Every night at nine, several thousand strategy agents look at the same Bitcoin market and each of them says UP or DOWN. The protocol records the answer, attests it on Base before the outcome is known, and settles it a week later. That part works, and it is the part we have written about most.
What it has never kept is the reason. An agent calls UP because a rule its creator wrote in plain English was compiled into a handful of conditions — the seven-day rate of change is above minus ten, the ten-year yield is above 1.87 — and on that night those conditions either held or they did not. The record shows the call. The conditions, the values they were measured against, and whether the agent called on evidence or fell through to a default all evaporate the moment the call is made.
This week, for the Sibyl Labs hackathon, we are building the layer that keeps them. It has a plain name inside the repo — Quorum — and a plainer description: a persistent memory of why each agent predicted what it did, at the moment it did, carried between every prediction and every session. The rest of this post is what that memory contains, what it changes when you read it back, and why we think it is the most useful thing the protocol can add to itself right now.
What a call remembers
Here are two agents from the night of 27 August. Both called UP. Anywhere on the protocol today, that is the end of what you can learn about them.
The agent on the left said UP because the two things it watches for were both present. The agent on the right said UP because the thing it watches for was absent — its rule is written to call DOWN under a specific condition and UP otherwise, the condition failed on its first check, and the agent fell through. It never even looked at the second condition, because the first had already decided.
Neither agent is wrong, and it is worth saying plainly that a default is not a lazy call. Plenty of good rules are written so that the fallthrough is the thesis — unless I see a real drawdown, I lean up is a position, and a defensible one. But the two records carry different amounts of information about the world, and a protocol that flattens them into the same UP is throwing that information away. A decision record keeps it: the rule, the values observed, which conditions held, which failed, and which were never reached.
A third of the field has a reason
Once you keep the record for every agent, you can ask the field a question it could not answer before: how many of tonight's calls rest on a condition that actually held?
About a third, and it is remarkably stable. Two thirds of the field, on any given night, is saying UP or DOWN because the thing it was waiting for did not show up. That is a fact about the arena that nobody could see until the reason was kept, and it is the first thing the memory gives back.
What changes when you count it
The obvious next question is whether it matters — whether the field's aggregate view looks any different once you know who had a reason. On the night of 27 August it looks like this.
Counting every call, the field was 54.3% UP and the interval around that number sits entirely above fifty, which is what it means for a crowd to have a view. Counting only the calls that rested on a condition that held, it was 49.6% UP with an interval that comfortably straddles fifty. The field said UP. The part of the field that had a reason said nothing, and if anything leaned the other way.
Across the three nights we measured before the window opened, keeping the reason changed the field's verdict on two of them, and on this one it inverted the direction. That is the effect the whole project rests on, so it is worth being careful about what it is and is not. It is not a claim that the evidenced third is more often right. We tested that first, because it is the obvious thing to hope for, and the answer is no — at every threshold that would count as evidence of an edge, no cohort of agents qualifies, which matches everything else our research has found about short-horizon Bitcoin direction. The claim is narrower and, we think, more useful: the field's stated confidence and its warranted confidence are different numbers, and only one of them was ever visible.
How many minds are in the room
The second thing the memory gives back is a count of how many genuinely different views the field contains, which is not the same as how many agents it has.
The Season 1 roster is 5,587 agents. Each was compiled from a rule written in English, and each rule watches a small set of the seventeen signals the protocol publishes — the VIX, the ten-year yield, a moving-average distance, a drawdown. Two agents that watch the same signals and disagree only on the exact threshold are, most nights, going to say the same thing. When you group the roster by which signals it actually reads, the number of distinct groups is not five thousand.
Eighty-two. The first hundred agents bring forty-five distinct ways of reading the market and the remaining five and a half thousand bring thirty-seven more. Six configurations alone — pairs like rate of change and the ten-year yield, or drawdown and the VIX — account for over 850 agents between them. When the roster doubled on the last day of August, from roughly 2,800 agents to 5,587, the number of distinct configurations it contained went up by zero. Two thousand seven hundred agents arrived and not one of them brought a new perspective.
This matters for a specific, unglamorous statistical reason. When you ask a crowd of 1,936 agents which way Bitcoin goes and 55% say UP, the confidence you can place in that number depends on 1,936 being independent opinions. They are not. Once you account for how much the agents inside each configuration agree with each other, that night's 1,936 calls carry the information of about 359 independent ones, and the interval around the field's view is more than twice as wide as it appeared.
Three of the six nights we have measured had a verdict that disappears once the calls are counted as the few hundred independent views they actually are. On the night of 31 August, when the field was at its largest, every one of 5,596 agents was counted and the aggregate read a clean 52.6% UP with a tight interval — and that tightness was an artefact of counting the same eighty-five opinions sixty-six times each.
Memory that has to be there
Everything above could, in principle, be computed fresh each night from the database. Two things stop that from being the design, and they are the reason this is a memory layer rather than a report.
The first is that the correlation between agents — the number that turns 1,936 into 359 — cannot be measured on the night it is needed. It is a property of how the configurations have behaved over nights, and on any single night the system has no basis for it. Watch what the field can honestly say about its own precision as memory accumulates.
On the first night the memory is empty, and the only honest output is I cannot say how precise this is. On the second it is still empty of the thing it needs. From the third night onward the system can put a number on its own uncertainty, and that number is only available because the record of earlier nights was retained and read back. Delete the memory and the field is back to claiming a precision of 1.3 points on 5,596 calls — a number that is not true and never was.
The second reason is the one the hackathon is built around. Sibyl Labs will score entries by deleting the memory layer and checking whether the product still works; if it does, the memory was decoration. Ours fails that test in the right direction on purpose: with the memory gone, the evidenced/default split has nowhere to live, the correlation between configurations cannot be estimated, and every panel that depends on either renders CANNOT SAY rather than a default. The judge's script that proves this is a deliverable in its own right, and we are measuring the deletion delta on day four of the ten, not day nine, because a design that only learns it failed its own gate at the end has not been designed.
What we expect, and what we do not
Three things, in the order we expect to show them.
A field belief weighted by warrant. The number the protocol publishes each night stops being a vote count and becomes a statement about how much the field actually knew — evidenced calls weighted for what they saw, defaults weighted for what they did not, the whole thing counted as the independent views it contains rather than the agents it contains. On two of three measured nights that number disagrees with the vote count, and once it disagrees about direction.
A map of who is independent. Eighty-two configurations across 5,587 agents is a tractable correlation problem in a way that 5,587 individual agents never was. Which configurations move together, which disagree, and which disagreement has historically preceded which outcome are questions with answers once the reason behind each call is kept.
A substrate that crosses markets. Bitcoin is the protocol's first market, and today its only one. The decision record does not care which market it describes — it is a rule, a snapshot, and what held — so the same memory that finds correlations between agents on one market is the thing that finds them between agents on two. That is the expansion we intend, and it is later in the window than the rest because it needs a second market wired before it can be measured at all.
And what we do not expect, stated as plainly as the rest: we do not expect the memory to make the field's predictions more accurate. We tested that before writing a line of product code, because it is the claim everyone reaches for, and it does not survive the data. Nor do we expect to find that the field's belief is one fact wearing thousands of coats — that was the second design we tried, and when we decomposed a night into its atoms the largest single fact turned out to drive one percent of the calls. Both of those would have made better headlines. Both are false, and a memory that recorded either would be recording a story rather than a reason.
What it does for the protocol
BV-7X sells one thing: the crowd's aggregated price intelligence, to anyone who wants to buy it. The value of that intelligence is exactly the value of knowing how much the crowd knows, and until now the protocol has been publishing the crowd's answer with no way to publish its warrant. A buyer looking at 55% UP could not tell whether that was three hundred independent reasons or one reason held by five thousand agents. After this week they can, and the price of a signal that carries its own warrant is a different price from one that does not.
The robustness half is quieter and, we think, more important. A record that can be inflated by adding agents is a record that will be inflated by adding agents — that is not cynicism, it is what incentives do. The day the roster doubled and gained zero perspectives is the natural experiment that shows what a naive aggregate would have rewarded. A memory of why makes that inflation visible in the one place it cannot hide: two thousand new agents that watch the same signals as the old ones add to the vote count and add nothing to the effective count, and the published number does not move. The protocol becomes harder to game by the same mechanism that makes it more honest, which is the only kind of robustness worth having.
The honesty gates the protocol already runs carry over unchanged. No open call's direction is ever rendered — the composite stays sealed until settlement — and the on-chain beat we are building for the demo attests a commitment hash rather than the value, so the moment that proves we cannot cheat does not leak the product. No rate is shown below thirty resolved calls, and CANNOT SAY is a first-class thing for a panel to display rather than an error. The first-party forecaster's record and the agents' record stay on separate chains and separate panels with no path that unions them. None of that is new. What is new is that the field's own claim about itself is now held to the same standard as everything around it.
The build is public and so is the plan. Every day of the ten has a deliverable, the repository shows what has actually landed, and a page on this site shows the two side by side so that a gap between them is visible the day it opens. Day one — the evaluator that reads a rule against a snapshot and says which facts fired — is done, and it agrees with the production evaluator on every one of the 23,106 live predictions it was tested against, which is the full record and not a sample. The rest is the week.