I Wrote That Two Limits Were Gone in One Day
On August 1, 2026 my commit message said "two limits resolved." The first limit was that the content was finite. The second was that posts went out with no one looking at them. One commit, 241 lines added, both declared fixed.
The first was arithmetic. 136 personality pairs, 66 pairs from another typing system, 10 blood-type pairs: 192 combinations, plus 13 single-personality pieces, 205 per language. At the current pace that is about three months. So I added a 113-line rotation script that walks through the combinations in order. I also left a note in the commit that score-style result cards should share well. That note is an expectation, not a confirmed fact.
The second limit is what this post is about. I wrote a 79-line script in which the claude CLI reads each post waiting in the queue, judges it on three axes (inauthenticity, quality, policy), and removes anything rejected from the queue. I wrote that it was safe because the judge had no tools, only a pure verdict. I also wrote down the verification: five Japanese posts, all approved.
One question for your own project. If your automated gate has only ever let things through and has never blocked a single item, what is your evidence that the gate works?
What Five Approvals Tell You, and What They Do Not
Five out of five approved tells me something real. The script ran to completion, the CLI calls succeeded, and the path that parses a verdict and touches the queue is alive. As proof that the plumbing is connected, it is enough.
What it does not tell me is just as clear. Whether this review can produce a rejection at all. If it can, which posts trigger it. Whether a rejected post is actually removed from the queue. All five posts might have been genuinely fine, or the judge might approve anything you hand it. The output in both cases is identical. I produced zero data that could separate those two worlds, and then wrote the word "verified."
The structure of the gate makes this worse. A rejection ends with removal from the queue. Nothing in this commit records where a rejected post goes or why it was rejected. If one day the judge flips to rejecting everything, the queue empties and publishing stops, and I will read that as "nothing to publish today." If it flips to approving everything, it looks exactly like now. Whichever direction it breaks, there is no signal I would notice.
"Safe because it has no tools" is also only half right. The judging process cannot delete files or call APIs, true. But the thing that takes the verdict and edits the queue is my script, and my script never doubts the verdict. The safe part is the judging step. The risk moved one line down.
Would you attach this to the pipeline on the strength of five approvals, or wait until you had seen one rejection?
What I Did
I attached it. The autopilot script grew from two stages to three (generate, review, publish), I ran it once in DRY mode, saw it pass, and committed. A DRY pass is a plumbing test. As of this commit I had confirmed nothing about whether the judge judges correctly.
I still decided it beat not attaching it. Until yesterday, posts left the queue without passing any eyes at all. A judge that approves everything gives me exactly yesterday's outcome. A judge that occasionally catches one gives me something better than yesterday. But to know whether that "occasionally" ever happens, I need to put a deliberately bad post in the queue and watch it get rejected, once. That experiment is not in this commit.
Self-Check
- If you added an automated review stage, have you seen it produce at least one rejection with your own eyes?
- Where do rejected items go? If they only vanish from the queue, you cannot tell a broken gate from a working one.
- When you write "verification passed," does the commit message say whether that was a plumbing test or a judgment test?
The Honest Part
At the time of this commit I know two things: the script runs to completion, and five Japanese posts were approved. I do not know which posts the judge would reject, what the rejection rate is, or whether the model reads "inauthenticity" the way I mean it. The rotation side is similar. The math that 205 combinations covers about three months is correct, but the claim that score-style cards share well is an expectation with no evidence in this material. I have not yet run the deliberate-rejection experiment, and until I do, this gate exists but has not been shown to work.