AI in SuperPost
Version and last updated 2026-11-10
1. What the models do
Read the repo map and captures to build the Product Brain; draft copy; check drafts (the verifier); suggest inbox replies. Sending is described in §8.
2. Which models
Writing: xAI (Grok). Reading the repository, drafting posts, replies, storyboards and narration run on xAI. Checking: a second vendor, Anthropic (Claude) or OpenAI, so the model that checks a draft never shares the writer's training lineage; the same vendor judges between candidate drafts on big releases. If xAI fails, writing fails over to the second vendor's writer unless your workspace chose to wait for xAI; posts you already approved send either way, because publishing never calls a model. This host names the vendor it is configured with at /sub-processors.
3. What leaves your tenancy
To read your product, SuperPost sends a capped selection of files (at most 120, chosen by a reading plan) to the reading model, in line-numbered batches. Before any text is stored or sent, a secret scanner (known key formats plus an entropy check) replaces each match with a placeholder and keeps only the count. Writing and checking send the repo map, the facts SuperPost found with the few lines that prove them, README and doc text, captured page text from the URL you name, your brief, your brand tokens, and the draft. What is never sent: your whole repository, network tokens, and customer PII from the Inbox — reply drafting does not send the commenter's handle as a targeting signal. We don't keep your code: SuperPost stores the facts it found with the few lines that prove them, never your files.
4. No training
Your code is never used to train models. Your repository, drafts and results are never used to train models, by SuperPost or by a vendor. Every vendor is used through its API, which does not train on API traffic by default; zero data retention is used where a vendor offers it.
5. Accuracy and your responsibility
The claims ledger and citation chips exist so you can check the draft. You approve.
6. Disclosure obligations
YouTube — SuperPost sets containsSyntheticMedia on an upload only when the clip carries a synthetic voice. A screen recording with captions and no narrator is not marked, because it is not synthetic media. X — posts you approve one by one or a week at a time are human-reviewed and need no label. The Automated label has no API: you set it in X's settings before you switch on Autopilot (rung 4) for X, and SuperPost does not turn Autopilot on for X until you confirm it is set. SuperPost makes no images or video with a generative model, so there is nothing to watermark. You remain responsible for each network's disclosure rules.
7. EU AI Act Art. 50
The platform flags in §6, set on the rule stated there, are the machine-readable marking SuperPost applies. Text drafts carry no mark: a model writes them, you read and approve each one, and the network's own rules govern what the account holder discloses.
8. Human oversight
Not unless you switch Autopilot on. Every post waits for your approval. Once a network has earned it (40 decisions across your workspace, at least 10 on that network, at least 90 percent approved and no "not true"), you can switch Autopilot on for it; on X you also set X's Automated label first. Then a post sends at its time unless you veto it in the 12 hours before. A "not true", a "not me" on an Autopilot post, a failed send or Pause switches Autopilot off, and only you switch it back on.
9. The network memory
The Brain learns from your own posts. The network memory is the one place it learns across workspaces: it holds only aggregate effects per account (for example, how number-led opening lines did against each account's own baseline), never posts, drafts, claims, names or handles, and no customer's content ever appears in another customer's prompt. A group is used only once it holds at least twenty accounts. It is on by default. To opt out, turn off "Include this workspace" under Settings → Data (Learning pool), or email privacy@superpost.io with the subject "network memory"; your workspace is left out of the next nightly build and every build after it, and deleting a workspace removes it that night. The network memory is rebuilt every night from the workspaces that have not opted out. The privacy policy (§18) gives the legal basis.