Controls that support an accountable decision.

PDF text (up to 50 pages), DOCX main-body text and UTF-8 TXT, Markdown and CSV are accepted up to 4 MB and 50,000 extracted characters. Extraction runs in a bounded worker; users inspect text before saving. Original files are not retained. Product briefs and conversations are workspace-scoped and retained with history. Exact quote checks reject references outside selected evidence, and stale sources block proposal saving. OCR, image/video interpretation and automatic website ingestion are not supported. AI replies remain advisory.

Provisioned accounts, expiring sessions, password hashing, same-origin write checks and database-backed rate limits. Owner, author, reviewer and viewer roles are enforced by the API. Signing out revokes the session.

Application queries and composite database relationships enforce workspace scope. PostgreSQL row-level security also applies to tenant content for the restricted runtime role. A request sees the signed-in workspace, the public publication reader sees one approved page, and the worker can process queued jobs. Table owners and roles that bypass row security are not used by the application. Invitations and account tokens are looked up before a workspace is chosen. Privileged database administrators remain trusted.

Authors cannot approve their own claim or version. All required attestations must pass. When retained source text continues after a linked excerpt, approval is refused unless the reviewer attests that they read that later text. The review desk does not offer that confirmation until each such source has been opened. The attestation is stored with the decision and shown in its history. It records the confirmation; it does not prove the reading. A hold does not require it. High-risk claims require two distinct independent reviewers. Concurrent edits reject stale versions rather than overwriting a newer decision.

Database triggers reject updates and deletion of protected version, decision, audit and measurement history during ordinary application operations. This is not externally notarized storage or a retention lock, and privileged database administrators remain trusted.

Source changes, revocation and expiry invalidate dependent approvals. The hosted page is withdrawn before that action returns. Connected CMS withdrawal is attempted in the same request, up to three jobs, and then retried up to eight times if the destination does not accept it. The workspace lists a withdrawal that is still open, and the owner can attempt that withdrawal again in the same request. A publication retry waits for the scheduler. A connector response does not confirm the live page. A CMS outage can leave the external copy in place. The daily scheduler is the retry backstop.

The deployment uses OpenRouter for advisory jobs; the implementation also supports direct OpenAI routing. Requests send the claim, linked evidence text and configured brief to the selected provider. Results are schema-validated and stored with run metadata. Four live synthetic job checks passed; this does not establish comparative quality or contractual data-retention guarantees. Provider terms and data suitability require assessment before sensitive information is submitted.

A workspace owner can save an HTTPS OpenID Connect issuer, client ID and encrypted client secret. Sign-in uses state, nonce, PKCE and token signatures, and only an explicitly mapped issuer and subject can enter. An owner can check the saved issuer's public discovery document and can record a member's identity provider subject. The discovery check does not send the client ID or secret. Neither that check nor a recorded subject proves the customer's identity provider has accepted Superlimen. Saving the connection does not prove acceptance either. Requiring company SSO is refused until the owner's own subject is mapped. SCIM tokens are scoped and expire; deactivation revokes sessions. Group membership does not grant application roles automatically. Automated tests use a mocked identity provider. SAML is not implemented.

CMS credentials are encrypted with a deployment key. Delivery jobs use leases, bounded retries and visible failure states. Connectors have transport-contract tests; a real destination, content model and publish/withdrawal acceptance test are required before customer use.

Outcome entry and authenticated ingestion are available. Visibility collection separately requires a direct OpenAI credential, currently unconfigured in production. It targets the Responses API with web search, not consumer ChatGPT or Google AI Mode. Cohort comparisons and pipeline influence do not establish causal revenue impact. Dedicated CRM and analytics OAuth connectors are not implemented.