Urd Atlas reads raw blockchain data every day, runs it through a calculation pipeline, and publishes an answer to one question: does this chain still look normal, or is something meaningfully changing?
Raw data gives you numbers. It does not tell you whether those numbers are high, low, normal, or important for that specific chain right now.
You can read raw chain data for free in explorers and dashboards. What you usually do not get is a daily, documented answer to whether the current move still looks like noise or has become structural.
Noise is short-lived movement in on-chain data that looks dramatic in the moment but does not last and does not represent a lasting change in the chain's state.
A real regime change is a persistent shift in the chain's descriptive state. It is not just one unusual print; it is a situation where demand, friction, or capacity has changed enough, and long enough, to justify a named published state.
A spike is a short shock in the daily data. A real change is something that remains visible after the data is smoothed and compared with the chain's own recent baseline.
Regime is a compact description of the chain's current operating state relative to its own recent history. It is not a prediction, a valuation, or a recommendation.
It means the product describes what is happening now rather than what will happen next. A state label tells you how the chain looks, not what price will do.
STABLE means the chain looks close to its normal recent operating range. HEATING means pressure is building. CONGESTED means conditions look materially tighter or costlier than usual. CHEAP means conditions look looser or cheaper than usual.
Because the product would rather say 'not enough evidence' than publish a confident-sounding answer built on weak data.
Confidence is a 0 to 1 measure of how well-supported the published reading is by the available data. Higher means stronger evidence quality. Lower means the evidence base is weaker.
A score like 0.847 means the current row is backed by relatively complete, recent, and historically sufficient data. A score like 0.675 still clears the gate, but the evidence is weaker and should be read a bit more carefully.
It means the product has a minimum evidence threshold for publishing a named regime. If the row does not clear that bar, the product withholds the label instead of pretending the answer is reliable.
Because the product needs a documented minimum evidence level before it will publish a normal label. Under that level the safer answer is to degrade rather than guess.
The product uses public on-chain transaction and block data from AWS Public Blockchain Data for Bitcoin, Ethereum, Arbitrum, and Base.
Urd Atlas is generally scheduled to publish updated artifacts around 09:00 and 21:00 Europe/Oslo.
Baseline means the chain's own recent history. The product does not compare Bitcoin to Ethereum or compare today's numbers to some universal market norm.
Today's values are compared against the same chain's own recent historical distribution. The point is to know whether the current move is unusual for that chain, not whether it is large in the abstract.
The window lengths are the same, but the data inside them is chain-specific. Bitcoin is compared to Bitcoin history, Ethereum to Ethereum history, and so on.
Because they are not the same kind of system. Ethereum has gas utilization and failed transactions in a way Bitcoin does not. Bitcoin has different capacity and fee mechanics.
Normal z-scores get distorted by extreme outliers. On-chain data has a lot of those. A robust z-score uses the median and median absolute deviation so a single whale day or fee shock does not bend the whole baseline out of shape.
Coverage factor tells you how much of the expected evidence was actually present for that dimension. If too much data is missing, the score is pulled back toward neutral.
Gold is the raw daily observation layer. Meta is the analytical interpretation layer. Derived is the smoothed trend layer built from Gold.
Mostly Meta. That is where the actual analytical output lives. Gold and Derived matter because they let you verify, compare, and build around the same state layer.
The label is the short answer. The scorecard dimensions are the structured explanation of what is driving that answer.
Meta saves you from implementing the difficult middle layer yourself: normalization, confidence, score construction, and explainable driver output.
In the current version, Derived is primarily rolling trend context built from Gold plus a confidence carry-through field useful for chart overlays.
Seven days is long enough to reduce single-day noise but short enough to react to actual change. Thirty days gives a broader medium-term reference.
The free surface lets you inspect the published state. Single Chain gives you the actual reference data JSON for one chain so you can use it in your own tools and models.
Research expands from one chain to all four and gives you a deeper history window. It is for users who need cross-chain context or broader research coverage.
Because the site is for reading and inspecting. The subscription is for using the artifacts directly in your own workflow.
It means the product does not tell you what price will do and does not use price data in its core labels. It describes network conditions, not market direction.
Use it when you need structured context on whether current on-chain conditions still look normal, are heating up, are congested, or are unusually cheap.
Primarily for analytically oriented users who are comfortable with data and want reusable on-chain reference data JSON, not just visual dashboards.
No for the public site. Yes, at least a little, if you want to get the full value from the subscriber reference data JSON.
It is a fingerprint of the published analytical row. It exists so the output can be checked for reproducibility rather than treated as a soft editorial opinion.
It is a traceability field that helps identify a specific published historical row or revision context.
Corrections are possible in principle, but the product is designed so that changes are not silent. Versioning and traceability exist precisely so users can detect meaningful changes over time.
Because it is meant to show what the system actually published, including weak-evidence days, not just the days that make the product look good.
Because a chain-state product should show whether its own published rows are fresh enough and trustworthy enough to read with confidence.
Because the product treats those chains with a different freshness policy rather than pretending their source data behaves exactly like L1 data.
At a high level, the final regime label is rule-based. The underlying evidence layers include scores and normalized signals, but the published label is produced by deterministic rules rather than a black-box model.
They are a ranked explanation layer: the product is surfacing the metrics that most help explain the current published reading.
It usually means the product did not have strong enough or coherent enough evidence to surface meaningful drivers for that row.
Above the gate, the product publishes a normal named regime when the rule conditions are met. Below the gate, it withholds the normal label and publishes UNKNOWN/DEGRADED instead.
The hash is built from the key state output and the information needed to reproduce it consistently.
Because the product does the hard part between raw data and usable state: aggregation, normalization, confidence gating, explanatory ranking, and artifact publishing.
You are accepting the product's definitions of baseline, persistence, confidence, and regime thresholds instead of defining all of those yourself.
Ninety days is enough for immediate context. A year or more becomes much more useful for serious historical analysis and regime-conditioned research.
