Engine

What this is, and how it computes

The whole path from a chain reading to a row on the board, in the order it runs. Everything further down this page is a grade on one of these six steps.

Vantias reads every live Pendle market on 8 chains, keeps its own dated history of them, and says what a fixed-yield buyer would actually receive at a size they choose -- after the cost of getting in, after the cost of getting out, and with the risk that is being taken for it named rather than scored out of ten.

It is two processes over one PostgreSQL database. A worker sweeps every 15 minutes whether or not anybody has the page open, and writes what it read with the moment it read it; the page only reads what the worker has already written. Nothing on the board is computed while you wait for it, which is why a figure can always be asked when it was taken.

1. 1
   What comes in

   Three sources, and the browser is none of them. Pendle's hosted index gives the market list, the prices and the two rates Pendle itself publishes. Each chain's own contracts, read directly, give the wrapper, its exchange rate, and who is permitted to change either. Pendle's router gives a quote at a real size. Each reading is stored with its timestamp and its source; nothing is averaged before it is stored.

   measured
2. 2
   The fixed side

   A PT is a bond: bought below par, redeemed at par, and the rate is entirely in that discount over the days remaining -- (1+i)^(-t/365). That is the definition of an annualised rate, not a model of how Pendle prices anything, which is why it is recomputed here instead of copied. Pendle's own published rate is then checked against it, and a market whose two figures disagree by more than 0.30 points of yield is rejected rather than shown with a number nobody can defend.

   derived
3. 3
   The floating side -- the guess the ranking rests on

   What the yield averages from here has not happened yet, and every YT figure depends on it. The estimator is today's reported rate, clamped to 2x the 30-day median: on the stored history the spot reading is the more accurate one on a typical day and much the worse in its tail, and the leash is there to keep the first without the second. Further down this page both ingredients are scored against what followed, and that section says whether the rule itself has been scored yet. Where an asset reports a rate the market will not pay for -- 4x the traded fixed rate, above 25%, and only when the 7-day median also agrees past 2.5x -- every figure built on that rate is withheld and says so, instead of being printed in the same typeface as one that was checked.

   estimated
4. 4
   What it costs to get in, and out

   Every yield on the board is net at the size in the box, not at zero size. Entry is a router quote at that size, with fee and price impact taken together as one payment per transaction rather than as a rate -- a large buy that arrives in two pieces is still one payment. The exit is a second quote at the size the holder actually has, and a market the router refuses at that size costs the position points for it instead of being scored as though leaving were free.

   measured
5. 5
   The risk band

   6 layers, each scored on its own evidence and each allowed to disagree with the other five. The band is the worst layer, never their average, and the card names which layer set it. There is deliberately no single number out of a hundred: a score that blends a redemption path with a governance key hides exactly the thing worth reading.

   estimated
6. 6
   What is kept, and what is rejected

   A market is rejected when its pool is under $100,000, when its net yield over the whole term is under 0.3%, when its risk points pass 50, when its composite risk band is worse than high risk, or when the position would be large against the pool itself. A rejected market is not deleted -- it stays in the same list, carrying the rule that rejected it, so it can be sorted beside the rest and argued with. Inside a section the order is absolute net yield rather than the annual percentage, because at a five-day maturity an annual rate is an extrapolation of a rounding error. Between sections there is no order at all: a PT yield and an LP APY are not the same unit and will not be stacked as if they were.

   derived

The six layers, in the order the card shows them: yield and structure, exit, control, contract, denomination, position.

Two of these six steps end in a claim rather than a reading: what the floating rate will average from here, and where the risk band falls. The first is scored against what actually followed, and that record is printed below. The second has never been scored at all, and this page says so in its own section rather than leaving it to be found. Everything else on the board is either somebody else's reading copied with the moment it was taken, or arithmetic over those readings with no free parameter in it.

Which of these numbers can be wrong

A card puts a chain reading, an arithmetic identity and an unscored model in front of you in the same typeface. Only one of the three can be wrong on its own.

• 13 measuredSomebody else's reading, copied: a rate off a contract, the size a router quotes. Wrong only if the source is.
• 6 derivedArithmetic over measured inputs with no free parameter. PT plus YT is one unit; there is no judgement to get wrong.
• 13 estimatedA guess about something not observed: a rate that has not happened yet, a price at a size nobody traded. The only class worth scoring.

The estimates of one rate, against what followed

Buying YT pays the fixed rate to receive the floating one, so the only question is what the floating rate averages from here. These are the ways to estimate it, and how each has done.

62,511 scored days across 492 maturity cycles, over a 30-day horizon. The error is in percentage points of yield; lower is better. · measured 07 Nov 2026

The 62,511 scored days are not 62,511 separate tests: each reads 30 days either side of its day, so neighbours share almost every reading. 1,353 of them share none. They run from 15 Apr 2023 to 28 Oct 2026; the brackets hold the typical error in 95% of 200 redraws of whole cycles.

Today's reading was closer on 53% of the 43,472 days that separated them; the median on 47%.

On the typical day today's reading missed by 0.141 points and the median by 0.165; at the extreme, 15,416.4 against 3,600.5. The estimator in use keeps today's reading on a leash of 2x the median, to take the typical-day accuracy of one without the tail of the other. Scored on the same days, the leash itself missed by 0.142 points on the typical day and 3,600.8 at the extreme.

Where the dispute threshold actually sits

When an asset reports a yield the market will not pay for, every figure built on that rate is withheld. This is the line that decides, re-derived from the accumulated history.

540,072 observations across 408 maturity cycles: the ratio of the yield an asset reports to the fixed rate the market trades it at. · measured 07 Nov 2026

0.783% of all observations sit above the threshold in force.

61 of 408 cycles have ever been above the line and 0 stay above it, so this is a state a market enters and leaves rather than a property it has. That is why the hold band below the threshold is load bearing.

How the risk band is put together

Six readings, one band. The rule that combines them is the whole of it.

Each layer is scored on its own evidence and can disagree with the others. The card names which one set the band.

1. yield and structureWhere the yield comes from, and where this token stands when whatever produces it fails - senior, junior, or unstated. An asset with no checked dossier is never “low” here: an unverified claim is not a verified one.
2. exitThe way out, and how long it takes. A redemption wait counts against a position only when the router will not sell the token at your size - a seven-day unstaking wait costs nothing if the market takes you out today.
3. controlWho can change the contract holding the money, read from the chain: code that cannot be replaced, a timelock standing in front of an upgrade, a multisig, or one plain key. A controller that could not be classified is the middle, and says so.
4. contractThe token itself: the contract that holds the money, how it has been audited, whether its price has moved away from what it is meant to track, and whether the protocol behind it has ever been hacked. A confirmed hack or a broken peg sets a floor this layer cannot go below.
5. denominationWhat one token is worth in. A dollar token and a token priced in ETH are different risks to a book measured in dollars, and anything unresearched is treated as the middle rather than as a dollar.
6. positionThis card only: what YOUR size does. It is read at the amount in the box at the top of the page - the same market is a different risk at $5,000 and at $200,000.

The composite is the WORST layer, never an average. An average lets five clean layers hide one that is not, and on a risk reading that is the only direction that matters.

What choosing was worth, over every matured cycle

A rule that picks the highest rate on offer against one that takes everything. The answer is about SELECTION, not about money received -- and the reason for that distinction is the second line below.

418 matured cycles, each entered 30 days before its expiry at the rate the market was showing that day.

WHAT THIS CANNOT SAY: what the position was worth in dollars. A PT redeems at par, so its return in the underlying IS the rate locked -- a tautology. The question that is not a tautology is what happened to the asset underneath, and the price history cannot answer it: 712,367 readings carry a rate, 649,604 carry a price and only since 09 Jul 2026, and of the matured entries that would need both, 83,393 have both. For a cycle that has already matured the price appears on one row, written by a backfill months after expiry.

Picking the highest rate on offer locked 4.9 points more than taking everything, at the median, over 418 cycles. That is what choosing was worth in RATE. What it was worth in money depends on what the assets did, which this service will be able to say for cycles maturing from here on and cannot say for these.

The band against what happened

The claim this service is least able to defend, said here rather than left to be found.

Every day, what the page said about a market is written down, and the outcome is filled in once the horizon has passed. A band scored against an outcome is the only thing that can turn it from a hypothesis into a measurement.

Until the bands can be graded -- at least 3 bands each carrying 30 settled verdicts from at least 10 different markets -- every risk band on this service is a hypothesis with a colour on it. It is built from measured layers and it is applied consistently, and neither of those is the same as having been right. Bands with the days but not the markets: low risk, extreme risk -- fewer than 10 different markets each, so they do not count yet.

The month after, beside the week: the lowest premium recorded in the 30 days after each verdict, below fair value when negative. Shown, not graded; it still cannot see an exit that closes while the price stands still.

And the door in the same month: how many verdicts saw the router refuse to sell $10,000 of PT for that market on any of the 30 days that start on the verdict's day. A refusal is the router answering no; a rate limit or a router that could not be reached is not one. A verdict whose market the router did not answer in its month is counted apart, as not answered, never as an open door. Shown, not graded. Counted by market and, beside it, by maturity: each verdict and each maturity the router answered for in its month is one door, so a refusal on a thin second maturity does not close the first one. The median of answered days says how much of the month each door was actually asked.

What this measure cannot see: a market with no way out barely trades, so its price hardly moves, and the band for no exit reads calmest. A trapped exit is the loss that matters there, and a week of price movement does not count it.

Whether it is getting better

The same scores, kept dated, so a change can be argued about afterwards rather than asserted.

One reading per estimate per day. Lower is better -- the figure is an error in percentage points.