# CoherenceGate's 12 Axes: 14 Manuscripts, 3.74 Passes, p<0.001

Brooklyn Bishop · August 7, 2026

> CoherenceGate's 12 Axes: 14 Manuscripts, 3.74 Passes, p

| Takeaway | Detail |
| --- | --- |
| The 40% reduction is a pass-canceller, not a drafting accelerator. | Validation deletes 40% of revision passes before they start; the saving appears as hours, not faster prose. |
| The saving is measured in editing labor, not word count. | For a long manuscript, professional editing ranges from $2,000 to $5,000, and from $2,100; a 40% pass cut shrinks the labor bill. |
| The gate's deterministic layer eliminates unnecessary consistency checks before LLM judgment is needed. | In a sample report, numerical consistency showed no findings—so the 40% cut comes from not rechecking clean work. |
| The same validation economics apply across manuscript lengths. | A short novella costs $600 to $1,500, from $690; epic fantasy costs $2,400 to $6,000, from $2,520—and the 40% reduction applies to both. |

Forty percent of revision passes vanish before they begin. In the Stanford benchmark that inserted CoherenceGate validation between passes, the mean revision-pass count fell by 40%—a cut that looks like a speed gain but is actually a cancellation effect. The gate deletes work before it starts, which is why the time saving shows up as hours on a project calendar, not as a faster word-processing cursor.

The mechanism is simple and deterministic. Before each planned revision pass, the gate evaluates whether that pass would address a real inconsistency; if not, the pass is cancelled. This is not an accelerator for drafting prose, nor a stylistic shortcut. It is a pass-canceller. Writers and editors do not type quicker; they simply start fewer passes. The 40% reduction is the direct result of those cancellations.

The economic context makes the effect concrete. A long manuscript costs $2,000 to $5,000 for editing, with a common starting price of $2,100. When 40% of passes never begin, the labor bill shrinks accordingly—without any change to the per-word rate. That is why CoherenceGate's benefit is measured in hours saved, not in words per minute.

![CoherenceGate's 12 Axes](https://static.mm-ais.com/article-images-ai/coherencegate-s-12-axes-14-manuscripts-3-ai-109f28c1.jpg)

## CoherenceGate's 12 Axes

“Grammar checker” is the wrong frame. In the 2026 Stanford Self-Publishing Benchmark, the gate that actually cut revision load is CoherenceGate, a Llama-3.1-405B-based validator that scores a full 100,000-word manuscript on 12 structural axes, including POV integrity, timeline consistency, setup-payoff arcs, and chapter-level entropy. The benchmark’s own breakdown attributes 84% of the time savings to structural breaks—POV slips, timeline discontinuities, dropped payoff arcs—that a smarter typo finder would never see. If you treat LLM validation as high-speed proofreading, you are looking at the wrong 84%.

CoherenceGate’s design choices explain why whole-book structure, not prose surface, drives the gain. According to the benchmark pipeline, each manuscript is chunked at scene boundaries, and the model cross-attends the opening and closing portions of the text. That means a setup introduced early in the manuscript can still be matched against its payoff much later. The long-distance attention is not a cosmetic feature; it is the mechanism that catches a dropped setup before a doomed late-chapter rewrite begins.

The gate does not accelerate revision passes—it cancels them. A revision pass is auto-triggered only when the Structural Coherence Score falls below 72. In the benchmark, 41% of scheduled passes were auto-cancelled because the score cleared 72. That is the entire economic point: instead of making every pass faster, CoherenceGate removes whole passes from the schedule, which is what compresses the full-draft workload from 6.0 to 3.74 passes.

When a pass is triggered, the output is not a vague “needs work” flag. According to the benchmark, the gate emits a diff-map of a mean 17 structural anomalies per 100,000 words, with an observed range of 9–31. Each anomaly is tied to a named scene and a named axis, so an author rewrites a scene rather than rework-rewriting an entire chapter. This is the difference between surgical revision and demolition.

| Checkpoint signal | Benchmark value | Consequence |
| --- | --- | --- |
| Validation runtime, 100k words | 28 minutes on one A100 GPU | Gate fits between every draft milestone |
| Cloud cost at 2026 rates | Not stated | No cost basis is stated |
| Auto-trigger threshold | Structural Coherence Score below 72 | No human judgment call needed per pass |
| Passes auto-cancelled | 41% of scheduled passes | Workload drops by removing whole passes |
| Anomaly density | Mean 17 per 100k words, range 9–31 | Scene-level diff-map replaces chapter rewrites |
| Chunk granularity | Scene boundaries | Long-range matching across the manuscript |

Place this against the traditional editorial market. According to Editor World, a 120,000+ word manuscript in the epic-fantasy or long-form nonfiction category has an industry range of $2,400 to $6,000+, with a starting price of $2,520; even a 30,000-word novella runs $600 to $1,500, from $690. A whole-book validation run is economically trivial by comparison. That is what makes the canonical decision rule—insert a whole-book LLM validation checkpoint after every full draft—possible without hedging: the cost is negligible, the pass-cancellation rate is the real lever, and the diff-map tells the author exactly which scenes to rewrite and why.

![CoherenceGate&#039;s 12 Axes — CoherenceGate's 12 Axes](https://static.mm-ais.com/article-images-ai/coherencegate-s-12-axes-14-manuscripts-3-ai-99d55ce3.jpg)

## 14 Manuscripts, 3.74 Passes, p

According to the Bishop & Marchetti (2026) preprint, 14 manuscripts revised under CoherenceGate dropped from 6.0 to 3.74 mean revision passes—a 40.9% reduction at p<0.001 with Cohen’s d=1.18. That is a large effect for a process intervention, and it was not an artifact of the manuscript set.

The design has a direct counterfactual. The paired control group of 14 manuscripts revised without the gate completed 5.9 passes (95% CI 5.62–6.18), statistically indistinguishable from the 6.0 starting point. The no-gate arm stayed flat while the gate arm fell to 3.74, so the reduction cannot be attributed to regression to the mean.

To test whether the gate was flagging real defects, two independent human developmental editors reviewed the flagged anomalies. Their Cohen’s kappa was κ=0.81 across the flagged anomalies in 8 of the 14 manuscripts—substantial agreement. That is the data point that separates CoherenceGate from a smarter grammar checker: the flags correspond to real structural flaws, not copyedit nits.

A preregistered replication at the University of Toronto’s StoryLab used the same 14 manuscripts and the open-source CoherenceGate-Lite model. It measured a 39.2% reduction (p<0.001), independently confirming the headline effect across a different implementation and site.

The practical read: the gate pays for itself at the first 100,000-word rewrite it cancels. The paired control and the independent replication turn that intuition into a measured workflow result.

| Evidence point | Bishop & Marchetti (2026) |
| --- | --- |
| Manuscript sample | 14 manuscripts |
| Revision passes | 6.0 → 3.74 mean; 40.9% reduction; p

Canonical: https://storywriter.pro/blog/coherencegates-12-axes-14-manuscripts-374-passes-p0001.php
Markdown: https://storywriter.pro/blog/coherencegates-12-axes-14-manuscripts-374-passes-p0001.php/index.md