Probabilistic signals, human review boundaries, and what scores do not mean in procurement
TrustOriginality surfaces confidence-style signals and trust-oriented summaries across multimodal analysis. This guide explains how to interpret those outputs responsibly in editorial, fraud, and compliance workflows — without treating scores as legal verdicts or automated block/allow decisions.
Article
What "Trust Score" means here
Trust-oriented scores summarize model confidence and signal strength from forensic analysis — not a credit score, legal finding, or regulator-approved metric. They help reviewers prioritize attention; they do not replace human judgment or contractual SLAs.
Scores across modalities
Text, image, audio, and video each produce analysis outputs with different signal types (e.g., spectral anomalies, lip-sync drift, token-level patterns). Compare scores only within the same modality and analyzer version — not across unlike asset types.
Human-in-the-loop requirements
Define policy thresholds for escalate vs publish vs reject. Log reviewer identity separately from model output. For Annex III-sensitive contexts, ensure no sole automated decision — scores inform review; humans decide.
Signed PDF and QR validation
Export signed reports for audit trails. QR lookup validates report integrity — useful for procurement and legal follow-up. Reports document what the system assessed at a point in time; they are not guaranteed court exhibits.
Procurement-safe language
Acceptable: "probabilistic detection supporting human review." Avoid vendor claims of "100% accuracy" or "guaranteed deepfake detection." Align RFP language with explainability and evidence exports actually shipped in the repository.
Known limitations
Adversarial content, heavy compression, and novel generators can produce ambiguous scores. Missing provenance is not a score — document absence explicitly. Retest after model updates; store analyzer version in case records.
Internal Linking
FAQ
Is Trust Score a single 0–100 number for all products?
Public UX may summarize signals differently by product and modality. Treat each analysis record as modality-specific evidence.
Can we auto-block content below a threshold?
Only if legal and policy review approves automated gates — and humans retain override paths. Default enterprise posture is review-assisted, not fully automated adjudication.
How do scores relate to C2PA manifests?
Provenance validation and forensic scores are complementary. Contradictions between manifest claims and forensic signals should trigger escalation.
