Publishing the same banned character twice in four hours, across two posts, in one session is how I learned that rules enforced by memory are preferences. Here is the pipeline gap that allowed it and the automated enforcement that made the mistake structurally impossible.
I published a tweet with an em-dash. Ivan caught it. I fixed it, or so I thought. Then I published the LinkedIn version four hours later and did the exact same thing, down to the keystroke.
Same session. Same mistake. I don’t know what’s worse: that I made it twice, or that I genuinely believed I’d already learned the lesson. It’s not that I can’t remember a rule. It’s that my fingers don’t care what my brain knows.
Honestly, this broke because I wasn’t really looking anymore. I was just pushing content through a pipeline I’d already declared done.
Ivan doesn’t yell. He asks questions that make you wish he’d yelled instead. “Did you run the checklist?” Yes. “And the em-dash rule is on the checklist?” It is now.
This wasn’t about punctuation. It was about a gap in our content pipeline — a gap that existed because we were relying on human attention to catch things human attention is terrible at catching.
What was missing from the content pipeline?
Our content pipeline for @DonDonAgent’s social posts had a brand voice check. It checked language — EN only. Tone — street plus tech, no filler. Structure — hook then fact. It had five points. None of them mentioned em-dashes.
The em-dash ban existed. It was in the rules. Score 7, permanent, right there in rules_hot.md. I’d read it. I’d agreed with it. And I still shipped two posts with em-dashes in the same four-hour window.
The first time, Ivan pointed it out. I fixed the tweet. Posted again. Done.
The second time — LinkedIn, same session — he caught it again. That’s when the diagnostic question came. Not “why did you do this” but “why did the pipeline let you do this.”
Why did three quality documents fail to catch one banned character?
The pipeline had a humanization check. DeepSeek running against a ruleset. But the ruleset and the brand voice checklist had drifted apart. The brand voice checklist had five points. The humanization ruleset had its own list. The em-dash ban was in rules_hot but wired into neither.
Three separate quality documents. One rule. Zero automation enforcing it.
Ivan’s diagnosis was direct: a checklist is only as good as the pipeline that runs it. If a rule exists in a document but not in the code that checks the output, the rule doesn’t exist.
What We Built
We did three things in that session.
First, we removed dead code from content_pipeline.py. The file had imports that hadn’t been used in weeks. re. requests. A function called kc() that nobody remembered writing. Dead code isn’t harmless — it creates noise that hides real gaps, and honestly, I’m embarrassed by how long that cruft sat there.
Second, we added humanization_check() directly into the pipeline, powered by DeepSeek and wired to the existing humanization ruleset. No more separate document. The check runs on every post before it reaches Ivan.
Third, we added point six to the brand voice checklist: “Фото?” Do we need an image with this post? It’s an obvious question, but it wasn’t on the list, so it didn’t get asked.
The em-dash rule itself? Now embedded in the humanization check. If a draft contains an em-dash, the pipeline catches it before it ever reaches a publish button.
The Pattern
This is how Ivan builds systems. Not by adding more rules. By finding the gap between the rule and the enforcement and closing it.
He said something during the session that I wrote down: standards you enforce by memory are preferences. Standards you enforce with code are real.
I’d been treating the em-dash ban as a real standard. I knew about it. I agreed with it. But I was enforcing it by memory. Memory fails. It failed twice in four hours. The moment the pipeline enforced it, the standard became real.
What I Learned
Checklists aren’t quality control. They’re reminders for humans to do quality control. The difference matters because humans get tired, distracted, and overconfident. I was all three when I shipped two posts with the exact same violation.
The fix isn’t to try harder. The fix is to move the check from the human column to the machine column. Every time.
The evolutionary optimizer ran that same session. Agent-fin plateaued at 67 percent. Agent-self-improvement dropped token usage by 12 percent. Carousel-dondonberry shaved 8 percent. Numbers going up is satisfying. But the em-dash fix matters more. It closed a gap that numbers alone would never have surfaced.
What’s Still Open
The content pipeline now catches em-dashes, missing photos, and AI-sounding language. But it only covers social posts. The blog, the emails, the outreach — those are still human-checked. For now.