Our Scientific Approach

Why layer separation matters

Many analytics tools mix data retrieval, scoring, and interpretation into an opaque pipeline.Stratensight keeps these layers separate so that users can audit each step independently. If you disagree with an interpretation, you can still trust the underlying scores. If you question a score, you can inspect the raw data it was computed from.

Momentum Index™

A composite score of filing dynamics. Above 70 the pace is high; below 45 it is low. The score states a level, never a direction: direction is carried by the compound annual growth rate alone, shown next to it. A low Momentum does not establish a decline — it is also obtained on a small corpus whose growth rate could not be published.

VARIABLES

• CAGR (35 %) — compound annual growth rate of filings, between the first and last year of the window
• Filing peak recency (30 %) — where the busiest year sits relative to today, corrected for the slope of the last three years
• Recent year-on-year changes (20 %) — mean of the last three annual changes
• Corpus size (15 %) — document count, log-normalised between 100 and 5,000

CONCEPTUAL FORMULA

No curve is fitted. The 30 % term classes peak recency into four steps — peak in the past year 0.80; two to three years 0.60; four to six years 0.40; beyond that 0.25 — then adds the slope of the last three years. When the overall trend is near flat (under 3 % per year) the value is capped at 0.55: a peak on a flat series is usually noise. The S-curve remains the reference model for adoption cycles; it is not the measurement performed here.

Example, verified against the computing function. A corpus of 450 documents filed from 2016 to 2024, annual series 30 · 34 · 38 · 42 · 48 · 54 · 61 · 68 · 75. The CAGR is 12.1 %, the peak is on the last year, and the mean of the last three annual changes is 11.6 %. The four normalised components are 0.88 · 0.95 · 0.87 · 0.38, and the resulting Momentum is 83 (HIGH).

Lifecycle Position™

Places the corpus on five maturity rungs, from Inception to Late. A rung is a position, not a movement: it does not say whether filings are rising or falling — the compound annual growth rate says that. An Advanced position on a receding corpus is therefore not a contradiction.

VARIABLES

• Age factor (30 %) — years elapsed since the first patent in the corpus
• Filing deceleration (25 %) — slowdown of the annual pace
• Growth signal (25 %) — derived from the CAGR: neutral at 0 %, saturated at ±20 %
• IPC classification diversity (20 %) — Shannon entropy of the classes covered. CPC is read only as a fallback, when a document carries no IPC code
• Citation density — blended with the value above for 40 % when available

CONCEPTUAL FORMULA

THE 5 PHASES

Transition detection: the model monitors growth deceleration, CPC diversity shift, and assignee concentration changes to detect phase transitions automatically. Transitions are not instantaneous — a technology may exhibit characteristics of two adjacent phases during transition periods.

Technology Maturity Priors: For well-established technologies, Stratensight applies known maturity baselines to prevent misclassification when temporal data is limited. When a prior is applied, a transparency flag is surfaced so the reader knows the lifecycle was adjusted from the industry consensus, not from patents alone.

Priors de maturité technologique : pour les technologies établies, Stratensight applique des niveaux de maturité connus pour éviter une mauvaise classification.

EPO Indexation Delay: Patent filings from the last 18 months may not yet be fully indexed. CAGR on short windows may appear negative while the market is actually growing. Stratensight surfaces a warning flag in this case rather than silently reporting the skewed value.

Délai d’indexation EPO : les dépôts récents (<18 mois) peuvent ne pas être indexés ; le CAGR peut sembler négatif alors que le marché croît en réalité.

Openness Score™

THE MEASURE OF HOW FILINGS ARE DISTRIBUTED ACROSS APPLICANTS

Measures how filings are distributed across applicants. A high score means filings are dispersed across many actors; a low score, that they are concentrated on a few. The score describes concentration, not the ability to enter: freedom to operate depends on claim scope, which this score does not read.

VARIABLES

• Share of the top five applicants (45 %) — the dominant term. Piecewise-transformed, calibrated on the fact that a 25 to 40 % combined share is normal in patents
• HHI (20 %) — Herfindahl-Hirschman Index, the standard antitrust concentration measure used by the FTC and EU DG COMP
• New entrant rate (20 %) — first-time filers, over the number of distinct applicants
• Number of distinct applicants (15 %) — on a log scale

CONCEPTUAL FORMULA

In Stratensight, share is measured by the document count attributed to each applicant, over the documents carrying an identified applicant. The HHI is inverted so that a higher Openness score corresponds to less concentrated filings. A second regime applies to highly fragmented corpora — more than 300 applicants and an HHI under 0.05: the weights become 0.25 · 0.40 · 0.20 · 0.15, the HHI becoming the leading term because the top-5 share over-penalises in that regime.

The HHI itself is the standard antitrust concentration measure. The 0–100 Openness bands above — and the OPEN / CONTESTED / CONCENTRATED / DOMINATED cut-offs — are Stratensight’s own scale, set by expert judgement; they are not the antitrust authorities’ HHI thresholds, which use different cut-offs.

Intelligence Grade™

Measures the reliability of the analysis itself. Higher grades mean more trustworthy signals. Unlike the other scores which measure the technology, this score measures the quality of the data used to compute everything else.

VARIABLES

• Dataset size — volume of patents analyzed (more data = higher confidence)
• Temporal coverage — year span of filing dates (wider span = better trend detection)
• Data completeness — fill rate of key fields (abstracts, CPC codes, assignees, dates)

Clustering quality was a fourth variable until July 2026. It was removed: the silhouette score of a given dataset can differ from one machine to another, so it could move the Grade without any data changing. It is still measured and displayed as a cluster-quality indicator — it qualifies the clusters, it no longer scores the signal.

CONCEPTUAL FORMULA

WHY IT PENALIZES INCOMPLETE DATA

Missing abstracts reduce cluster quality. Missing filing dates impair Momentum calculation. Missing CPC codes weaken Lifecycle detection. Missing assignee fields make competitive analysis unreliable. The Grade reflects these real analytical impacts — it is not an arbitrary penalty but a direct measure of what the algorithms can and cannot compute.

CONTEXTUAL ADJUSTMENTS

Three contextual adjustments may reduce the Intelligence Grade™:

• Source retention: factor 0.65–1.0 depending on what percentage of patents passed filtering (≥60% retention = no penalty, <25% = ×0.65)
• Geographic coverage: up to −8 points if CN patents are underrepresented vs domain baseline
• Academic fallback: ×0.75 if academic sources substitute primary patent sources

These adjustments are applied automatically and reported in the transparency trace.

Signal Context

When a signal is degraded or unavailable, Stratensight displays the cause and recommended corrective action directly in the report. This transparency is deterministic — computed automatically, never AI-generated.

Quand un signal est dégradé ou indisponible, Stratensight affiche directement dans le rapport la cause et l’action corrective recommandée. Cette transparence est déterministe — calculée automatiquement, jamais générée par IA.

Decision Engine™

Combines all four scores into a single strategic verdict with confidence. The weighting reflects the relative importance of each dimension for technology investment decisions.

WEIGHTING AND FORMULA

LAYER A — VERDICT THRESHOLDS (AND-LOGIC)

The attention levels sit on a single ordered scale, from weakest to strongest signal: LOW COMPOSITE PACE → ISOLATED SIGNAL → PARTIAL CONVERGENCE → CONVERGING SIGNALS. ISOLATED SIGNAL is not an invitation to act — it is the weakest level above outright rejection. The cards below are listed strongest-first for readability.

All conditions must be met simultaneously. A high Momentum alone is not sufficient for CONVERGING SIGNALS. The Decision Score shown in the UI is an indicative composite — the level is determined by the multi-criteria AND logic above.

LAYER B — BLOCKING GUARDS

Eight canonical guards detect data conditions that compromise verdict reliability. When triggered, a guard either downgrades the level (e.g. CONVERGING SIGNALS → PARTIAL CONVERGENCE) or attaches an explanation surfaced in the UI alongside the result.

• bad_quality — Intelligence Grade™ below threshold; signal reliability reduced.
• small_dataset — Patent corpus too small for high-confidence conclusions.
• mature_declining — Late lifecycle position with a composite filing pace below 40.
• ip_concentration — Few actors control the IP landscape; entry barriers raised.
• research_phase — Technology still in research; commercial validation pending.
• lifecycle_na — Lifecycle stage could not be determined from the dataset.
• openness_na — Openness signal could not be computed from the dataset.
• ig_na — Intelligence Grade™ computation failed; signal reliability unknown.

Note: momentum_na is handled by a separate path (Coherence Validator rule C8) rather than the canonical guard set, because momentum unavailability triggers a deterministic downgrade of one level (CONVERGING SIGNALS → PARTIAL CONVERGENCE, PARTIAL CONVERGENCE → ISOLATED SIGNAL) before guard evaluation.

Absolute rule: no verdict is ever produced without a confidence score. Low Intelligence Grade™ triggers mandatory disclaimers on the verdict itself.

LAYER C — TIER GATE (CONFIDENCE)

Layer C subordinates the user-facing verdict to evidence_certainty — a categorical reliability label (HIGH / MODERATE / LOW / VERY_LOW) derived from the Intelligence Grade™ framework. A verdict is never presented without a coherent confidence signal attached. The mapping is categorical, not threshold-based.

TIER MODERATE — LEVEL PRESERVED + CONFIRM-FIRST CAVEAT

When evidence_certainty is MODERATE (~70%), the attention level stays intact — it is surfaced exactly as computed. What changes is a confirm-first caveat: the reading is real, but you are asked to confirm it before any strategic commitment. verdict_conditional stays false.

TIER LOW — THE SAME LEVEL, MARKED CONDITIONAL

When evidence_certainty is LOW, the dataset is rich enough to compute every score, but Signal Integrity warnings constrain the audit-grade confidence. The attention level does not change word — what is added is an explicit qualifier, rendered — Conditional in amber (for example PARTIAL CONVERGENCE — Conditional). Resolve the flagged Signal Integrity warnings to lift certainty before any strategic commitment.

TIER VERY_LOW — INSUFFICIENT EVIDENCE

When evidence_certainty is VERY_LOW, the input itself is genuinely unusable (typically n < 10 patents, temporal coverage < 1 year, or structurally corrupted data). No attention level is issued at all; the page reads INSUFFICIENT EVIDENCE: no signal can be defended. This is distinct from Conditional: Conditional means « we have a reading but warnings »; INSUFFICIENT EVIDENCE means « the dataset cannot support any reading ».

Why Layer C exists: prior to its introduction, a reading like « CONVERGING SIGNALS + Confidence 50 » was structurally possible because the AND-logic operates on raw scores while evidence_certainty derives from orthogonal signals (critical issues, temporal bunching, source coverage). Layer C subordinates the user-facing verdict to the certainty label so every output is auditable: here is the verdict, here is the confidence band that supports it.

The verdict is a synthesis tool, not a recommendation. It condenses multi-dimensional analysis into a single directional signal. Strategic decisions should consider the individual scores, not just the final verdict.

Critical Reader — Signal Integrity

Every analysis audits itself before it is presented. The Critical Reader layer runs between scoring and interpretation: it verifies that the verdict is mathematically consistent with the underlying scores, that the dataset has not produced silent artifacts, and that no rule of analytical hygiene has been violated. Runs on every analysis, on every plan, with no gating.

13 DETERMINISTIC RULES

• CE1 — Momentum vs YoY divergence (recent_yoy_mean coherence with CAGR)
• CE2 — Level vs AND-logic criteria (CONVERGING SIGNALS without all three thresholds met fires critical)
• CE3a — White Space saturation on mono-CPC clusters
• CE3b — White Space saturation on poly-CPC clusters
• CE4 — CPC diversity contradiction (concentration vs dispersion)
• CE5 — Temporal span consistency with phase detection
• CE6 — Momentum HIGH but the most-recent filing year already declining (grade-neutral coherence flag — CE1 covers the 3-year mean, CE6 the single latest year it masks)
• S1 — Source coverage anomaly (retention vs domain baseline)
• M_CAGR_LAST_YEAR_ARTIFACT — CAGR distorted by last-year filing lag
• L_ACADEMIC_DOMINANCE — Lifecycle bias from academic source fallback
• D_ABSTRACT_FILL_CRITICAL — Abstract fill rate below clustering threshold
• D_GRADE_PARTIAL_METADATA — Partial assignee/abstract/date fill rates trigger a narrative disclaimer

AI AUDITOR — CONTEXTUAL LAYER

Claude Haiku 4.5 reads the full analysis context (scores, metadata, source mode) and may surface up to 8 additional issues that the deterministic rules cannot express. Hard guardrails apply: allowed_values whitelist, ±0.5 float tolerance, 15-second timeout, 6000 max output tokens. The auditor never invents a fact and never re-scores — it can only flag.

SEVERITY MODEL

• critical — mathematical or pipeline inconsistency (e.g. CONVERGING SIGNALS with a failing AND criterion). The level remains visible but should be reviewed before action.
• warning — data-quality concern that may bias the signal (small dataset, low coverage, abstract fill below threshold).
• info — legitimate downgrade by a Layer B guard (research-phase, IP concentration, mature+declining). Surfaces the constraint, never blocks the result.

RELIABILITY GUARANTEES

• Fail-safe: an audit failure never blocks the pipeline — the analysis is delivered with a Signal Integrity unavailable badge.
• No gating: the audit is shown on every analysis, every plan. Scientific credibility cannot be fragmented.
• Cache-safe: audit results are cached 24h keyed by (scores + metadata + source mode + locale).
• Budget kill-switch: 500 LLM audits per day across the platform — a safety bound, not a cost optimization.

WHY A CONFIDENT VERDICT CAN STILL CARRY A CAVEAT

Some audit flags are grade-neutral: they leave the verdict and the confidence band untouched, because the data is sound. Only its reading would be incoherent. CE6 is the clearest case. The Momentum Index™ can stay HIGH on a strong multi-year trend while the most recent filing year is already declining. The level can legitimately remain CONVERGING SIGNALS at HIGH certainty — but printing « act now » next to that contradiction would be dishonest. So when a grade-neutral flag fires, Stratensight keeps the level and the confidence, both score-driven, then surfaces the contradiction in the « Why this decision » rationale and tempers the recommended action to « strong signal, but reconcile the flagged contradiction before committing ». A confident verdict and a « reconcile first » caveat can coexist, by design: the page never contradicts its own audit.

Why this layer exists: a verdict you cannot audit is a verdict you cannot trust. Stratensight presents the audit before the verdict, not after.

IP Barrier Index™

Composite score measuring IP barrier strength within a technology domain. Quantifies how difficult it is for new entrants to compete based on existing patent portfolios.

FORMULA

IP Barrier = citation_fortress × 0.35 + market_control × 0.35 + remaining_term × 0.30

SCALE

• HIGH ≥ 70: Strong IP barriers — significant portfolio required to compete.
• MEDIUM 40–69: Moderate barriers — selective entry possible.
• LOW < 40: Accessible landscape — limited IP concentration.

Available when n ≥ 50 patents AND Intelligence Grade™ ≥ 60%.

White Space Score™

Measures uncontested innovation zones within technology clusters. Identifies where filing opportunities exist with minimal competitive pressure.

INTERPRETATION

• HIGH ≥ 60: Many open clusters — significant room for new filings.
• MEDIUM 30–59: Mixed landscape — selective opportunities exist.
• LOW < 30: Most clusters contested — limited white space available.

Computed per cluster using diversity, citation pressure, dominance, and freshness. Global score = (open patents / total patents) × 100. Requires minimum 50 patents, Intelligence Grade™ ≥ 60%, micro-clusters < 3 excluded.

Data Requirements & Signal Guards

MINIMUM DATASET SIZE

• Upload mode: ≥ 50 patents recommended for reliable clustering. Below 50, Momentum Index™ is flagged as low-confidence.
• Query mode: ≥ 150 patents recommended. Below 150, the Signals page shows a data quality warning.

TEMPORAL COVERAGE

A minimum span of 3 years is required for Momentum Index™ calculation. Datasets with fewer than 3 years of filing dates receive a temporal bias warning and a degraded confidence score. Lifecycle Position™ needs at least 2 years of span (below that it reports “insufficient temporal data”); a span under1 year drops evidence certainty to VERY_LOW. (Source: score_engine.py.)

MOMENTUM FALLBACK POLICY

• When Momentum cannot be calculated (insufficient data), the score is set to N/A — never estimated.
• Attention levels are automatically downgraded by the Coherence Validator (rule C8) when Momentum is unavailable: CONVERGING SIGNALS → PARTIAL CONVERGENCE, PARTIAL CONVERGENCE → ISOLATED SIGNAL.

INTELLIGENCE GRADE THRESHOLD

A reading of CONVERGING SIGNALS requires an Intelligence Grade™ ≥ 40. Below this threshold it is downgraded to ISOLATED SIGNAL regardless of the other scores.

EPO Open Patent Services (OPS)

The European Patent Office's free API provides structured access to worldwide patent data covering EP, US, PCT/WO, CN, JP, and KR filings with international visibility.

STRENGTHS

• Structured metadata with reliable CPC classifications assigned by patent examiners
• Access to the latest indexed patents (subject to a 3–6 month publication indexing delay)
• Multi-jurisdiction coverage (EP, US, PCT, WO, CN, JP, KR)
• Official source maintained by the European Patent Office

LIMITATIONS

• Results capped at 2,000 patents per query — broad technology domains may be truncated
• Recent filings may have a 3–6 month indexing delay
• Coverage of Chinese patents filed outside the PCT is partial with this source.
• Assignee name normalization is limited — same entity may appear under different names

How Stratensight uses it: Primary source for Explorer analyses. Queries are built using CPC codes for precision. When results approach the 2,000 cap, a warning is displayed and users are encouraged to narrow their query or use Upload mode for complete coverage.

Google Patents BigQuery

Google's worldwide patent database, accessible via BigQuery. It is not queried for technology searches: it serves company searches only, which find the documents published under an applicant name.

Technology searches rely on EPO OPS, and on it alone.

STRENGTHS

• Broadest geographic coverage of any single patent database
• No per-query result cap — full dataset queries possible
• Cross-jurisdiction family deduplication available
• Includes full-text abstracts and claims for most patents

LIMITATIONS

• Indexing delay of approximately 6 months for recent filings
• Assignee names are auto-normalized — precision may vary for large corporate groups and subsidiaries
• CPC codes may lag behind EPO’s official assignments for recently classified patents

Current status: Explore analyses run on EPO OPS. The BigQuery branch is implemented and operational, but a single Explore query scans ~173 GB — above the per-query ceiling of every plan we sell — so it is switched off. Measured complementarity when it did run: +500 patents on a shared window, 93% Chinese.Stratensight names the source it actually uses.

OpenAlex (Fallback only)

OpenAlex is an open-access database of 250M+ academic publications.Stratensight uses it only as a last-resort fallback when patent data from EPO OPS is insufficient for a meaningful analysis.

IMPACT ON SCORES

• Key actors reflect academic institutions instead of companies
• Momentum Index™ may be underestimated (publication cadence differs from patent filing cadence)
• The competitive landscape reflects research activity, not commercial IP positioning
• Lifecycle Position™ is unreliable (academic publication patterns differ from patent filing patterns)
• Openness Score™ reflects academic collaboration, not industrial competition

This is always signaled explicitly in the UI with a non-dismissible amber warning banner. Intelligence Grade™ is automatically reduced when OpenAlex is the primary source.

User-uploaded datasets (Premium)

The most reliable source. Users export patent data from professional databases and upload it directly, bypassing all API limitations.

ADVANTAGES

• Full control over scope and date range — no arbitrary caps
• Reproducible results — same dataset always produces the same scores
• Source is auto-detected and Intelligence Grade™ is adjusted based on known completeness of each export format
• Supports larger datasets than Explorer (no 2,000 patent limit)
• Date-stamped analysis that can be archived and compared over time

COMPATIBLE SOURCES

What is CPC?

CPC is a patent classification system jointly managed by the EPO and USPTO. It assigns one or more technology codes to every patent, organized in a hierarchy:

Sections A through H cover the full range of technology, from human necessities (A) to electricity (H), with Y codes for cross-sectional technologies. Using CPC codes instead of free-text keywords eliminates linguistic ambiguity and provides consistent, language-independent technology mapping across all patent offices worldwide.

How Stratensight maps queries to CPC

When a user enters a technology topic in Explorer, the system maps it to relevant CPC codes through a multi-level fallback strategy. Each level has a different confidence level, and the system always selects the highest-confidence match available.

5 FALLBACK LEVELS

Query Control Levels

Every Explorer query runs at one of three control levels, from fully automatic to full manual control — with rising transparency and user control at each step.

Keyword examples

How real Explorer queries resolve through the fallback ladder, with expected dataset size and noise level.

Why CPC is better than free-text keywords

Patent data coverage varies significantly by jurisdiction. Stratensight’s coverage depends on the data source used and whether filings have international visibility.

Temporal limitations

Patents filed within the last 3–6 months may not yet appear in EPO OPS indexes. This affects recent momentum calculations for fast-moving domains. For time-sensitive analyses, upload a dataset with complete, recent filing data.

The 18-month secrecy period (standard in most jurisdictions) means that the most recent patent applications are inherently invisible to any analytics tool, not just Stratensight.

Data quality limitations

Patent databases contain inconsistencies that affect analysis quality:

• Assignee names may be recorded differently across offices (e.g., “Samsung Electronics Co., Ltd.” vs “Samsung Electronics”)
• CPC codes are updated periodically — older patents may carry outdated classifications
• Abstracts may be missing for some jurisdictions, reducing cluster quality
• Filing date vs publication date discrepancies can affect temporal analysis

Stratensight applies normalization (assignee name matching, family deduplication) to mitigate these issues, but perfect accuracy is not achievable with any automated system.

What Stratensight does NOT do

Stratensight is a patent analytics platform. It is not a substitute for legal, financial, or strategic consulting. Specifically, Stratensight does not provide:

• Commercial valuation of patents or technologies
• Revenue or market size predictions
• Legal advice or freedom-to-operate assessment
• Patent validity or enforceability opinions
• Guarantee of future technology outcomes
• Prior art search or novelty assessment
• Claim-level analysis or infringement detection
• Regulatory or compliance guidance
• Competitor revenue or financial performance data
• Licensing or transaction recommendations

Comparing analyses over time

To track technology evolution, archive your uploaded dataset with its date. Upload the same domain from the same source at different dates to observe score changes. Explorer snapshots should be compared directionally only — exact score differences may reflect index updates rather than real technology shifts.

Algorithm versioning

Stratensight’s scoring algorithms are versioned. When weights or formulas are updated, the version number changes. Analyses run under different algorithm versions are not directly comparable. The algorithm version is recorded with each analysis for traceability.