Contradiction Detection¶

Context: Cross-chapter consistency checking for a 12-chapter, 81,000-word manuscript. Found and resolved 26 contradictions between main text and companion materials in a single pass. Most were stat conflicts or terminology drift -- the kind of problems that compound silently across months of writing.

Chapter 1 says Harvey reached $100M ARR. Chapter 9 says $75M. Both were written with citations. Both were accurate at the time of writing -- the number changed between research sessions. Without cross-chapter detection, both numbers ship. A careful reader catches the conflict. Your credibility takes the hit.

In a 12-chapter book with 81 sections written across multiple agent sessions over weeks, contradictions aren't a risk. They're a certainty. The question is whether you catch them before your readers do.

Types of Contradictions¶

Not all contradictions are equal. Some destroy credibility. Others just create confusion. The detection system categorizes five types:

Type	Example	Risk Level
Stat conflicts	Harvey valuation: "$5B" in Ch9, "$8B" in Ch1	High -- readers notice numbers
Terminology drift	"AI gateway" in Ch4, "AI router" in Ch5 for the same concept	Medium -- creates confusion over time
Advice contradictions	"Start with pilots" in Ch2, "Go all-in" in Ch8	High -- undermines trust in recommendations
Example inconsistency	Air Canada damages: "$650.88" in one section, "$812.02" in another	High -- suggests sloppy fact-checking
Framework conflicts	Different step counts or ordering for similar frameworks	Medium -- weakens structural clarity

Every high-severity contradiction in the actual book had a legitimate explanation -- a funding round happened mid-writing, or different sources reported different components of a ruling. None of those explanations matter to a reader who notices two different numbers for the same fact.

Detection Methods¶

Three approaches, from cheapest to most thorough:

1. Grep-Based Scanning¶

Search for key terms, company names, and numbers across all chapter files:

# All mentions of a company with context
grep -rn "Harvey" Book/drafts/Draft\ 3/ --include="*.md" -C 2

# All dollar amounts to check for conflicts
grep -rn "\$[0-9]" Book/drafts/Draft\ 3/ --include="*.md"

Catches obvious conflicts -- same company, different numbers. Misses terminology drift and advice contradictions that require semantic understanding.

2. Pre-Scan Deduplication¶

The publish-review skill's deduplication scan runs before editorial work begins on any chapter. It checks for:

Stat repetition -- the same statistic appearing in multiple chapters (even if consistent, over-repetition weakens impact)
Stat conflicts -- the same metric with different values across chapters
Company over-reliance -- the same company example appearing too frequently (Harvey in 8 chapters was a flag)
Verbal tics -- the same transition phrase or opening pattern appearing across multiple sections

This catches more than grep because it understands context. It knows that "$100M ARR" in Chapter 1 and "$75M ARR" in Chapter 9 refer to the same company's revenue, even if the sentences are structured differently.

3. Full Manuscript Review¶

The most thorough approach: a reviewer agent with the complete manuscript loaded, specifically tasked with finding cross-chapter inconsistencies. This catches advice contradictions ("start with pilots" vs. "go all-in") and framework conflicts that require understanding intent, not just matching strings.

This is the approach that found the GitHub Copilot discrepancy -- Chapter 2 cited "78% suggestion acceptance rate" while Chapter 5 cited "30% suggestion acceptance rate." Both numbers were from legitimate sources. But they measured different things: 78% was the task completion rate, 30% was the actual suggestion acceptance rate. The fix wasn't changing a number -- it was clarifying what each number measured.

Resolution Process¶

Finding contradictions is step one. Resolving them without introducing new problems is the real work.

Step 1: Identify. Note every location where the conflicting information appears -- including companion materials, not just the main text.

Step 2: Verify. Go back to the source. Don't guess, don't average, don't pick the more impressive number. The Air Canada damages were standardized to $812.02 (total tribunal order) because the $650.88 figure was only the fare difference.

Step 3: Standardize everywhere. Every location gets the same number, same terminology, same framing. Miss one and you've created a new contradiction while fixing the old one. The Harvey valuation fix touched Chapter 9's intro, data moats section, companion case studies, and exported book copies.

Step 4: Use "primary mention + back-reference." The chapter where a fact is introduced gets the full explanation. Later chapters reference back: "Harvey, which reached $100M ARR (Chapter 1), demonstrates..." One authoritative source per fact. Later chapters point to it, not restate it.

Scale and Results¶

The full contradiction scan across the book and its companion materials found 26 contradictions:

Severity	Count	Examples
HIGH	4	Harvey valuation ($5B vs $8B), Harvey ARR ($75M vs $100M), Air Canada damages ($650 vs $812), cost metric misattributed as latency
MEDIUM	12	Copilot acceptance rates (different metrics conflated), ROI timeframes, terminology inconsistencies
LOW	10	Missing nuance in statistics, ordering differences, attribution details

All 26 resolved in a single pass. HIGH issues required updating book text, companion materials, and exported copies. MEDIUM and LOW were mostly companion updates to match the book as source of truth.

Prevention¶

Catching contradictions is the safety net. Preventing them is better.

Handoff protocols help. Each new agent session receives a summary of what previous chapters established -- key statistics, terminology choices, framework structures. This reduces the chance of an agent inventing a different number for the same metric.

Centralized research helps more. When every chapter pulls statistics from the same research files, conflicts are less likely. The Harvey ARR contradiction happened because research was updated mid-writing and earlier chapters weren't refreshed -- a process gap, not a data gap.

Regular cross-chapter scans are the real prevention. Run them every 3-4 chapters, not just at the end. Catching a conflict between Chapter 3 and Chapter 5 while writing Chapter 6 is far cheaper than catching it during final verification.

The contradictions will still emerge. Months of writing, evolving data, multiple agent sessions -- structurally inevitable. The question isn't whether you'll have contradictions. It's whether you have a system for finding them.