Methodology

How an entity enters the graph

An entity begins as a candidate found in public records, company material, procurement and contract data, reputable reporting, or a source submitted for review. We resolve its identity, attach the source and retrieval date, normalize the record, and link it to products, people, deals, capabilities and deployments when the evidence supports those connections.

A company can enter the directory before every category has been researched. Sparse records are published as sparse records: missing work remains NOT_ASSESSED, while an active search that produced no reliable result is recorded separately as SEARCHED_NOT_FOUND. Entity names and slugs are deduplicated against the existing graph; merges preserve the surviving identity and redirect old references.

Evidence language

These labels describe how strongly a displayed claim is supported. They do not describe how complete the whole dossier is.

CONFIRMED

Supported by a primary record or by multiple independent sources that agree.

REPORTED

Attributed to a named source, but not independently confirmed by us.

INFERRED

Our analytical conclusion from cited facts. The evidence and the reasoning should be visible.

Assessment states

Assessment state answers a different question: what happened when we evaluated a fact or category? Absence and unfinished research are never collapsed into the same empty value.

SUPPORTED

We found evidence that supports the assertion.

SEARCHED_NOT_FOUND

We looked for the information and found no reliable public evidence as of the stated date. This is negative evidence, not a claim that the thing does not exist.

NOT_APPLICABLE

The question does not apply to this entity or category.

NOT_ASSESSED

We have not yet completed a search. It must not be read as “none.”

CONFLICTED

Credible sources disagree. We retain the conflict instead of silently choosing one account.

Coverage estimate

A dossier’s coverage signature shows four separate axes. The composite is versioned as v2-20-30-25-25 and is calculated as:

0.20 × Identity + 0.30 × Capability + 0.25 × Relationships + 0.25 × Scale

Evidence and analysis are deliberately not terms in the composite. Evidence quality belongs beside the assertions in the trustline; analysis is a publication state. The composite may support ordering and corpus statistics, but it is never an entity headline.

Why there is no headline percentage

A prior single-score model gave both ANYbotics and NEWCOM Wireless Services a score of 68. That equality hid two very different records.

One percentage made them look identical. The four-cell signature plus trustline shows the useful difference: what is well documented, what is thin, how the evidence is sourced, and whether analysis exists.

Sources and trustline

The trustline is separate from coverage. It reports the number of source records attached to the company, the last verified date, the next scheduled review when available, and whether provenance backfill is still in progress. Individual facts carry their own source, date, confidence and stable citation as the fact layer is populated.

During the current company-level provenance phase we do not claim “N of M facts sourced.” A company can have sources without every displayed assertion having a source link. That ratio will appear only when per-fact provenance is available.

Refresh cadence

Different facts decay at different rates, so there is no blanket freshness promise. These are the review triggers and target sweep cadences; a material signal can prompt an earlier update. Each dossier’s dates are more specific than this general schedule.

Corrections

If a fact is wrong, incomplete or stale, email editor@robotics.press. Include the dossier URL, the fact in question, the proposed correction and a primary source where possible. You do not need an account.

We check the submission against the cited and existing sources. Accepted changes update the structured record and its verification date. If credible evidence remains in conflict, we mark it CONFLICTED rather than erase the disagreement.

AI-assisted publishing

AI systems are used throughout robotics.press. They monitor public sources, identify candidate entities and events, extract fields, normalize names, propose graph links, compare sources, and draft summaries and analysis from the structured graph. Generated text is a rendering and interpretation of that evidence; the AI system is not itself a source.

What is checked

• —Required fields, identifiers, dates, units and links are checked against the data contract.
• —Claims retain source attribution and retrieval dates; multiple sources are compared where available.
• —Conflicting accounts are flagged for retention or review instead of being averaged into false certainty.
• —Analytical statements are separated from reported facts and use explicit confidence language.

What is not guaranteed

Not every published claim is independently confirmed, and not every page is reviewed by a person before publication. Public sources can be wrong or become stale; extraction and entity matching can also fail. The evidence labels, assessment states, dated trustline and correction path are how those limits remain visible and repairable.