Sweep the Whole Document Set for Contradictions in One Pass

Workflow area: Requirements Quality · Tier: Core · Hands-on time: ~35 minutes

Documents diverge because they are written at different times, by different hands, for different audiences — and no human rereads the full set side by side once it passes fifty pages. The contradictions hide in the seams: a threshold updated in one document and not another, "customer" meaning three different entities across four artifacts, a requirement that exists in the BRD and appears nowhere downstream. Long-context models are built for exactly this cross-referencing work, and it is among the highest-value, lowest-risk AI tasks in the BA toolkit — because every finding is a checkable claim about two pieces of text you possess. The discipline: findings arrive as citations, you verify both sides at source, and resolution is a human decision recorded with a name, because which document is right is a judgment about intent, not text.

When to use — and when not to

• Use it before every baseline or major reviewthe sweep is the pre-flight check that catches divergence while it is an edit, not a change request.
• Use it when documents have different authors or erasa vendor FRD against your BRD, or a Phase 2 pack against Phase 1 decisions, is where contradiction density peaks.
• Use it after every substantial change roundeach CR or re-scope is a fresh chance for one document to move and its siblings to stay behind (pair with pb-22).
• Skip it on drafts still in churnsweeping documents that change daily produces findings that are stale by review time; sweep at stabilization points.
• Skip the auto-fixnever let the model "harmonize" the documents. A contradiction is a symptom; the cure is a decision about which position is intended, and that decision has an owner.
• Skip it as a substitute for reviewthe sweep finds inconsistency between documents, not wrongness within them. A perfectly consistent set can still be perfectly wrong about the business.

The recipe

1. 1.

   Freeze and label the set. Collect the exact versions in force: BRD, FRD, story export, process-map narratives, glossary, decision log. Note version and date for each — findings will be cited against these, and a sweep over mixed versions produces phantom contradictions.

2. 2.

   Run the sweep. Attach the set to Claude (long context) with this prompt:

   Copy

   You are a consistency auditor for a requirements document set. Find
   problems BETWEEN documents; do not critique individual quality.
   Document set: [LIST with versions/dates]
   Precedence if known (e.g. decision log > BRD > FRD): [STATE OR
   "none declared"]
   
   Find and report:
   1. DIRECT CONTRADICTIONS — two artifacts stating incompatible
      values, rules or behaviors. Quote BOTH passages verbatim with
      document + section/line.
   2. TERMINOLOGY DRIFT — one concept under multiple names, or one
      term meaning different things. Cite each usage.
   3. ORPHANS — BRD requirements with no downstream FRD item or
      story; stories tracing to no BRD requirement.
   4. STALE REFERENCES — citations of superseded decisions, dead
      section numbers, renamed artifacts.
   
   For each finding: SEVERITY (HIGH = would change build or scope;
   MED = would confuse delivery; LOW = cosmetic) and a one-line
   suggested resolution QUESTION (not a fix).
   Rules: quote verbatim, never paraphrase inside quotes. If two
   passages are reconcilable under a reasonable reading, mark
   POSSIBLE and say what reading reconciles them. Do not invent
   section numbers.
   

3. 3.

   Verify both sides of every HIGH. Good output looks like this:

   Copy

   HIGH — DIRECT CONTRADICTION
   BRD v2.1 §4.3: "Vendor invoices are auto-approved up to $10,000"
   FRD v1.4 §7.2.1: "Auto-approval threshold: $25,000 (per Finance
   workshop)"
   Stories: JIRA-214 acceptance criteria assume $25,000.
   Question for resolution: which threshold did the sponsor approve,
   and where is that decision recorded?
   
   MED — TERMINOLOGY DRIFT
   "Vendor" (BRD §2), "supplier" (FRD throughout), "payee" (JIRA-190)
   — glossary defines only "vendor".
   

   Open both cited passages for every HIGH and a sample of MEDs. You are checking that quotes are verbatim, citations land, and the contradiction is real rather than two statements about different scopes. Long-context models occasionally manufacture a plausible section number — that is why citations are the contract.

4. 4.

   Adjudicate with owners, not with edits. Take verified HIGHs to the people who can rule: sponsor for scope-level conflicts, product owner for story-level, architect for technical. Each ruling gets recorded — winner, rationale, decider, date — in the decision log. Terminology drift gets one canonical term ratified into the glossary; orphans get traced, descoped, or promoted deliberately.

5. 5.

   Propagate and re-sweep. Apply the rulings across every affected artifact (the losing document changes too — that is the point), then re-run the sweep. The second pass should return only LOWs and known-open items; attach it to the baseline as evidence the set is coherent.

Make it yours

• Claude Project as the standing auditor: keep the current document set in a Project and re-run the sweep at each stage gate; the findings trend is a quality metric for the program.
• Rovo/Confluence for Atlassian estates: where BRD and FRD live as pages and stories in Jira, grounded Q&A can chase specific suspected conflicts; use the full sweep prompt for the systematic pass.
• Chunked sweeps for huge sets: beyond one context window, sweep pairwise (BRD×FRD, FRD×stories) and union the findings — pairwise misses three-way drift, so keep the glossary in every batch.
• Sweep the vendor's pack on arrival: run the same pass over a vendor's own document set before fit-gap work; their internal contradictions become your clarification list and negotiating leverage.

Quality check

• Both sides of every finding were opened at sourcequotes verbatim, citations landing on real sections; any finding that failed verification was deleted, not softened.
• Severities reflect build impactHIGHs would genuinely change scope, build or money; a sweep where everything is HIGH was not triaged.
• Every HIGH has a named rulingwinner, rationale, decider and date in the decision log; a contradiction "resolved" by editing one document without a decision will return.
• Fixes propagated to all artifactsthe losing documents were corrected, traceability links updated, and the glossary now carries the ratified terms.
• Orphans were dispositionedeach one traced, deliberately descoped, or promoted; none merely noted.
• A clean re-sweep is on filethe post-fix pass with only LOWs/known-opens is attached to the baseline as coherence evidence.