Research foundations

Built on nine research domains. Not one model.

Most detectors are a classifier with a marketing layer. This system is the convergence of nine research streams — forensics, speech science, conversational authenticity, probabilistic evidence fusion, resilience and governance, privacy-by-design, and consumer/analyst workflow. Fused, they do something a classifier can't: evaluate whether a communication is trustworthy enough to act on.

What we actually evaluate

Forensics

Does the media carry synthetic or manipulated traces?

Interaction authenticity

Does the exchange behave like a genuine human interaction?

Identity & authority

Is the claimed person, role or source sufficiently linked?

Policy & reliance

What action is justified under this much uncertainty?

Forensics → interaction authenticity → identity & authority → policy & reliance. That sequence is the difference between an artifact classifier and a trust-evaluation decision system.

The research streams

01Multimedia forensics

Manipulated media leaves traces across image, video, audio, compression and provenance. We read them as distinct evidence planes — signal interpreted within evidential logic and diagnostic limits, not a single black-box score.

02Speech & acoustic science

Source–filter reasoning, spectral and cepstral analysis, LPC-style features, channel and microphone conditions. Audio is its own discipline with a dedicated engine, not an afterthought to video.

03Interaction authenticity

Conversational timing, repair behaviour, interruption and turn-taking, response-latency realism, challenge–response dynamics. Beyond static media: does the exchange coordinate like real human conversation?

04Identity & authority reasoning

Identity linkage, entity continuity, role plausibility and authority claims — moving from "is the media fake?" to "is the claimed person, role or source sufficiently linked to act on?"

05Bayesian evidence fusion

Heterogeneous evidence combined with decision theory into a calibrated score, an explicit uncertainty band, and the named factors that drove it — not raw-score theatrics. Deterministic rules stay separate from probabilistic fusion.

06Systems architecture

Explicit trust boundaries, contracts-first interfaces, separate control and data planes, and auditability from day one — designed to withstand enterprise deployment, legal scrutiny and adversarial inputs.

07Resilience & governance

Practical drift, normalisation of deviance, and work-as-imagined vs work-as-done. The reason this is built as a governance system: evidence accumulates across weak and strong indicators, overconfidence is dangerous, and abstention is a feature.

08Privacy by design

Data minimisation, ephemeral media, retention discipline, signed access, production-vs-training separation, immutable audit — and raw payloads that never leak into logs.

09Product & workflow

Action-first clarity and plain language for consumers; evidence-by-plane, replayable timelines, override/notes and exportable case records for analysts. The interface is research, not decoration.

How we talk about this

These are the research foundations and influences the architecture is built on — presented as the intellectual basis of the system, not as third-party endorsement or validation of any specific result. Where a result isn't calibrated, the engine abstains rather than assert. That is the same discipline that lets us publish a per-use-case error rate instead of a marketing claim.