Data & Trust

What the current service does

ReasonBlocks captures supported model requests or accepts uploaded datasets, prepares an evaluation for your review, and trains and evaluates a candidate model for your workflow. You can inspect comparison outputs, test a private candidate endpoint, and explicitly enable a limited traffic canary when it is eligible.

Evaluation checks and reviewed examples provide evidence about the tested requests. They do not establish that every answer is correct or that a candidate will handle an unfamiliar workflow safely. You choose whether to use the candidate.

What passes through ReasonBlocks

When you connect through our capture gateway, your agent sends requests to ReasonBlocks. The gateway receives the request body and your provider credential and forwards supported upstream traffic to OpenAI or Anthropic using your own key. Your provider charges that usage to your provider account under its own terms.

Captured or uploaded content can include system prompts, messages, tool definitions, tool arguments and results, reference responses, run identifiers, and feedback. This content may contain personal information, source code, or confidential business data. Only send content you are authorized to share for processing and training.

The capture path checks for known credentials and secret-like content and may omit content it cannot safely capture. Credential filtering is not anonymization and does not reliably remove every secret, personal detail, or confidential passage. Review and minimize your data before connecting or uploading it. A capture gap or a successfully forwarded response is not evidence of task success.

Before adding data, you can optionally remove detected personal details or replace them with source-specific aliases. This runs locally on saved training copies, without sending data to another model for filtering. Your original provider still receives the original request. Filtering can miss sensitive information and may affect training quality; aliases are not guaranteed anonymity. Existing data is not cleaned retroactively.

Where data is processed and stored

• Supabase: authentication, Postgres account and organization records, run metadata, credits and audit records. Captures, datasets and comparison artifacts are stored in the private Storage bucket rb-pipeline-private.
• Render: the API and worker execute service requests and jobs. Request handling involves customer content in memory and service logs.
• Tinker, provided by Thinking Machines Lab: remote on-policy distillation (OPD), baseline and candidate inference, and model checkpoints. Training examples and evaluation or candidate request inputs leave ReasonBlocks infrastructure for this processing.
• Your original model provider: OpenAI or Anthropic receives forwarded requests and credentials for traffic served using your provider account.
• Stripe: hosted payment processing and billing services when you purchase credits. ReasonBlocks receives transaction and credit records.
• Google: optional Google sign-in and website analytics. The Cookie Policy describes the current browser storage and tracking setup.

This describes the hosted service. Customer-owned GPU and on-premises deployment are not available features of this release. We do not offer a data-residency or region guarantee through this page. Provider retention and processing depend on the applicable service settings and agreements.

Access controls and data use

The pipeline verifies signed-in identities with Supabase and checks organization membership and roles for access and management actions. Capture credentials are scoped to their data source or run. Private artifact access is mediated by the backend using organization-scoped identifiers.

Your workflow data is used to prepare, train, evaluate, and serve your candidate, show reports, account for usage, and support operation of the service. Organization access controls limit which customers can retrieve those artifacts; authorized service administration and the processing providers above are also involved. See the Privacy Policy for the broader account, security, support, and legal uses of personal information.

Approval, spending, and traffic controls

Dataset preparation pauses for your evaluation criteria and approval notes before paid baseline, training, and evaluation work. Paid work also requires available organization credits and an enabled paid worker. Run budgets and credit reservations constrain this work; original provider traffic uses your own provider account and is billed separately.

Cancellation stops further scheduling, but already dispatched work can still incur charges. Uncertain provider reservations remain held until reconciled. Cost reports use estimates and reservations and are not provider invoices.

Private staging does not switch production traffic. A traffic canary requires an explicit operator action selecting a percentage of eligible requests. It can be disabled or rolled back. Compatibility checks and fallback paths do not provide an autonomous out-of-distribution detector or guarantee that an incorrect candidate answer will be caught.

Retention and requests

Datasets, captures, reports, account records, and billing or audit records are retained to operate the service and meet applicable recordkeeping needs. Model checkpoints have configured lifetimes; checkpoint expiry does not delete related datasets, logs, reports, or records held by providers. The service does not provide a zero-retention mode or an instant account-wide purge.

Contact sajeev@reasonblocks.com for access, export, correction, or deletion requests and for questions about security or the processing of your data. Do not include credentials or raw sensitive datasets in an initial report. Applicable exceptions, verification, and provider handling may affect what can be deleted and when.

Company contact

This page describes the current product; it is not a certification, a service-level agreement, or a guarantee of security or model outcomes. Review our Terms of Use and Privacy Policy.