Given N PDFs on the same subject, do NOT summarize each one separately. Instead:

1. Inventory all files → 2. Extract every substantive idea with source tags (I001, I002...) → 3. Compare across documents (consensus / unique / duplicate / contradiction) → 4. Assign a disposition to every idea (INCLUDE/MERGE/ADAPT/VALIDATE/DEFER/REJECT/UNRESOLVED) → 5. Write one synthesized master document organized by topic, not by source → 6. Attach audit appendices (idea register, comparison matrix, conflict log, source index) → 7. Run coverage/consistency/rendering QA → 8. Deliver.

Minimal example of a traced idea record:

I014 | "Use feature flags for rollout" | Sources: D2(p.7), D4(p.3,p.19)
Category: Implementation | Agreement: same substance, different rationale
Disposition: INCLUDE (merged with I009) | Evidence: reasoning only, no data cited

Treat each run as a persistent project with a running idea register — never discard analysis between batches or documents.

Progress:
- [ ] Stage 1: Inventory every file (name, source ID, pages, structure, integrity issues)
- [ ] Stage 2: Full extraction of every accessible page from every PDF (no executive-summary-only shortcuts)
- [ ] Stage 3: Normalize ideas into a canonical taxonomy, assign IDs, link occurrences
- [ ] Stage 4: Cross-document comparison — consensus, unique contributions, duplication, contradiction, gaps
- [ ] Stage 5: Critical evaluation — assign a disposition to EVERY substantive idea
- [ ] Stage 6: Write master synthesis organized by topic/decision, not by source
- [ ] Stage 7: Assemble required PDF structure (exec synthesis, detailed analysis, audit appendices)
- [ ] Stage 8: Traceability pass — every material claim cites [Source: Dx, p.y]
- [ ] Stage 9: QA gates — coverage audit, consistency audit, PDF production audit
- [ ] Stage 10: Deliver A/B/C/D with completion report

Stage details

Inventory (Stage 1). For each file record: filename, source ID (D1, D2...), page count, title, apparent purpose, structural elements (tables/figures/appendices), and extraction risks (scanned pages, embedded images, missing text layer). Flag anything requiring OCR or image inspection. Do not proceed to synthesis until gaps are identified and disclosed.

Extraction (Stage 2). Read every page. Capture: arguments, recommendations, alternatives, reasoning, stats/evidence, assumptions, examples, trade-offs, risks, terminology, tables/diagrams, caveats hidden in footnotes/appendices. Assign each extracted item an ID (I001...) with source doc, page/section, and a faithful paraphrase (short exact quotes only when wording is essential).

Normalization (Stage 3). Build a subject-appropriate taxonomy (e.g., principles, problems, recommendations, risks, metrics, assumptions, open questions — adapt freely). Create canonical records linking all original occurrences of equivalent ideas. Keep different ideas under similar labels separate; keep same idea under different labels linked.

Comparison (Stage 4). For every significant idea, explicitly determine:

• Consensus: which docs share it, same substance or just similar language, same or different rationale.
• Unique: what's distinct, whether it's compatible with shared recommendations, what disposition it merits.
• Duplication type: exact / semantic / complementary-to-merge / similar-with-differences / related-but-different-problem.
• Contradiction: conflicting positions, who holds them, why (factual/strategic/definitional/assumption-based), and whether resolvable.
• Gaps: omissions shared by all docs or raised by one and ignored by others.

Never treat majority agreement as proof of correctness, and never resolve conflicts by picking the majority, the longest explanation, or the most confident tone.

Evaluation & disposition (Stage 5). Judge each idea on strategic alignment, expected value, evidence quality, feasibility, risk, novelty, and complementarity — qualitatively, not with invented numeric scores. Assign exactly one disposition: INCLUDE, MERGE, ADAPT, VALIDATE, DEFER, REJECT, or UNRESOLVED, with a stated reason. Rejected/deferred/unresolved ideas stay visible in appendices — never silently dropped.

Synthesis (Stage 6). Organize by topic/problem/decision. Lead with strongest conclusions. Merge overlapping recommendations into unified explanations while preserving distinct reasoning threads. Fold in unique insights where they add value. Explain trade-offs and disagreements openly. Keep caveats attached to their claims. Label each claim as source-derived, synthesized, or inferred.

Example 1 — Consensus with different rationale: Input: D1, D3, and D5 all recommend "implement rate limiting on the API." Output: Master doc states the recommendation once, in the Implementation chapter, noting: "[Source: D1 p.4, D3 p.11, D5 p.2] All three sources recommend rate limiting; D1 justifies it via abuse prevention, D3 via cost control, D5 via SLA protection — the convergence from independent rationales strengthens the case. Disposition: INCLUDE (merged as I022)."

Example 2 — Unique contribution preserved: Input: Only D4 proposes a "circuit breaker pattern" absent from the other six documents. Output: Included in the Technical Requirements chapter with full explanation: "[Source: D4, p.9] Unique to D4. Addresses cascading-failure risk not discussed elsewhere. Compatible with the shared resilience recommendations. Disposition: ADAPT — generalize from D4's single-service example to the multi-service architecture discussed by the consensus sources."

Example 3 — Contradiction surfaced, not hidden: Input: D2 claims migration will take "2 weeks"; D6 claims "2-3 months" for what appears to be the same migration. Output: Logged in the Conflict and Decision Log: "Conflict C07 — D2 [p.6] vs D6 [p.14]. D2's estimate assumes an existing CI/CD pipeline; D6 assumes building one from scratch. Disagreement is assumption-based, not factual. Recommended resolution: treat D6's estimate as authoritative unless pipeline already exists; disposition UNRESOLVED pending confirmation of current infrastructure state."