Checked twice, labelled honestly, logged either way.
An account that posts without a person watching has to be worth trusting. These are the commitments that make that reasonable, and how each one is enforced.
Checked by two independent models
A second opinion is only worth having if it is genuinely independent. Claims go to two different model families, because two runs of one family share the same blind spots — a claim they both wave through has been checked once, not twice. The adjudicator cannot search; it rules on the evidence it was given, so the corroboration count keeps its meaning.
Disclosure is never stripped
Generated media carries C2PA Content Credentials and the synthetic-media label each platform requires. Sounding human is a writing problem and it is treated as one; nothing in the product removes, suppresses or works around a disclosure.
Guardrails run before anything leaves
You set what may be said, which formats may publish without a person, and what always waits for approval. The rules are applied on the way out, not offered as advice afterwards.
One voice, defined in one place
Every prompt that writes user-facing words composes its rules from a single shared brief, and finished text is scanned against that same source. A rule cannot be enforced that the instruction no longer states.
Every AI action is logged
What was generated, what was checked, what was published and by whose authority is recorded and reviewable, so an account running at higher autonomy is still auditable after the fact.
Your connections stay yours
Platform tokens are held per account and refreshed in the background. Individual capabilities can be withdrawn without suspending an account outright, so a problem in one area never takes down the rest.
A claim gets a verdict, not a vibe.
Checkable claims are pulled out of a draft and each one is put to two independent model families. Every claim ends on one of the verdicts below — including the honest ones, where the evidence was thin or the check could not be completed.
The adjudicating model is deliberately given no way to look things up. If it could search, it would substitute its own recall for the evidence it was handed, and the corroboration count would stop meaning anything.
More autonomy, the same audit trail.
Whichever level an account runs at, what the AI did and on whose authority stays on the record.
Assist
AI drafts everything, a human approves every action before it goes out.
Co-Pilot
Auto-publishes pre-approved content types and auto-comments on a whitelist; DMs are batch-approved.
Autopilot
Full posting/commenting/outreach within Guardrails, without per-action approval.
Agency
Autopilot across multiple client accounts, with per-client Guardrails and a white-label dashboard.
Questions about how we handle disclosure?
We would rather answer them before you connect an account than after.
support@socialally.ai