EVIDENCE CERTAINTY

GRADE Methodology — How Stratensight Assesses Evidence Certainty

Why we grade every decision’s evidence level before issuing a strategic recommendation.

GRADE (Grading of Recommendations Assessment, Development and Evaluation) is a transparent system for rating the certainty of evidence and the strength of recommendations. Stratensight applies GRADE to patent intelligence so that every verdict carries an auditable certainty label, not just a score.

Two questions sit behind every recommendation: how much can we trust the underlying signal? and how confidently should we recommend acting on it? GRADE is the discipline of answering each separately and disclosing both.

02 / FRAMEWORK

The GRADE framework in brief

GRADE was developed by the GRADE Working Group, an international collaboration of methodologists, clinicians, and statisticians. It is the standard adopted by the World Health Organization (WHO), the National Institute for Health and Care Excellence (NICE), UpToDate, and Cochrane for grading clinical evidence.

Core principle

GRADE separates two concepts that are routinely conflated:

Evidence certainty — how much we trust the underlying signal.
Recommendation strength — how confidently we recommend acting on it.

A recommendation can be strong even when certainty is moderate (when downsides of inaction outweigh uncertainty), and conditional even when certainty is high (when context-specific tradeoffs dominate).

gradeworkinggroup.org →

03 / DOWNGRADE FACTORS

The five downgrade factors

Stratensight starts every analysis at HIGH certainty and applies up to two-level downgrades per factor. The floor is VERY_LOW. Each factor has a documented threshold in the codebase — we publish what we measure today, and we publish what we have not yet covered.

01Factor 01 — Risk of Bias

HOW STRATENSIGHT MEASURES IT

OpenAlex fallback ratio. When more than 20% of patent records required OpenAlex fallback to recover bibliographic metadata, we apply a minor downgrade. When the ratio exceeds 50%, a major downgrade.

NOT YET COVERED

Citation completeness, partial patent records. On our research roadmap.

02Factor 02 — Imprecision

HOW STRATENSIGHT MEASURES IT

Dataset size. Below 150 patents triggers a minor downgrade (sample may be too narrow for stable clustering). Below 50 patents triggers a major downgrade.

NOT YET COVERED

Confidence intervals on individual metrics. On our research roadmap.

03Factor 03 — Inconsistency

HOW STRATENSIGHT MEASURES IT

One internal coherence signal.

Temporal bunching: filing ratio concentrated at 80% or more in a single year window suggests artefactual bursts rather than sustained innovation (minor downgrade).

NOT YET COVERED

Cluster separation. The silhouette score is measured and displayed — it qualifies the clusters — but it deliberately does not downgrade the Grade. The same dataset can produce a different silhouette on a different machine, and a measure that is not reproducible must not decide a verdict that has to be. Also not covered: cross-metric contradiction detection (e.g. high momentum but low maturity). On our research roadmap.

04Factor 04 — Indirectness

HOW STRATENSIGHT MEASURES IT

Recognised-territory match. We maintain a curated reference set of technology domains. When the analysis domain is in this set, the signal keeps full confidence; when it falls outside, a deterministic downgrade is applied to reflect lower confidence in extrapolation. Continuous empirical validation of this territory is on our research roadmap.

NOT YET COVERED

CPC scope breadth, geographical representativeness. On our research roadmap.

05Factor 05 — Data Integrity

HOW STRATENSIGHT MEASURES IT

Integrity signals surfaced by our Critical Reader module — single-CPC artifacts and data-completeness gaps (partial metadata, last-year filing artifact), not academic publication bias. Structural defects apply a firm downgrade; data-shape warnings a softer, capped one. Above 1 critical issue or 2 warnings, minor downgrade. Above 3 critical issues or 4 warnings, major.

NOT YET COVERED

Academic vs industrial filing ratio as a standalone signal. On our research roadmap.

04 / MATRIX

Certainty × Strength matrix

Once the five factors have been applied, the resulting evidence certainty is paired with a recommendation strength. The combination shapes the verdict the user sees on every analysis.

EVIDENCE CERTAINTY	RECOMMENDATION STRENGTH	INTERPRETATION
HIGH	STRONG	Act with confidence
MODERATE	STRONG	Act with standard due diligence
LOW	CONDITIONAL	Act only after validation
VERY_LOW	CONDITIONAL	Do not act on this signal alone

05 / APPLICATION

How Stratensight applies GRADE

GRADE assessment runs automatically inside the analysis engine. No analyst intervention. The pipeline is auditable end to end.

Certainty starts at HIGH and is downgraded factor by factor. The floor is VERY_LOW; no upgrade factors are applied in the current version.
Each downgrade threshold is a named constant in grade_assessor.py and is covered by anti-regression tests. Changes go through code review, not configuration drift.
Recommendation strength becomes CONDITIONAL automatically when certainty is LOW or VERY_LOW, or when any decision criterion sits within five points of its threshold.
Every assessment is exposed verbatim on the verdict page (Certainty Assessment panel) — starting level, final level, and the explicit list of downgrades applied.

Independent audit report — in preparation.

HOW GRADE FEEDS LAYER C

The final GRADE level (HIGH / MODERATE / LOW / VERY_LOW) drives the Tier gate (Layer C) directly — and what it changes on screen is narrower than the machinery behind it. HIGH, MODERATE and LOW all display the same attention level: the gate substitutes an internal code at MODERATE and LOW, but both resolve to the wording the reader already saw at HIGH. What LOW adds is an explicit qualifier, — Conditional, pending Signal Integrity resolution. Only VERY_LOW changes the words: the input is genuinely unusable (n < 10, temporal < 1 year, or corrupted), no level is issued, and the page reads INSUFFICIENT EVIDENCE.

This closes the loop: evidence_certainty is not just a metadata badge — it gates the verdict surface itself. See the methodology page for the Layer C thresholds and what each tier displays.

06 / REFERENCES

Scientific references

The GRADE framework is widely documented. The five sources below are entry points for readers who want to verify our application against the canonical literature.

Sibling methodology page

Evidence certainty meets the wider epistemic contract behind every Stratensight verdict.

/methodology/grounding →

Four-layer epistemic contract (hedging, grounding, refusal, source tagging) plus the eighteen user-facing tokens that appear across decision narratives and persona insights.

Stratensight provides patent intelligence signals, not legal opinions or freedom-to-operate assessments. Not a substitute for IP counsel.