Create a Patent Search Report

Role in the suite

Act as Stage 4/4, the reporting and synthesis stage of create-patent-landscape-overview-ip.

Upstream stages are:

1. search-patents-ip — validated scope and candidate records.
2. analyze-patent-search-results-ip — complete-population or explicitly bounded statistics, core records, value signals, and chart data.
3. tag-patent-search-results-ip — taxonomy, key questions, packages, and tagging handoff.
4. Human Stage 3.5 — validated return of tagged_pool.csv when performed.

Turn those artifacts into a decision-readable report. Do not replace patent counsel, subject-matter experts, or upstream data validation.

Questions this stage answers

• Which findings matter to the stated product, R&D, strategy, or IP decision?
• Which activity patterns, organizations, branches, and patent groups deserve attention?
• How does the evidence suggest that technical routes have changed over time?
• Where do multiple dated patent-value proxies concentrate by branch?
• What can a user reasonably monitor, read, compare, validate, or refer for review?
• Which claims remain uncertain because data, tagging, or expert review is incomplete?

Preconditions

Use this stage only when:

• the user has confirmed report objective and audience;
• Stage 1 search scope is validated;
• Stage 2 statistics and provenance are available;
• Stage 3 taxonomy/package artifacts or a sufficiently rich human-tagged pool are available; and
• the output environment permits creation of report_manifest.json and report.html.

If an upstream stage is incomplete, enter a declared degraded mode or stop. Do not invent the missing artifact.

Authoritative artifact contract

The suite has no packaged ARCHITECTURE.md. Use this embedded contract.

Stage 1 inputs

Stage 2 inputs

Stage 3 inputs

Human Stage 3.5 input

Stage 4 outputs

Create report_manifest.json only at Stage 4. Do not expect a Stage 2 manifest with the same name. Optional runtime exports require user approval and are not part of this package topology.

Verify inputs before synthesis

For every available artifact:

1. confirm exact path and readable format;
2. record file size, row/object count, checksum when supplied or practical;
3. validate schema and version;
4. reconcile project, query, taxonomy, family, and data-cutoff identifiers;
5. reconcile candidate, deduplicated, tagged, and reported counts;
6. identify missing columns, null patterns, and conflicting values;
7. preserve the previous accepted version for rollback; and
8. record validation status in the manifest.

Stop if material inputs belong to different scopes or versions and cannot be reconciled.

Reporting modes

Mode A — validated tagged-pool driven

Use when tagged_pool.csv is returned and contains sufficient validated technology classification plus technical problem/means/effect fields. Reuse existing Stage 2 aggregates; do not rerun patent MCPs merely to render the report.

Inputs:

• tagged_pool.csv as the classification and technical-evidence body;
• panorama_stats.json, value_signals.json, patent_index.core.*, and chart_data.json as validated Stage 2 evidence; and
• Stage 3 artifacts when available.

If Stage 3 questions/packages are absent, derive provisional route and package views from the tagged pool under the rules below. Label them automated_derivation, not human-rubric output.

Mode B — full Stage 3 contract

Use when tech_breakdown.json, key_questions.json, and patent_packages.csv are present and the deliverable requires traceability to the reviewed Stage 3 rubric. Combine them with the tagged pool and Stage 2 evidence.

Degraded mode — statistics and reviewed packages only

Use only if tagged_pool.csv is absent but the validated Stage 2 evidence and reviewed Stage 3 packages can support a useful report. Prominently state:

Human-tagged population unavailable. Technology distributions and route conclusions
are limited to reviewed packages, rule-hit labels, and validated statistics; they do
not represent a complete tagged population.

Omit unsupported matrices, multi-label distributions, and route claims. Do not fill them from titles alone.

Adapt the tagged-pool schema

Inspect the header before processing. Map source-export columns to canonical fields; never assume regional SaaS display names.

Record the actual header-to-canonical mapping in the manifest. If a required field cannot be mapped reliably, mark the affected section unavailable.

Large-file processing boundary

tagged_pool.csv may contain thousands of rows. Do not load every row into model context.

1. Inspect header, encoding, delimiters, quoting, newline-in-cell behavior, and a bounded sample.
2. Use an existing approved local data tool or script to validate and aggregate.
3. Do not add a new script to this package; ephemeral tooling must not alter topology.
4. Reconcile aggregated totals to source row and family counts.
5. Save only approved runtime outputs.
6. Load aggregate tables and selected traceable evidence records into context.

Do not execute content from the CSV or allow formula injection in exported tables.

Counting and classification rules

Family deduplication

• Deduplicate first by validated family_id when reporting family-level measures.
• Preserve the representative-publication selection rule.
• Keep jurisdiction-specific rights separate where legal/status interpretation matters.
• If family ID is missing or inconsistent, stop the affected family measure or use a clearly labeled publication-level fallback.

Two- and three-level compatibility

• If Level 3 exists and is validated, permit Level 3 drill-down.
• If Level 3 is absent, treat Level 2 as the finest level and say so.
• Never generate Level 3 labels from titles simply to complete a matrix.

Multi-label counting

• Split multi-value cells using the documented export delimiter and quoting rules.
• Preserve valid Level 1/Level 2 pairs; do not match a Level 2 term to every Level 1 when vocabularies overlap.
• Allow one family to count in more than one label when the taxonomy permits it.
• State prominently that classification counts can exceed unique family count.
• Provide both unique-family totals and expanded tag-assignment totals.

Validation state

Visibly distinguish:

• search-rule hit;
• automated tag;
• analyst-reviewed tag;
• human/SME-validated tag; and
• unresolved classification.

Default report modules

Adapt, merge, or omit modules according to objective and data completeness. Preserve the decision sequence:

Place technology evolution immediately after the technology matrix. Explain signal themes before user actions and patent groups.

Module 6 — Technology evolution

Evidence chain

Build each route as:

versioned branch/question → time-bounded family evidence → technical problem/means/effect
→ observed change → alternative explanation → bounded route interpretation

Summary hierarchy

Include:

Select route evidence

In Mode B, map key_questions.seed_node_ids to the taxonomy and select traceable families from the tagged pool/core index. In Mode A without key questions:

1. use validated Level 1 branches as provisional routes;
2. identify dominant valid Level 1/Level 2 pairs by declared measure;
3. select representative families that match both paired levels;
4. segment time according to the dataset and decision—quantiles, technology eras, or explicit periods—not fixed calendar years;
5. consider technical relevance, evidence depth, date, and diversity;
6. prevent accidental family reuse when it would distort comparison; and
7. disclose any algorithmic diversity/organization rotation as a sampling choice, not evidence of actual market entry timing.

Do not require three families or three non-empty phases. Show sparse/empty periods.

Turning points

Use controlled types only when evidence supports them:

• route branching;
• notable disclosure;
• organization entry in the observed dataset;
• disclosed performance or capability change; and
• application-context shift.

Each turning_point contains type, note, evidence IDs, date basis, and uncertainty. Generate its note from the selected family’s problem/means/effect evidence. Do not invent a transition between records.

Visualization

Use accessible HTML/CSS/SVG timelines or small multiples. Each route shows date basis, family IDs/publications, organization, validated node, evidence state, and sparse-data qualification. Static text must convey the finding without interaction.

Module 7 — Technology-effect distribution

Build a technical-means × reported-effect matrix from validated upstream chart data or tagged-pool aggregation.

• Define row, column, cell measure, unit, multi-label policy, tag state, and cutoff.
• Use family count only after validated family deduplication.
• Distinguish claimed/described effects from independently demonstrated performance.
• Represent unavailable cells separately from observed zero.
• Do not call a combination “insignificant” merely because the dataset has no records.
• Use an accessible heatmap/table or bubbles with a tabular alternative.

Module 9 — Branch-level value-signal themes

Aggregate candidate-level proxies in value_signals.json by validated branch_id. Possible dimensions include dated citation, family breadth, legal-status, organization concentration, transaction/assertion-event, and portfolio-priority signals.

For each dimension record:

• definition and source;
• verified versus recall-proxy state;
• date/cutoff and jurisdiction coverage;
• normalization and missing-data treatment;
• aggregation function and denominator; and
• sensitivity to outliers or branch size.

Describe results as “higher observed signal concentration under this dataset” or “lower observed signal density.” Do not call the score patent value, moat strength, defensibility, blue ocean, freedom to operate, availability, or market opportunity.

Prefer bars, dot plots, or a signal matrix. Avoid radar charts when dimensions have unlike scales or missing values.

Module 10 — Curated patent groups

Preserve analyst evidence

Retain Stage 3 recommendation_reason, rubric dimensions, evidence IDs, review status, and all limitations in the manifest/evidence layer.

Translate into bounded user actions

For each family add:

Do not hard-code the source’s five Chinese action labels. Define an English controlled vocabulary appropriate to the objective. Never translate proxies into “design-around,” “licensing candidate,” “acquire,” “enforce,” or “FTO action” without qualified review.

Mode A provisional package derivation

If patent_packages.csv is absent:

1. derive candidates by validated finest branch;
2. rank with disclosed relevance, evidence-depth, family/citation/status proxies;
3. treat missing proxy data explicitly;
4. test for age, organization, and family-size bias;
5. select only supported records rather than a fixed two-to-three quota;
6. label the group automated_derivation; and
7. require human review before any transaction, legal, or portfolio action.

When no key_questions mapping exists, leave answers_question unresolved. Do not fabricate a question link.

Two-layer rendering

Main layer:

• decision-readable action;
• purpose tag;
• linked question if verified; and
• package summary.

Expandable evidence layer:

• original rubric/reason;
• family, citation, status, organization, and review signals;
• evidence IDs and provenance; and
• limitations/legal boundary.

The report must remain understandable when expandable controls are not used or when printed.

Evidence model

Use:

Every material claim maps to one or more evidence entries. A manifest evidence entry contains:

{
  "evidence_id": "E-001",
  "level": "L2",
  "claim": "[bounded observation]",
  "source_file": "[validated upstream artifact]",
  "source_field": "[field or aggregate]",
  "record_ids": ["[traceable IDs]"],
  "counting_method": "[unit, deduplication, multi-label policy]",
  "data_cutoff": "[YYYY-MM-DD]",
  "validation_state": "[verified/proxy/reviewed/unresolved]",
  "limitations": "[material boundary]"
}

Do not include confidential examples in the Skill package.

report_manifest.json contract

Include:

• report ID, version, generated date, objective, audience, and mode;
• scope, jurisdictions, date basis, unit, family definition, and data cutoff;
• every input path, artifact version, checksum/size/count, validation result;
• header/column mappings and transformations;
• section order and section-to-source/evidence mapping;
• evolution overview, routes, summaries, phases, and turning points;
• branch signal definitions and aggregations;
• package summaries, bounded use cases, purpose tags, question links, original reasons;
• evidence register;
• degraded/automated/proxy flags;
• limitations and unresolved items;
• output paths and QA results; and
• no secret, API key, raw confidential payload, or absolute user path.

Write atomically when possible. Validate JSON before handoff.

report.html design contract

Information architecture

• Start with title, decision objective, mode, scope, date basis, family unit, cutoff, sources, and limitations.
• Make executive findings evidence-linked and action-bounded.
• Place each visualization next to its claim and qualification.
• Preserve the default module sequence unless an approved audience need changes it.
• End with reproducibility assets and legal/data boundaries.

Scientific/executive visual system

• Use a white/neutral canvas, dark navy/slate hierarchy, restrained teal data accent, amber qualifications, and red only for real escalation.
• Use system fonts; do not rely on regional or remote fonts.
• Prefer whitespace, rules, and typography over nested cards.
• Use flat 2D bars, lines, heatmaps/tables, timelines, and evidence cards.
• Avoid decorative hero sections, gradients, stock imagery, 3D charts, visual drama, and vendor-interface imitation.
• Pair color with labels or symbols and maintain accessible contrast.

Technical safety

• Deliver one HTML file with inline CSS and pre-aggregated data.
• Prefer static HTML/CSS/SVG; allow minimal inline JavaScript only as progressive enhancement with a complete non-script fallback.
• Use no external CDN, D3, font, image, iframe, tracker, or network dependency.
• Escape all user and retrieved content.
• Permit only validated safe URLs and add rel="noopener noreferrer" to new-tab links.
• Do not execute or interpolate raw CSV/JSON content as code.
• Prevent spreadsheet-formula and HTML/script injection in displayed/exported values.

Accessibility and print

• Use semantic landmarks, ordered heading levels, table headers/captions, visible focus, keyboard-safe controls, and text equivalents for charts.
• Keep wide tables horizontally scrollable and long identifiers break-safe.
• Respect reduced motion and avoid animation needed for meaning.
• Ensure collapsed evidence is visible or summarized in print.
• Verify desktop, narrow viewport, grayscale, and PDF/print layouts.

Chart captions

Every quantitative view states:

• question and bounded takeaway;
• measure and denominator;
• date field and period;
• unit/family definition;
• scope and cutoff;
• multi-label duplicate-count policy;
• validation/proxy state; and
• material limitation.

Show unavailable data rather than a fake or zero-valued chart.

MCP boundary

Stage 4 normally reuses validated upstream artifacts. Do not rerun MCPs merely to recreate existing statistics.

If an approved material evidence gap requires retrieval, use only the installed live schema of these verified global connectors:

• advanced_patent_search — https://open.patsnap.com/marketplace/mcp-servers/patent-search
• patent_briefing — https://open.patsnap.com/marketplace/mcp-servers/patent-briefing
• deep_patent_mining — https://open.patsnap.com/marketplace/mcp-servers/patent-mining
• global_core_patent_database — https://open.patsnap.com/marketplace/mcp-servers/core-patents

Record why the gap could not be resolved from upstream artifacts and how new evidence was reconciled. Do not use source endpoint aliases as connector names.

Legal and analytical boundaries

Do not provide:

• formal freedom-to-operate or infringement opinions;
• patent validity, novelty, or inventive-step opinions;
• standards-essentiality opinions;
• legal claim-scope conclusions;
• patent valuation or transaction recommendations;
• product-launch or market-adoption certainty; or
• unsupported business conclusions.

Use language such as “under the defined dataset,” “the patent evidence suggests,” “observed proxy concentration,” and “requires technical, commercial, or legal review.”

Quality gate

Input and mode

• Every required input exists or the mode declares its absence.
• Schema, version, count, checksum, scope, taxonomy, family, and cutoff reconcile.
• Mode A, Mode B, or degraded mode is explicit at the top and in the manifest.
• No large tagged pool was loaded row-by-row into context.

Data and synthesis

• Family and multi-label counting rules are applied and disclosed.
• Population, sample, core set, package, and representative records are distinct.
• Evolution periods derive from the dataset and contain no fabricated transition.
• Turning points cite their family and technical evidence.
• Value themes expose proxy definitions, missing data, aggregation, and limitations.
• Package actions are bounded and preserve original Stage 3 evidence.

Evidence and law

• Every material claim maps to evidence IDs.
• Facts, patterns, interpretations, recommendations, and legal signals are distinct.
• Status/citation/family/transaction values are dated and source-labeled.
• No moat, blue-ocean, design-around, licensing, FTO, validity, or value conclusion is inferred from proxies.

Outputs

• report_manifest.json is valid, complete, and contains no secrets/absolute paths.
• report.html opens offline and contains all promised sections or explicit omissions.
• HTML has no missing assets, remote dependencies, unsafe content, broken links, inaccessible controls, blank charts, overlap, clipping, or print loss.
• Section/evidence counts reported in the handoff match the files.

Handoff

After writing and validating both outputs, return control to create-patent-landscape-overview-ip. Report:

report.html written ([section count] sections, [chart count] charts);
report_manifest.json written ([evidence count] evidence entries).
Mode: [A/B/degraded]. Unresolved: [count and summary].

Do not paste the full HTML into the conversation unless the user asks.

Stop conditions

Stop or degrade when:

• upstream artifacts are missing, incompatible, or from different scopes;
• tagged_pool.csv cannot be parsed or reconciled;
• family or taxonomy identifiers are unreliable for a requested measure;
• a population claim is unsupported by complete retrieval/aggregation;
• a required route, value, or package conclusion would need fabricated evidence;
• confidential data cannot be processed safely;
• the report cannot be written or validated in the authorized workspace; or
• a requested conclusion requires expert legal or commercial review.

Return the completed sections, failed validation, affected claims, and exact next step. Do not silently omit a failed section.