DIVERGENCE

A price can be wrong
while every venue is honest

Most of this venue's machinery answers one question: how much do the places that quote a thing agree about what it costs right now. When they stop agreeing, the number in the middle is not a price any more, it is an average of two different stories. This page is every minute in the published corpus where that happened, and what the engine did about it.

What this page is not. It is not a crash detector, and the corpus says so plainly. On 12 March 2020 bitcoin fell 52% while three venues agreed all the way down: mean confidence 98.78, the book held Full for most of the window, and what cut the ceiling there was volatility, not disagreement. An ordinary cascade is not an oracle failure. Both are worth surviving and they are different events. The replay measures the other one.

Every run where the venues came apart

Peak disagreement is the widest gap between the venues in any single minute of the window, measured after each venue's structural basis is learned and removed, in basis points of the mid. Twenty basis points is roughly where a normal book stops looking normal.

EventInstrumentPeak disagreementMinutes past 20 bpsLowest ceilingLiquidated: tier / here
10 October 2025TIA-USD717 bps271×83.3% / 48.1%
10 October 2025WIF-USD461 bps271×85.2% / 34.4%
10 to 11 October 2025SOL-USD297 bps91×85.6% / 80.1%
10 October 2025ONDO-USD271 bps261×80.8% / 44.2%
9 to 12 May 2022SOL-USD208 bps51×95.5% / 80.3%
16 to 22 August 2026TRUMP-USD167 bps21×92.2% / 88.4%
12 March 2020ETH-USD160 bps131×81.4% / 66.7%
12 March 2020BTC-USD98 bps141×81.7% / 66.6%
19 May 2021ETH-USD91 bps71×71.5% / 64.2%
8 to 10 November 2022SOL-USD77 bps371×89.1% / 71.3%
19 May 2021BTC-USD59 bps51×63.6% / 55.2%
2 to 3 February 2025ETH-USD59 bps41×39.2% / 38.7%
9 to 12 May 2022ETH-USD54 bps11×85.3% / 82.1%

What it costs to be wrong about this

The ceiling falls when the venues disagree, which means a trader is stopped from adding risk at the exact minute the price is least knowable. That is the point, and it is also the cost: on a book that recovers in ten minutes, the restriction bought nothing and the trader felt it. Across the 13 runs in the corpus where the venues disagreed by more than 50 basis points, the flat tier liquidated 79.6% of entry minutes and this engine 63.1%, a difference of 16.5 points. That is the axis earning its keep.

The honest reading

It fires rarely. 13 of 134 published runs reach 50 basis points of disagreement and 5 reach 200. On the other 121 the venues agreed, and whatever this engine did for a trader there, it did not do it with this measurement. A venue that sold itself on this axis alone would be selling an umbrella for one storm a year.

What the venue is seeing right now

The same statistic, live, on the books this venue quotes. It is published every minute at /api/v0/instrumentation and summarised on the status page, so the claim above can be checked against the market rather than against another replay.

MarketRegimeConfidenceWidest disagreement
loading…

How it is computed

Each venue carries a structural offset, a USDT pair against a USD pair or a future against spot, learned on a fifty second half-life and removed before anything is judged. What remains is deviation. The statistic charged is the widest of the median absolute deviation scaled to a standard deviation, the second largest deviation when four or more venues are live, and the excess of any single venue beyond three sigma, so one venue running away from a panel is seen rather than averaged into it. Half the prevailing spread is subtracted, because venues cannot agree tighter than their own quoting granularity, and the result is divided by what this book's disagreement normally looks like, learned on a two hour half-life. Confidence is exp(-bps / 120) on that number, multiplied by freshness and by the share of expected venues that are live.

The code is api/riskcore.mjs, the same function the live venue calls on every tick, and the corpus behind this page is the one published on the replay page, run on the deployed engine.