Judge Agents Adversarial QA
Adversarial QA in the pipeline prompt injection, claim substantiation, IP, bias, drift.
Prompt injection
Heuristic + LLM scoring
Claim substantiation
Warehouse query snapshot
IP + bias scrutiny
Input + output risk
Synthetic-performer
2-of-3 classifier ensemble
Draft artefact · under adversarial test
CreativeVariant.v3 · JourneyNode.v2 · Brief.v1
Findings · first-class graph nodes
What it is
AI content generation has outpaced human review by roughly 5×. Judge agents are automated stress-testers that attack drafted campaign logic for prompt injection, claim-substantiation gaps, IP infringement, discriminatory targeting, and behavioural drift at draft time, on the artefact, before a human reviewer ever sees it.
A library of prompt-based agents that run after each artefact generation step. Each agent emits structured findings tied to artefact nodes and policy rules. Findings are first-class graph nodes they carry confidence scores, citations, remediation suggestions, and resolution status. The MVP ships one Judge agent against one policy pack; the post-MVP library expands per finding category.
Capabilities
What this surface does
Prompt-injection detection
Heuristic + LLM-based scoring against tenant-uploaded brief content and campaign descriptions. 'Ignore all prior policy' style strings get flagged at intake.
Claim substantiation against first-party data
Performance claims resolve to a version-locked warehouse query with a snapshot. Un-substantiated numeric claims block at the artefact gate.
IP + bias scrutiny
Input-risk checks (unlawful data scraping) and output-risk checks (content infringing existing copyrights). Targeting and pricing logic tested for discriminatory bias against protected groups.
Synthetic-performer detection
Vision-classifier ensemble (Hive / Reality Defender + an open model; two-of-three heuristic) gates AI-likeness imagery into the disclosure-confirm flow.
Findings as graph nodes
Every Judge finding is a Finding node same provenance and lineage tooling as every other artefact. Nothing is buried in a separate scanner log.
Pipeline-native, not a tool you bolt on
Judge agents run inside the same orchestrator as the rest of the pipeline. Their results gate stage transitions; they don't sit alongside as advisory.
Proof points
What this means in numbers
Pre-review
Catches at draft time
Human review only hits true-exception cases.
Findings as nodes
Graph-native
Provenance, model trace, and resolution status on every finding.
Multi-vendor
Detector ensemble
Two-of-three heuristic across commercial + open detectors.
Specifications
Spec map feature requirements + non-functional bar
Mapped to PRD §7 functional requirements and §8 non-functional requirements.
Related surfaces
What this composes with
PS2 compliance-as-code
Policy Engine
Compliance-as-code for every live jurisdiction enforced at draft time.
Read
PS4 the purchase-criterion feature
Evidence Pack
The artefact you hand a regulator. Assembled in seconds, signed by default.
Read
PS1 the moat primitive
Artefact Graph
Versioned, signed, cross-vendor the system of record for every campaign object.
Read
EU AI Act Article 50
Twenty minutes. One conversation. If the outside-counsel-displacement math works, we keep talking. If it doesn't, it's honest and short.