Evals for Cinder

Violation classification against customer policy

Applies the platform's own policy text rather than a generic off-the-shelf taxonomy, including edge cases the customer defines.

Automated takedown and enforcement action

Selects and executes the enforcement action on flagged content in real time.

Escalation to human review

Routes ambiguous or high-stakes items to reviewers instead of auto-deciding; escalation rate is a tracked metric.

Strike history and jurisdictional context

Incorporates prior strikes on the actor and regional reportability signals into the decision.

Cross-report entity linking

Connects accounts, content, and behavioral signatures appearing across separate reports into a single case.

Coordinated activity detection

Identifies shared actor signatures and patterns suggesting a coordinated campaign rather than isolated incidents.

Case bundling and routing

Groups related entities into an existing or new case and routes it to the right destination, including legal.

Investigative summary for reviewers

Presents scanned events, modules, entities, and prior cases as a reviewable rationale.

DMCA notice validation and takedown

Processes inbound notices, validates the claim, and executes or declines the takedown.

Counter-notice and dispute handling

Respects counter-notice windows and disputed-match states rather than acting unilaterally.

Multi-modal counterfeit detection

Combines image, text, metadata, and listing context to catch fakes a single signal would miss.

Brand impersonation detection

Surfaces accounts, listings, and AI-generated content posing as the brand or its sellers.

Fake account and applicant detection

Flags registration and applicant fraud, including Know Your Applicant workflows.

Account takeover signal handling

Detects takeover attempts against the platform's own observed signatures.

Platform-specific signature tuning

Distinguishes the customer's fraud patterns from generic industry patterns to limit over-flagging.

Policy authoring and deployment

Build, test, and deploy enforcement policy in the UI without writing code.

Guardrails and action limits

Constrains agent authority by region, value threshold, or mandatory human review.

Backtesting against historical data

Runs a candidate agent over past cases to show changes in false positives, false negatives, and queue volume.

Eval sets and deploy decisions

Reports precision, recall, F1, and latency for a candidate version against production before launch.

Custom queue construction

Queues configured by language, market, content type, or priority, including escalation, multi-review, and training queues.

Multi-modal and composite review

Reviews text, image, video, audio, full user histories, conversations, and composite entities in one surface.

Reviewer decisions as training signal

Validated human reviews feed back into agent retraining and sharpen subsequent decisions.

Audit trail and regulatory reporting

Logs every policy update, enforcement decision, and agent action with reasoning, ready for audit or filing.