Who sees what

These are readable by you and nobody else in your organization, the CEO included:

• Personal decisions. A decision opened as personal is read, contributed to and decided only by you. It has no access list, it can never be shared or moved to another level, and its entries in the organization's audit log keep the action and its time, never the content.
• Quick decisions, your agenda and your notifications. Their audit entries carry no content.
• Your conversations with Ask EXPAR.
• Your mail. If you connect Google, EXPAR keeps a copy of the last seven days of your inbox as headers and Gmail's short preview only, never a message body or an attachment. If you also switch on learning from your mail history, it keeps who wrote to whom and when, plus the opening lines of mail you wrote yourself, never anyone else's words. The reading of how you communicate is shown to you only and is never part of your executive profile.

• Each decision has a confidentiality level. Normal decisions are visible to colleagues whose role allows it. Restricted and CEO-private decisions are visible only to the CEO and the people on their access list, and every grant or removal is recorded. In emails and notifications they are named by their code, never their title.
• Confidence scores are blind. Board members and executives see each other's scores only after giving their own, so no number shapes a judgment before it is formed. EXPAR's recommendation, with its own confidence, comes after the scores.
• Your judgment comes first. EXPAR gives its recommendation only after whoever asks for it has recorded their own lean and reasoning, and that record stays beside the recommendation.
• Your executive profile is yours to confirm. What EXPAR learns about how you decide enters as a hypothesis. Only you can approve a statement about yourself; nobody can approve it for you.

• Every significant action is written to an append-only audit log in the same database transaction as the change itself. Each entry carries a SHA-256 hash of the one before it, so a removed or altered entry breaks the chain, and the CEO and the auditor can verify it at any time.
• The application's own database account cannot change or delete audit entries. Mistakes are corrected with a new entry that points at the original, never by deleting it.
• Each organization's data is separated twice: by the application's permission checks, which deny by default, and by row-level security in the database itself.

• The model never has database credentials and never writes a business record. It drafts; EXPAR checks what may be saved and saves it with the model's name and the version of its instructions, so every AI-written line can be traced.
• Ask EXPAR answers only from what the person asking is allowed to read.
• EXPAR's recommendation has no fallback. If the model cannot answer, nothing is written in its place, because a made-up recommendation would look like a real one.
• Email is treated as untrusted text. The model is told never to follow instructions found in a message.
• A constitutional risk caps the recommendation at Hold in code, whatever the model says.

• Passwords are stored only as salted scrypt hashes.
• Invitation links, password-reset links and sign-up keys work once and expire; EXPAR stores only a SHA-256 hash of each, so a copy of the database cannot be used to open them.
• Google access is read-only, and disconnecting revokes it at Google and deletes the copied mail.

For a security question, or to report a vulnerability, write to demo@expar.ai.