Trust

Security & trust

This page describes how the CapiFact pilot is actually built, written by the team that built it. It is a description of current practice, not a third-party audit or a certification.

Last updated · Pilot release, description of current practice

Draft for legal review. This page describes accurately how CapiFact works today, but it is not legal advice. Capifact's own legal adviser must approve the final wording before real student or employee data is used on the service.

01

Account separation

Every record, piece of evidence, claim, endorsement and link carries the owner's account identifier, and access is enforced by row-level security rules in the database. A request from another account returns nothing at all, separation does not rely on the interface hiding anything.

02

Roles and privileged access

  • Two roles exist: member and CapiFact administrator. Roles live in their own table, never on the profile, so they cannot be self-assigned.
  • Role checks run server-side through a security-definer database function, not a client flag.
  • Administrator access requires a second sign-in factor (an authenticator app) for every session. There is no exemption for demonstration accounts, and no public button signs anyone in as an administrator.
  • Oversight is deliberately narrow: counts, statuses, endorsement and share activity, and audit events. It does not expose claim wording, evidence text, endorser identities or career outputs.

03

Files and uploads

  • Only PDF and Word (.docx) files are accepted, up to 5 MB. Anything else is refused at upload.
  • Each upload is screened for a mismatched file type and for active content such as embedded scripts or launch actions. A file that fails is quarantined: it is never stored as usable evidence, never sent to the AI model and never included in a shared record.
  • Source files are never downloadable or viewable through an employer or endorser link.
  • Profile photos are held in a private storage area scoped to each account's own folder and served through short-lived signed links.

04

Account protection

  • Email addresses must be confirmed before first sign-in.
  • Passwords must meet a strength policy, with live feedback as you type.
  • Password reset uses a single-use, time-limited link.
  • Any member can enrol an authenticator app for two-step verification in Settings.
  • You can permanently delete your own account, photo and records from Settings.

06

AI use and traceability

Claim suggestions and career drafts are produced by an open-weight model (openai/gpt-oss-120b) via Groq. Only extracted text and approved claim wording are sent; never files, credentials or account identifiers. Every AI-suggested claim stores the verbatim sentence it came from and stays a candidate until a person approves it. Support states are decided by inspectable rules in the application, not by the model, so the same inputs give the same answer.

07

Audit trail

Sign-offs, claim edits, endorsement invitations and responses, and link creation and revocation are recorded with a timestamp. Audit entries cannot be edited or deleted through the application, by members or by administrators.

08

Known limits at pilot stage

  • Endorser identity is not independently verified; an endorsement is that person's own word.
  • Uploads are screened by type and content pattern, not by a commercial antivirus engine.
  • An employer link is unlisted, so anyone it is forwarded to can open it until it is revoked.
  • Destructive operations on a member's real data are not available to administrators in the application, by design.

09

Reporting a vulnerability

If you think you have found a security issue, email security@capifact.com with steps to reproduce. Please test only against the seeded demonstration accounts, never against someone else's record. We will acknowledge your report and keep you updated.