# Data, judgment and access boundaries

Understand CandiKeep’s current permissions, AI review behavior, product limitations and public-site traffic measurement.

Web version: https://candikeep.com/docs/trust/

Keep the operator informed about what information enters the product and what each action actually does.

## Operator authorization comes first

Public documentation is readable without sign-in. Candidate records are behind an authenticated personal account. Reading these guides does not authorize an assistant to upload resumes, alter candidate information, export records or send them elsewhere.

Use candidate information only for an operator-authorized recruiting purpose. Existing access to a connected mailbox or drive is not, by itself, approval to copy all of its contents into CandiKeep. Review the intended input and destination with the operator.

CandiKeep does not offer shared team roles or a public agent authorization flow today. Do not invent API credentials, MCP endpoints or supported background-agent permissions.

## AI assistance needs review

AI-assisted processing extracts candidate information from resumes and notes. Source-evidence checks and review proposals help make changes inspectable, but do not prove that every interpretation is correct.

Non-conflicting additions can update the profile automatically. Changes to accepted information and contradictions require review. Search results are decision support: important qualifications, interest, compensation and availability should be confirmed.

The product does not automatically reject candidates, send outreach or make a hiring decision. The operator remains responsible for candidate assessment and communication.

## Know the current handling limits

Uploaded resumes, saved notes, accepted profiles and review history are stored so they can support future recruiting work. Selected-profile exports are available, but should not be described as a complete account archive or a self-service deletion mechanism.

CandiKeep uses Clerk for account sign-in and authentication. AI processing uses OpenAI: resume or note text and current candidate-profile context are sent for extraction and reconciliation, accepted profile text is sent to generate search embeddings, and optional AI refinement sends the search request for interpretation and embedding. Ordinary keyword search does not make an AI request, although earlier imports and indexing use AI processing.

These guides do not establish a retention period, model-provider training policy, service-level agreement, data-residency commitment or security certification. Confirm any required contractual or data-handling terms with the operator of the service before uploading sensitive information.

## What public-site traffic counts mean

The public site counts requests to public pages without tracking cookies. Server request measurement includes non-JavaScript requests to documentation and Markdown pages, which browser-only analytics may miss. Optional browser interaction collection is a separate feature and is disabled in this publication.

Automated traffic is classified from request signals such as user-agent strings. Daily request counts are not unique-agent counts. A label is an estimate: clients can omit or imitate identifiers, and a request does not prove that an assistant read, recommended or used CandiKeep.

Search indexing crawlers, user-triggered fetches and ordinary browser requests have different meanings. Treat measured requests as evidence of retrieval, and assess actual product adoption separately.

Open the signed-in workbench: https://app.candikeep.com/

Product access requires operator authorization; a documentation link grants no account or candidate-data access.

Documentation index: https://candikeep.com/llms.txt