Evidence before volume.
A source is a starting point. Our job is to show what it supports, where it stops, and when it was last checked.
Facts have an address
Every structured specification, commercial statement and benchmark result is attached to a source. Open a source marker to see the publisher, accessed date and evidence label. A manufacturer specification describes a published specification; it is not an independent performance test. If the reviewed public record does not support a value, we display “Not publicly disclosed”.
The publication standard
An entity enters the indexed database only after identity verification, a substantive original description, meaningful sourced fields, a source ledger, a review date and useful internal links. We normally require multiple credible sources. A documented exception allows a sufficiently authoritative primary source for a narrowly scoped record. An unresolved identity or claim conflict keeps the page out of search indexing. The research queue is not the published database.
Evidence labels describe origin
Primary spec and primary release identify manufacturer or developer documentation. Customer confirmed means a named customer acknowledges the particular activity. Independent report identifies an outside source, not an automatic proof of every underlying claim. Company claim is a vendor statement. Live demo and edited demo are demonstrations with their own limits. Conflicting sources means the disagreement remains unresolved.
Deployment maturity has to be earned
We distinguish announcements, prototypes, demonstrations, pilots, customer-confirmed pilots, paid deployments, recurring operation and production scale. A commercial agreement can establish a relationship without establishing installed fleet size or continuous autonomous operation. A customer report from an earlier product revision does not transfer to a newer version. Historical evidence stays dated.
Comparisons answer a specific decision
Only curated matchups receive comparison pages. Each declares a use case, exact version scope, review date, meaningful decision dimensions and evidence gaps. We show availability, task fit, integrations, cost and deployment evidence without combining them into a universal winner. Editorial decision paths are questions to investigate, not vendor endorsements.
Benchmarks stay in their own context
A result needs an original source, model revision, metric, environment, reported date and methodology limitations. Embodiment is recorded when disclosed. Simulation and real-robot results stay distinct. Comparability groups identify matching protocols; even a shared benchmark name is insufficient if task splits, tuning, observation inputs or evaluation budgets differ. No global Physical AI score is calculated.
Freshness, conflicts and corrections
Price and availability claims enter review after 30 days. Production status receives a 60-day window; general entity identity receives 90 days; evergreen research receives 180 days. An overdue claim is visible and prevents continued index eligibility until review. Corrections enter an editorial queue with their proposed value and source; submissions never change a public record automatically.
What the database cannot establish
Public-source research is not a physical product test, safety certification, legal assessment or guarantee of availability. We do not operate an independent robot testing laboratory. Buyers should confirm current terms and validate their own task and environment with the manufacturer or integrator.