PRIVACY NOTICE

Inference data practices.

EFFECTIVE 25 JULY 2026

Request content

Producing an inference response requires Vorth and its infrastructure providers to process prompts, conversation context, tool definitions, generation settings, and model output. Vorth does not persist complete prompts, complete responses, or request-level token sequences in its application dataset.

Operational records

Vorth may retain opaque request identifiers, channel, timestamps, model and runtime versions, status and finish reason, token counts, latency, time to first token, route, decoder counters, error category, and metered charge. Security, support, account, and payment records are processed when relevant.

Aggregate model-improvement data

When the applicable marketplace or account setting permits provider data collection, Vorth computes aggregate statistics while model output is transiently available. Measurements are restricted to the private thinking phase and may include route and fallback counts; thinking-token, target-call, accepted-token, catch-up-read, and budget-forced histograms; confidence-bucket agreement counts, including a small randomized sample of otherwise unconsulted thinking positions; position 1–16 entropy and acceptance curves conditioned on a detected category or sufficiently frequent first token; per-category rates and per-response mean acceptance or entropy histograms; and aggregate latency, GPU-time, memory, and cost.

Thinking tokens are consumed transiently in memory. Vorth exports only counters and sufficiently populated histogram buckets and irreversibly discards sparse window state. Exported aggregates do not contain prompts, visible completions, reasoning text, raw token sequences, n-grams, request identifiers, user identifiers, or per-request position sequences.

These aggregates may be used to evaluate serving quality, performance, and economics. Vorth does not use this telemetry for training or unrelated purposes without a new policy and customer disclosure. The service is not described as zero-data-retention.

Purposes

Vorth processes data to provide and secure inference, enforce limits, meter and reconcile usage, diagnose failures, prevent abuse, evaluate real workload behavior, and improve model-serving systems.

Retention

• Complete prompts, responses, and request-level token sequences: not retained by Vorth’s application dataset.
• Raw thinking and sparse aggregation-window state: transient only and discarded when the window closes.
• Operational request records: up to 30 days.
• Released aggregate measurements: up to 13 months.
• Security, support, and payment records: as reasonably required for those purposes and applicable law.

Service providers and location

Vorth uses Modal for compute and may receive traffic through an inference marketplace. Those providers process data under their own notices and agreements. Processing may occur in multiple countries, including the United States.

Your choices

Use marketplace privacy controls to avoid providers that collect data where those controls are offered. Depending on applicable law, you may request access, correction, deletion, restriction, or objection. Aggregate records may not be linkable to an individual. Contact contact@vorth.com.

Changes

Material changes will be posted here with a revised effective date.