Demonstration dataset · Synthetic records · Not real security assessments
COLDGUARDResearch Observatory

Methodology · version 0.1-draft

How COLDGUARD builds and qualifies its evidence

Written in the form of a methods section. Every rule here is applied in code, versioned, and cited on the pages it affects.

1. Problem formulation

Temperature controlled infrastructure converts digital signals into physical outcomes: refrigeration setpoints, alarm chains, release decisions, and inventory integrity. A cyber disruption can therefore become a public health or food continuity event. Public discussion of this risk is dominated by anecdote. COLDGUARD builds an auditable record of what has actually been documented, and is equally explicit about what has not.

2. Population and inclusion criteria

An organization is eligible when it operates in at least one of four sectors (cold storage, food storage, blood storage, blood banks), is located in an active research geography, and its identity can be resolved from eligible public sources. Inclusion is editorial and incremental; the included set is a research sample, never a census.

A finding is eligible for publication when it has at least one eligible source of grade A or B (or grade C labeled provisional), passes sensitivity screening, and receives human editorial approval. Grade D material is never counted.

3. Operational definitions

TermDefinition
OrganizationA legal or operating entity with a resolvable public identity.
FacilityA physical site operated by an organization. Facility level data is private research data and is published only as approved, coarse metadata.
Public presenceWhether a public website or online service is documented. Not evidence of operational technology exposure.
IncidentA dated event affecting an organization, documented by eligible sources. Multiple articles about one event are one incident.
FindingA reviewed claim with provenance: organization incident, public disclosure, sector advisory, or regulator notice.
Evidence completenessHow well the research record for an organization is supported (ECI), independent of the content of any finding.

4. Geographic scope

Availability is driven by a typed coverage manifest (2026.11.08-m49-global) that follows the UN M49 standard. North America comprises M49 Northern America, Central America, and the Caribbean. Europe comprises all M49 Europe sub-regions, not only the European Union. South America, Africa, Asia, and Oceania follow their M49 sub-regions. Every continent except Antarctica is active; activation is a navigation property and says nothing about evidence volume. Transcontinental policy: Russia is European; Turkey, Cyprus, Georgia, Armenia, Azerbaijan, and Kazakhstan are Asian; Greenland is North American; the French Southern Territories are filed under Eastern Africa. Kosovo, Taiwan, and the State of Palestine are listed as territories because they are not M49 member states and the basemap renders each as a unit; the basemap units Northern Cyprus and Somaliland are routed with Cyprus and Somalia. Boundaries are cartographic and express no political position.

Geographic functionality is independent of data availability. Every active country page exists, including countries with zero reviewed records, which display an honest empty state rather than a score.

5. Source registry and collection windows

Each source records license, retrieval method, permitted uses, attribution terms, URL, source date, first seen, last checked, and publication eligibility. Adapters are disabled unless the source permits the intended use. No logged-in services are scraped and no personal information is harvested. Vulnerability catalogs are imported as sector context only and never auto-linked to organizations.

6. Entity resolution and deduplication

Organizations are linked by stable public identifiers where permitted; fuzzy name matching is conservative and ambiguous matches require manual confirmation. Source article, incident event, organization, and finding are distinct records. Event, publication, and verification dates are stored separately. Corrections use supersedes_finding_id and retracted_at; retracted records stay visible as notices and are excluded from all counts.

7. Evidence quality rubric

GradeDescriptionEligibility
APrimary official disclosure, regulator report, organization notice, or direct confirmed documentationEligible after editorial approval
BMultiple independent credible reports corroborating the same eventEligible after editorial approval
CSingle secondary report with incomplete confirmationAggregate only or labeled provisional
DUnverified posts, rumors, model outputs, ambiguous matchesExcluded from counts

Source grade and entity attribution confidence are separate fields. A reliable advisory can carry zero support for attributing a vulnerability to a particular organization.

8. Measures and missingness

Public Finding Count (PFC)

Distinct, eligible, reviewed findings for a region, sector, and period, always shown with its denominator.

Evidence Coverage Index (ECI, eci-0.1)

ECI = 100 × Σ(weight × component) / Σ(weight) over assessed components. Weights: identityCorroboration 25, sourceTraceability 25, reviewRecency 20, sectorClassificationQuality 15, jurisdictionCompleteness 15. Null components are excluded from numerator and denominator and the metric is flagged incomplete; they never score as zero. Weights are a versioned research decision.

Published Signal Burden (PSB)

PSB = 100 × unique reviewed findings / included organizations. Suppressed when the denominator is below 10. It is not a population incident rate.

No composite security score

Version 1 publishes no single security score or country safety ranking. Public evidence does not support that precision.

9. Geographic comparison and coverage bias

Data scarcity, national reporting differences, organizational size, indexing, language, and publication practices produce large ascertainment biases. A country with more published findings may simply be more transparent or better researched. Charts state selection criteria, n, vintage, and missingness. Cross-country statistical comparison is disabled when source coverage differs materially, rather than decorated with false significance.

10. Ethics and disclosure

No live scanning, device fingerprinting, or exploit validation. No publication of IP addresses, ports, endpoints, precise coordinates tied to unresolved weaknesses, credentials, or exploit steps, even when an external source contains them. No personal data of patients, donors, or employees. Named-entity claims require second-person review for high consequence allegations. Every record carries a correction path.

11. Limitations and replication

Threats to validity include sample incompleteness, reporting delay, language coverage, and reviewer variability. Reviewer agreement statistics are computed only when a sufficient double-coded sample exists; none is reported until then. The replication package is the versioned dataset, this methodology, the coverage manifest, and the code that applies them.