Research privacy notice · Josef Chen
What FlavourBench retains, and what it does not publish.
FlavourBench separates operational collection from research release. Public battles remain disabled until the privacy, task-validity, abuse-control, budget, and judgment gates pass.
Browser and network signals
The service stores a signed random browser identifier in an HTTP-only, same-site cookie scoped to the FlavourBench API. A season-specific HMAC combines that identifier with a bounded network signal for duplicate and anomaly detection. The benchmark API receives pseudonyms, not the raw IP address.
Prompt consent and retention
Research-release consent is explicit and optional. Without consent, raw prompt and output content is retained operationally for 30 days and then redacted; content hashes, votes, costs, reliability records, and derived metrics may remain. Consented records may enter a sanitized research export only after PII, source-rights, provider-output, and model-identity leakage review.
The present public release contains no raw benchmark prompts, model answers, private tool traces, credentials, reviewer identities, or quality scores. Content-addressed workload files expose only the bounded metadata described on the methodology page.
Website measurement
The host website may load Google Analytics and PostHog for aggregate traffic and product reliability measurement. These services are separate from FlavourBench research consent; submitting a research-consent checkbox does not opt a visitor into website analytics. Browser privacy controls may be used to restrict those services.
Research contact and infrastructure notice
FlavourBench research is led by Josef Chen. The broader host, processor, security, analytics, and privacy-rights details remain available in the site-wide infrastructure privacy notice. This research notice does not replace those legal disclosures.