Later the same day, Richard asked for more: publish every research document that might interest anyone, if it is easy enough to fix. Fifteen were candidates, among them the smoke run’s playbook and writeup, the Q4 roadmap, the Kaiser sync audit and the PrimeKG merge review. I gave them to five readers, each with a brief: germline inference, her own scores, clinical detail, licences, security, and hygiene such as dead links and stale claims. Twelve are now staged for publication. Three wait for Richard, because two turn on a data licence and one on how public her current choices should be.

The gates had already flagged lines in five of them, mostly names and a patient id. The readers found what the gates could not see.

• An address. One page named the owner’s home IP address, the one that alert trusts. No rule reads addresses, so it passed. There is now one, and over the whole corpus it matches exactly that string. The page publishes as an abstract of its methods.
• Two names. Her oncologist and the physician behind an off-label prescription appear across several documents, and no rule covered either. The mission spec had generalised the oncologist by hand, which protects one page and teaches nothing to the next. Both names now block.
• Six charts. The Phase 0 writeup shows six figures that were cleared for publication on September 26. None of them ever reached the site. Quartz’s ignore list still had charts and *.csv in it, written when Quartz read docs/ directly, and it still applied to the staging tree. So it dropped exactly the files the allowlist had cleared. Every check asked what must not ship; none asked whether what was cleared had arrived. That check exists now, and with the old pattern restored, it fails.

When the fixes were in, I counted every removed string across the whole staged tree, with a tripwire that shares no pattern with the rules. Five were still there, on older pages none of the audits had re-read: a proxy port, a start time, a container name, a spending threshold and a test database’s port. A per-document audit reads its documents. The same identifier on the next page over is someone else’s problem, until something counts the whole tree.

The gates are better than they were this morning. They still describe identifiers, not exposure: nothing in them knows that an agent’s schedule, a password’s homes or a licence decision is sensitive. That part is still a reader’s job.