Every AI marketing tool we evaluated before building Pappus shipped with the same default: auto-publish unless you stop it. Sit down on Monday, find that the agent posted six tweets, two LinkedIn articles, and a newsletter you don't remember approving. The pitch was always the same — "save time, intervene only when something looks wrong."
We picked the opposite default. Nothing goes live until a human approves it. Every draft is recorded with an HMAC-signed audit trail, every AI-generated image carries C2PA cryptographic content credentials, and every claim in a draft is checked against your tenant-owned Claims-DB before a human ever sees it.
The case for the gate
Three concrete failure modes drove the decision. They are the kind of thing you hear about from a customer at 2am, after it has already happened on a public timeline.
- Wrong claim, public timeline. A drafting agent can confidently write "the leading provider of X" without knowing whether you have grounds to claim it. With the gate, Pappus's Substantiation Mode catches the unverified claim and either auto-hedges it or annotates
[citation needed]before the reviewer sees it. Without the gate, the post is already live by the time the legal team notices. - Off-brand image. Image generators get the subject right most of the time and the brand colors wrong some of the time. The cost of one wrong-colored hero image being tweeted by a brand is hours of cleanup. The cost of a one-second review is a one-second review.
- Tone misread. The day Slack went down, the on-brand cheerful "Mondays are for shipping!" copy your agent drafted at 8am is not the right thing to publish at 11am. A human looking at the calendar catches the misread.
What we learned from the "give me autopilot" requests
We have had this conversation many times. The request usually sounds like: "I review every draft anyway, I want one fewer click."
We dug into the actual data with the few teams who beta-tested an opt-in auto-publish toggle. Two things came back:
- The teams who turned it on caught fewer issues, but the issues they did catch were worse. The base rate of needing to delete a published post went from near-zero to roughly one per month per channel.
- The "one fewer click" did not save much time. Most teams already batched their approvals into one weekly sweep on Sunday or Monday morning. The marginal click is dwarfed by the batching.
We retired the toggle. The approval queue is the contract, and we invested instead in making the queue faster: keyboard shortcuts, cascade approvals (approve a LinkedIn → blog → newsletter sequence with one decision), and one-click "approve everything in this pillar that passed the Reviewer's checks."
What this actually buys you
Trust is the obvious one. The deeper one is leverage: the gate is the place where every other safety promise becomes provable. The Claims-DB check, the C2PA stamp, the brand-voice memory — none of them mean anything without a moment where a human says "yes, this is correct, publish it." The HMAC-signed approval is the load-bearing artifact. The publisher refuses to publish without one.
Read more on the trust posture on the trust section of the home page, or grab the long-form trust posture if you want the raw substantiation index.