Skip to main content

Legal & Data Handling · Effective November 17, 2026

Privacy Policy

This Privacy Policy explains how TrueSynthesis Inc d/b/a VerdictModels (“VerdictModels,” “we,” or “us”), headquartered in Austin, Texas, handles information across https://verdictmodels.com and https://api.verdictmodels.com.

1. Zero Prompt Storage & Zero Model Training

We never train, fine-tune, or distill models on your API prompts, tool definitions, or completions. Prompt payloads submitted to POST /v1/chat/completions or the same-origin browser playground (POST /api/playground) are processed transiently in memory solely to generate your requested completion and compute token-based billing settlement. Prompts and raw bearer API keys are never written to application logs or databases by verdictmodels.com or api.verdictmodels.com.

2. Streaming vs. Non-Streaming Completion Replay Retention

To prevent duplicate credit charges when clients retry requests with an Idempotency-Key, completion retention differs strictly between streaming and non-streaming modes:

  • Streaming requests (stream: true): Upstream Verdict Cloud Gateway persists usage and billing metadata only (request ID, model ID, token counts, microcredit debit) in RequestState.receipt_json—neither prompt text nor streamed assistant completion content is stored in upstream databases. When an Idempotency-Key is supplied and the SSE transcript is <= 64 KiB, api.verdictmodels.com holds that completion transcript transiently in process RAM for up to 15 minutes solely for warm idempotent replay, after which it is evicted.
  • Non-streaming requests (stream: false): api.verdictmodels.com never writes prompts or completions to its own logs or disk (caching responses <= 64 KiB in process RAM for up to 15 minutes when an Idempotency-Key is present). Upstream Verdict Cloud Gateway persists the completion response JSON (choices assistant message / tool_calls and usage, never the input prompt) in RequestState.receipt_json (and encrypted Cloudflare R2 receipt backup when enabled) solely to honor deterministic idempotent replays and billing auditability without re-billing. Customers requiring zero completion-output persistence should send stream: true without an Idempotency-Key.

3. Stateless GPU Inference & Infrastructure Subprocessors

TrueSynthesis Inc d/b/a VerdictModels uses the following subprocessors to deliver hosted inference, routing, and shared Verdict Cloud billing:

  • RunPod, Inc. (Stateless GPU Inference — Canada CA-MTL-1 / North America): Executes stateless llama-server GPU workers for GA vc/verdict/* SKUs (verdict-4-35b-toolterse-v5, verdict-4-30b-terse, verdict-4-4b-quill) with zero disk prompt logging.
  • Modal Labs, Inc. (Stateless GPU Inference — United States): Executes stateless container inference for private-beta / burst vc/verdict/* routes (vc/verdict/verdict-4-20b).
  • Fly.io, Inc. (API Gateway Hosting): Hosts the stateless api.verdictmodels.com proxy and upstream Verdict Cloud Gateway.
  • Vercel Inc. & Cloudflare, Inc. (Web Storefront, Edge TLS & DNS): Serves verdictmodels.com and encrypted receipt storage where configured.
  • Neon, Inc. & Stripe, Inc. (Shared Verdict Cloud Identity, Ledger & Payments): Stores account records, SHA-256 hashed API key metadata, credit balances, and payment transactions for the unified Verdict Cloud wallet.

4. Browser Storage & Telemetry

verdictmodels.com sets no advertising or third-party tracking cookies. If you manually pin light or dark mode against your operating system preference, only the string "light" or "dark" is stored in browser localStorage (removed automatically when reverting to system default). API keys entered into the interactive playground remain strictly in volatile browser tab memory and are cleared when the tab closes or reloads.

5. Contact

Privacy inquiries, data rights requests, and security reports may be sent to TrueSynthesis Inc d/b/a VerdictModels (Austin, TX) via privacy@verdictide.com.