How this surface is built

A pair is (collateral asset, debt denomination). BTC/USD means Bitcoin-representation collateral against dollar-denominated debt. Nothing ever aggregates across pairs. A tab is a filter over one schema, not a separate dataset.

The partition is by what the debt is denominated in, not by what the borrower intended. Debt in a euro stablecoin against Bitcoin collateral is BTC/EUR credit whether or not the borrower thinks in euros, because its risk decomposition differs: the position carries a currency basis that a dollar loan does not.

Denomination is resolved from a mapping of each debt asset to its currency, with a recorded valuation method — observed where the asset has its own market price, parity where a stablecoin is valued at par to its peg currency. Par is a documented convention, not a measurement, and where an asset behind a figure is outside parity monitoring the surface says so on the tab.

Currently partitioned: BTC/USD · BTC/ETH · BTC/EUR · BTC/BTC. Debt whose denomination cannot be resolved belongs to no pair and is excluded from every total rather than absorbed into the largest one; the count of excluded rows is published on each tab.

Every rate on this surface is annualised and stated as APR on an act/365 basis. Venues quote in at least four different ways — APR, APY, per-second compounding factors, and per-interval funding — and comparing them unconverted is the first error an analyst notices.

Where a venue publishes a different convention, the conversion is applied here and the venue's native quote is shown alongside on its own page, so a reader checking our figure against the venue's own site can see both and reconcile them.

This is the distinction most easily lost and the one that changes conclusions.

Tenor is how long the money is out for. It can be contracted — set at origination, as with a fixed-term loan — or realised, observed after the fact as how long a position actually persisted. Both answer "how long", one ex ante and one ex post, and marks on the term structure distinguish them.

Lookback is something else entirely: the window over which an average is computed. A 7-day trailing average of an overnight rate has a tenor of zero and a lookback of seven days. Reading that as a seven-day rate is the error.

So a tenor axis and a lookback axis never appear on the same chart, and the dashboard separates them physically rather than by footnote. Where a figure carries both — the funding rate does — the label states each.

The realised curve is built from what borrowers actually paid, over realised holding periods. It is a measurement, not an offer. A point on it says "borrowing for about this long has cost about this much", never "you can lock this rate for this long" — most of the venues behind it issue no fixed-term credit at all, and their rates float for the life of the position.

A curve is only drawn where there is enough behind it to draw one. Below that bar the surface shows the prints and says the bar is unmet, rather than fitting a line through too few points and letting the smoothness imply confidence the data does not carry.

Spans that are extrapolated, modelled or reconstructed render dashed rather than solid, and the count behind every surface cell is published on the cell.

Curve shape: one classification, and what it measures

The curve's shape — normal, inverted, humped or flat — is read from the curvature of the fitted curve: the sign and size of the coefficients in the log-quadratic regression through the bin means. It is a property of the fit, so it answers "what shape did the model find in this data".

The surface previously also published a second classification, read from the observed bin endpoints — comparing the short end against the long end directly, without reference to the fit. That measure answers a different question: "which way do the two ends point". The two can disagree about the same data, and did. A curve whose endpoints rise while the fitted curvature turns down is not a contradiction in the market; it is two questions with two correct answers.

Only the fitted classification is published. Two contradictory public labels on one fit is a defect, not a richer picture: a reader cannot act on both, and nothing on the page told them which to prefer. The endpoint comparison is retired from this surface and remains queryable — it is not deleted, and no history is discarded.

The change of answer is registered rather than smoothed. Where a reader previously saw the endpoint reading on this page, they now see the fitted one, and the two disagreed on the date recorded in the break register below.

Every aggregate carries a coverage percentage: how much of the venue universe contributed to it. An aggregate without one is a number whose denominator is hidden.

Three separate questions sit behind every venue, and reading any one alone misleads. Coverage level says what a venue's data can support once ingestion runs. Status says whether anything is ingested yet. Validation says whether what was ingested has passed its verification gate. Only the third gates publication: an unvalidated series shows nothing, however complete its coverage and however advanced its ingestion. A published wrong number is not fixed by validating it afterwards.

Known discontinuities in the underlying data. Breaks are annotated, never smoothed, and no chart joins a series across one without saying so.

Where documentation and chain disagree, the chain wins and the disagreement is recorded.

The venue attribute matrix is attributes, not scores. Each cell is a factual structural property with the contract, document or filing establishing it. Nothing is ranked, weighted or recommended, and no composite index is published: a weighted composite is editorial by construction, however well documented.

Where a venue does not publish a property, the matrix says undisclosed — distinct from not applicable (the question does not apply to that venue class) and from an empty cell (nobody has established it). Filling an unpublished cell with the industry convention would turn a matrix of sourced facts into a matrix of assumptions about named third parties.