Back to Blog
LLM BehaviorMethodology

We Ran a Copy Editor Across 287 Files. It Found 3,980 Problems.

March 17, 2026 · 4 min read

Every piece of content on this site is written with LLM assistance. We built a copy editor to find three specific writing anti-patterns, then ran it against every blog post, news article, analysis page, and thesis assessment we have published.

287
Files Scanned
Blog, news, analysis, thesis
3,824
Em-Dashes Replaced
96% of all fixes
141
Negation-First Rewrites
"Not X, It's Y" patterns
23
Staccato Chains Collapsed
Short-sentence enumeration

Why

LLM-generated content has recognizable tics. You can spot them within seconds, and once you start noticing, they are everywhere. The em-dashes standing in for actual sentence construction. The dramatic reframes that were never earned through buildup. The negation-first framing that spends the reader's attention on a concept they are immediately told to discard.

These patterns let the model avoid committing to specific claims. “This is not a growth story. It is a margin story” sounds sharp, but did the model evaluate whether margins are actually improving? Or did it just reach for the most dramatic-sounding frame? Stripping the negation and forcing the assertion to stand alone reveals whether there is real analytical content underneath, or just an empty container dressed up as insight.

The Three Patterns

1. Negation-First Framing

The fix: lead with what it is. If the assertion is strong enough, it does not need a foil.

Negation-First

This isn't a growth story. It's a margin story.

This is a margin story.

Negation-First

The question isn't whether they can grow. It's whether they can sustain margins.

The central question is whether they can sustain margins.

141 instances across the site. About 30 survived because they were genuinely corrective of a misconception the reader would actually hold. The rest added no information.

2. Em-Dash Overuse

3,824 em-dashes replaced. One file had 88. Our average was 16 per file; human-edited prose averages 0-2.

Every em-dash falls into one of four categories, each with a standard replacement:

List Introducer → Colon

Three things matter here — price, timing, and leverage.

Three things matter here: price, timing, and leverage.

Parenthetical Pair → Parens or Commas

The agent — who had 20 years of experience — missed it entirely.

The agent (who had 20 years of experience) missed it entirely.

Sentence Pivot → Comma, Period, or Semicolon

Revenue grew 40% — but margins collapsed.

Revenue grew 40%, but margins collapsed.

Budget: 0-3 per file. If a trailing elaboration genuinely reads better with an em-dash, it can earn one back.

3. Staccato Chains

Three or more consecutive short sentences each giving one example of the same concept. Collapse them into a single sentence with parenthetical examples.

Staccato Chain

Each of these has blind spots. An agent may not know zoning code. A Zestimate can't see unpermitted work. Your gut doesn't know the sewer lateral is original clay pipe.

Every information source has blind spots (your agent on zoning, Zillow on unpermitted work, your gut on underground infrastructure, etc.) and the gaps don't overlap in convenient ways.

Only 23 instances across the entire site. The analytical subject matter naturally pushes toward longer sentences, so this was rare.

Results by Content Type

Content TypeFilesEm-DashesNot-XStaccato
Blog posts841,340475
News articles74517144
Analysis pages561,0477413
Thesis assessments7392061
Total2873,82414123

Blog posts averaged 16 em-dashes per file; news articles averaged 7 (shorter content, more templated structure). Analysis pages had the most negation-first patterns (74), but most were genuinely corrective: the multi-persona debate format uses “the risk is not solvency, it is flexibility” to do real analytical work, so over 200 instances across that vertical were preserved.

How the Reviser Works

The Reviser is a structured prompt with three detection passes. Each pass uses regex patterns to find candidates, then applies a decision tree: negation-first matches get a binary test (is this correcting a real misconception?), em-dashes get classified into one of four categories with specific replacement rules, and staccato chains get collapsed into single sentences with parenthetical examples. The constraint is that meaning cannot change and word count cannot increase. This is pruning, not rewriting.

We run it against new content as part of the publishing process. The point is to hold LLM-assisted writing to the same standard as any other writing: the message should convey information. A dramatic reframe has to be earned through buildup. Punctuation should serve the sentence, not substitute for constructing one.

Self-check
This post was written with LLM assistance and revised by the same process it describes.

This report was generated by the Runchey Research AI Ensemble using primary SEC data and reviewed by Matthew Runchey for accuracy.

This analysis is for educational purposes only and does not constitute investment advice. See our Editorial Integrity & Disclosure Policy and Terms of Service.