At Kairos Signal, our topological mapping of Decentralized Physical Infrastructure Networks (DePIN) has cataloged 745 distinct networks operating across varying L1s and L2s. When we evaluate the empirical observability of these networks, the landscape fractures violently: 263 networks carry live, queryable state data, and a mere 171 expose first-party supply telemetry (e.g., geolocation, hardware status, uptime proofs). The remaining 482 networks—roughly 64.5% of the ecosystem—have no free public feed.

In traditional data engineering, the instinct when faced with a missing variable $X_{miss}$ is to impute. One might model the expected supply based on tokenomics, market cap, or social sentiment. This is an epistemic fallacy. In spatial-temporal infrastructure modeling, naming what you do not cover is fundamentally more honest, and mathematically more rigorous, than generating estimates from uncorrelated priors.

This post explores the architecture of the not_covered map—a topological surface of known unknowns—and why researchers must adopt it to understand where physical data infrastructure is genuinely missing versus merely dark.

The Epistemic Boundary of DePIN Observability

Let $S$ be the universal set of DePIN networks. We can partition $S$ into three disjoint subsets based on telemetry availability:

$$ S = S_{live} \cup S_{telemetry} \cup S_{dark} $$

Where:

When a quant analyst attempts to model the global supply of, say, GPU compute or IoT sensors across $S$, they inevitably encounter $S_{dark}$. The immediate temptation is to apply an Expectation-Maximization (EM) algorithm or multiple imputation to fill the $S_{dark}$ matrix.

This is a catastrophic failure of modeling.

Missing Not At Random (MNAR)

Data is typically categorized as Missing Completely At Random (MCAR), Missing At Random (MAR), or Missing Not At Random (MNAR). DePIN telemetry for $S_{dark}$ is overwhelmingly MNAR. A network obscures its telemetry because its supply is deficient, its node distribution is sybil-attacked, or its