proves a model was good before it shipped. Nothing proves what it did after. Field Recorder closes that gap.

EdgeGate Engineering Team

Edge AI CI/CD platform · Qualcomm AI Hub integration partners

Testing ends at the shipping door

Pre-deployment gates are the right place to catch a bad model version: run it on real hardware, block the release if it regresses. EdgeGate has done this from day one. But a gate only sees the moment before you ship. The device then runs for months — through firmware updates, thermal swings, silent model swaps, and quantization that behaves differently in the field than it did on the bench.

In the cloud, that gap doesn’t exist: every inference passes through a server you control, so you have logs by default. On a device — offline, on a quantized model, on whatever silicon the product line uses — there is no record. When a customer, an auditor, or a regulator asks “prove what your model actually did on that device,” a green checkmark from six months ago isn’t an answer.

A black box, in the aviation sense

Aircraft carry a flight recorder for one reason: after something goes wrong, you can pull the box and prove exactly what happened, in a form no one can quietly alter. Field Recorder is the same idea for edge AI. It wraps your inference call — one context manager, decorator, or line — and records a compact event for every consequential decision: the model’s hash, the device and chip, a privacy-safe digest of the input, the output, confidence, and latency.

Each event is hash-chained to the one before it, and the chain is signed with a per-device Ed25519 key. Alter or delete a single event after the fact and the chain breaks — verification fails. That’s the difference between a log file and evidence: a log you can edit; a tamper-evident chain you cannot, undetected.

Two timestamps, one evidence chain

The certified model your pre-deploy gate signs at CI time is the same reference Field Recorder replays production decisions against. Pre-deployment evidence (“we tested it”) and runtime evidence (“here’s what it did”) aren’t two products — they’re two timestamps on one signed chain. Replay catches the drifts that matter: a model quietly falling back from the NPU to the CPU, a quantized build diverging from the certified one, or the same input suddenly producing a different answer than the baseline recording did.

What leaves the device — and what doesn’t

Raw inputs and raw outputs never leave the device. Field Recorder stores only one-way digests of them, plus the signed metadata. You get a verifiable audit trail without shipping sensitive data off-box — and it keeps recording offline, so an air-gapped device loses nothing while disconnected.

Recording is local, offline, and free. The one step that needs an API key is syncing those signed events to your dashboard — where the timeline, replay-divergence, and Article 12 export live. Record all day for nothing; the evidence layer is the paid product.

Built for the regulation that’s already here

EU AI Act Article 12 requires high-risk AI systems to automatically keep records over their lifetime, and those obligations are phasing in now. Field Recorder’s export is shaped for exactly that — a signed bundle your compliance team can hand to an assessor. One honest note we hold to: Field Recorder is tamper-evident, not tamper-proof, and it produces evidence. Whether that evidence satisfies a specific regulation is a call for your auditor — our job is to make the record unforgeable-to-alter and verifiable years later.