A meme coin at full confidence, gold degraded
Our first honest confidence reading punished the most reliable market on the venue and waved through the least. The data was fine. The measurement was not.
Confidence in Openfield is built from three things: how far apart independent sources are, how fresh they are, and how many of them are actually alive. The first one carries most of the weight. If Binance, OKX and Coinbase all say a thing costs the same, we know the price. If they drift apart, we know less, and the engine reduces what you are allowed to do with that knowledge.
The first version measured drift the obvious way. Take each source, adjust it by its learned basis, compare it with the median, and record the largest gap in basis points. Fifteen basis points of disagreement is worse than one. That is true, and it is also where the trouble started.
The engine came up with this:
What the first version reported
| Market | Raw divergence | Confidence | Regime | Leverage |
|---|---|---|---|---|
| PEPE-USD | 15.33 bps | ~85% | Degraded | 1× |
| XAU-USD | 4.86 bps | ~89% | Degraded | 7× |
| BTC-USD | 0.22 bps | 99.3% | Full | 19× |
Real figures from the engine, with the divergence correction switched off.
PEPE dropping to a single turn of leverage looks defensible until you ask what those fifteen basis points actually were. PEPE trades around three and a half millionths of a dollar. The venues quoting it publish prices to a fixed number of decimals, and at that price the smallest step a venue can express is already worth about fourteen basis points of the price itself.
So three venues can agree completely, in the sense that there is no price between them that any of them could have quoted, and still be recorded as fifteen basis points apart. The engine was not measuring disagreement. It was measuring the size of a tick and calling it doubt.
A number can be computed perfectly and still answer a question nobody asked.
Meanwhile gold, quoted to two decimals on a four-thousand-dollar price, has a quoting floor of about a hundredth of a basis point. Every one of gold's 4.86 basis points is a real disagreement between two crypto venues and a COMEX future, and it is telling you something: the basis is moving, and the hedge you would put on is not currently free. That is exactly the case where leverage should come down.
The first version treated these two situations as roughly the same. One of them was noise. The other was the single most informative signal on the venue that day.
The correction
We estimate the quoting floor directly from the sources we already have. Each venue reports its own bid and ask, so the median spread across live sources is a decent measure of how finely this asset can be priced at all. We subtract half of it before judging divergence, and floor the result at zero.
Same tick, corrected
| Market | Raw | Quoting floor | Net divergence | Confidence | Regime |
|---|---|---|---|---|---|
| PEPE-USD | 15.33 | 14.39 | 0.95 bps | 99.1% | Full |
| SHIB-USD | 1.24 | 9.70 | 0.00 bps | 100% | Full |
| XAU-USD | 4.86 | 0.11 | 4.75 bps | 89.5% | Degraded |
| BTC-USD | 0.22 | 0.01 | 0.22 bps | 99.3% | Full |
Live from the venue on 1 September 2026. These numbers are visible at any time under the confidence figure in the terminal, and through /api/v0/markets.
The ordering inverts, and it inverts correctly. A meme coin priced in millionths sits at full confidence, because its sources genuinely agree as precisely as they are able to. Gold sits in Degraded, because its sources genuinely do not.
Note what does not change. PEPE is still capped at five turns of leverage and still graded D, because its risk genome says a token like that gaps and its liquidity is thin. Correcting the measurement did not make the asset safer. It moved the caution to the place that actually justifies it: the genome, which is a deliberate judgement we published, rather than an artefact of decimal places, which is an accident.
Why this matters beyond one token
Any venue that turns a data-quality signal into a trading constraint inherits this problem. The moment you say "we will reduce leverage when our sources disagree", you have promised that your disagreement metric means the same thing across every asset you list. It does not, unless you make it.
The alternative most venues choose is to not have the signal at all. Take one price feed, or a median of three with no opinion about it, and set leverage from a table written by a risk committee last quarter. That never produces an embarrassing inversion, because it never produces a reading. It just fails silently later, in the one hour when the sources actually diverge and nothing in the system notices.
We would rather have a measurement that can be visibly wrong than a table that cannot be.
The correction is not the interesting part of this story. The interesting part is that the venue printed its own reasoning, an outsider could read it, and the absurdity was legible in seconds: PEPE at one turn, gold at seven, both labelled Degraded for reasons the display could not distinguish. Had leverage been a number without a derivation, this would have shipped, and it would have looked entirely reasonable.
What is still unresolved
The half-spread estimate is a heuristic. It happens to fit well across seventy-one crypto markets and eleven macro ones, and we have no proof it is the right coefficient. A venue quoting a wide spread for reasons of genuine uncertainty rather than tick granularity will have some real divergence forgiven, which is the direction of error we like least.
The honest position is that this is version two of a measurement, not the final one. When it breaks again, the replay will be here.