Claremont Security seal Claremont Security Contact
// Restricted_Vault / Workbench

The Workbench

Client work does not reach an assistant raw. It moves through the Workbench: a briefing that disciplines the model, a clean room that converts every identifier into a reversible token, and a scan that proves the text is clean before it leaves. The entire pipeline executes on this machine. Nothing you type is transmitted or stored.

// Protocol_01 01Select a preset or set the six parameters 02Review the assembled directive 03Copy it, or attach it so every vault prompt carries it
// Presets
Parameters0 of 6 set
06. Constraints
Directive
Saved BriefingsStored to your vault account
// Protocol_01 01Pick a template, or define the advisor yourself 02Give it a name and at least one decision principle 03Copy the preface, or attach it so every vault prompt carries it

An Advisor Card is a reusable point of view. It turns a role your firm relies on, a reviewer, a skeptic, a client translator, into a standing instruction the assistant applies to whatever you give it next. Define the principles once; attach the card and every vault prompt inherits them.

// Templates
DefinitionName and one principle required
Decision Principles, applied in order
Advisor Preface
Saved AdvisorsStored to your vault account
// Protocol_02 01Paste the working text 02Sanitize to swap identifiers for tokens 03Give the assistant the sanitized block 04Paste its reply into Restore to re-insert real values
01Source Text0 chars
Firm DictionaryAlways redacted, saved on this device only
// Protocol_03 01Paste what you intend to send 02Read the verdict; it updates as you type 03Anything flagged goes to the Clean Room 04Download the record for your file
Text Under Review0 chars • live
Awaiting Text Twenty-two detector patterns with checksum validation, covering government IDs, financial and health identifiers, personal data, and credentials. 0 critical 0 elevated 0 low
Annotated TextClick a finding to jump to it
Findings
The scan record documents that text was checked before it reached an AI assistant: timestamp, verdict, and finding counts. It never contains the flagged values themselves.