Forty-seven fields. That is how many slots a nine-dimension due-diligence template returned last week when it was pointed at a live token — technical architecture, supply schedule, market structure, ecosystem position, regulatory exposure, team composition, risk matrix, narrative durability, transmission path. Forty-seven fields, every one of them stamped N/A.
The pipeline had a defect. Upstream extraction returned an empty array: no title, no source, no fact primitives, no project name. The tool had nothing to chew on, and it said so.
The correct institutional response to that document is not 'inconclusive.' It is 'stop.' Instead, inside the same 72-hour window, three separate research feeds published confident two-thousand-word theses on the same asset. Each carried a price target. Each carried a chart. Each cited on-chain data that traced back to a single tweet. The blank document was the only honest artifact produced that week.
Crypto research industrialized somewhere around 2021 and never looked back. Scorecards, dimension grids, risk matrices — the machinery exists because a desk that is forced to enumerate validator sets and unlock cliffs will occasionally notice something a price chart hides. I use these templates. I built two of them.
The problem is never the template. The problem is the deadline attached to it. A framework that returns forty-seven nulls is telling you the instrument is unpriceable. A framework that returns forty-seven nulls and a 'hold' rating is telling you the analyst had a quota.
This is why I insist on an anchor chain in every piece of research I sign. Every conclusion carries a citation to a numbered fact primitive — subject, action, datum, source. No anchor, no claim. It is the difference between an audit and an opinion, and in a market where a single mispriced liability can cascade through three layers of collateral inside an hour, the distinction is not academic.
I learned that the hard way in 2018, auditing the 0x Protocol v2 contracts for three months while the ICO carnival ran outside the window. Seven critical edge cases. Each one traceable to a specific function, a specific state transition, a specific line of Solidity. I filed pull requests. Half were rejected on first pass. The point was never whether I was right — the point was that my claim could be checked against a diff. Sentiment is irrelevant without mathematical integrity. A thesis you cannot anchor is a thesis you cannot defend when the funding rate flips.
So read the empty array correctly. No fact primitives means no anchors. No anchors means any output is training-memory recall dressed in the costume of analysis. The tool that refused to produce it did more work than the three feeds that did.
Here is the structural insight most desks still miss. Markets are priced binary — long or short, risk-on or risk-off. Risk is ternary: known-good, known-bad, and unknown. Those three states have wildly different loss distributions, and only two of them get modeled.
Known-bad is bounded. You can size it. A depegged stablecoin has a floor at zero and a recovery path you can handicap. Unknown is unbounded, because the tail you have not measured is by definition the tail you have not bounded. In a bear market, the tail dominates the mean, which means 'unknown' must be priced at the extreme, not the center. Every desk that fills a null with a midpoint estimate is quietly selling tail risk at par.
In financial engineering we have a name for this failure. NaN propagation. Put a single not-a-number into a covariance matrix and the optimizer does not return a conservative portfolio — it returns no portfolio at all, or worse, a garbage one with plausible-looking weights. NaNs do not average to zero. They contaminate.
Portfolio construction already knows this. An unassessable asset in an institutional book gets zero weight. Not two percent 'just in case.' Zero. But retail does the exact opposite: sizes the unassessable position small enough to feel prudent and large enough to hurt. That is not diversification. That is a NaN with a position size attached.
The compliance stack has the same bug, one layer up. Take the Howey test — money invested, common enterprise, expectation of profit, derived from the efforts of others. Four booleans. Most analysis treats them as a vote: three yes and one unclear becomes 'probably a security, leaning cautious.' Wrong. Three true and one null returns null. The instrument is not 'probably regulated.' It is unclassified, and unclassified instruments get delisted, not grandfathered.
I spent 2023 running exactly this kind of simulation — a five-person team modeling the digital euro's deposit-substitution channel under strict holding limits. Our output was a 15% potential shift of Spanish retail savings out of commercial banks and into central bank liabilities. The number mattered less than the structure: we published confidence intervals, not a point estimate, because a point estimate would have been a null dressed as a forecast. Regulators do not punish uncertainty. They punish false precision. That is the regulatory anticipation framework in one line — the hurdle you get hit by is the one you claimed to have cleared.
Now the case study that should have ended the debate. Terra/Luna, May 2022. Sixty billion dollars of stablecoin value evaporated in roughly 48 hours. Not an ideology failure. A liquidity cascade — a reflexive loop where the marginal buyer of the peg was the same entity issuing it. The information was public. The 19.5% Anchor yield was visible on-chain to anyone with a terminal.
What broke was not data availability. What broke was the anchor chain. The market read a subsidy as demand, because nobody traced the yield to its funding source. A blank got filled with a narrative, and the narrative got sized at institutional weight. That is what a null position costs when you refuse to hold it. Liquidity doesn't price what it cannot see — it prices the fact that it cannot see, all at once, in the same hour.
And this is about to become a machine problem, not a human one. In 2025 I built a prototype for verifying human-versus-AI wallet interaction, and the hardest layer was never identity. It was provenance. An autonomous agent executing transactions does not have the luxury of a deadline-driven guess. An agent that treats null as zero gets arbitraged to death in a single block. An agent that halts on null survives, but only if its counterparty can verify why it halted.
The selection pressure is already visible. The machine economy will not reward the agent with the most data. It will reward the agent whose claims can be traced. Which means the scarce infrastructure of the next cycle is not compute and not throughput. It is attestation — a signed record of what an agent knew, when it knew it, and what it refused to price.
Which brings me to the part that will make this essay unpopular.
The consensus says more analytical dimensions produce more alpha. I think the tenth dimension on a thin-data asset has negative expected value. It does not add resolution. It manufactures precision, and manufactured precision gets levered. 'Do your own research' has been inverted the same way: research volume has gone vertical while information gain has gone to zero. We are not drowning in ignorance. We are drowning in confident output with no anchor chain.
And the incentive gradient points the wrong way, structurally, not morally. 'N/A' does not sell subscriptions. A 2,000-word thesis with a chart does. So the honest analyst is systematically outcompeted by the fluent one — right up until the cascade, at which point the fluent one discovers that the blank they filled was the whole risk.
There is a second misattribution worth flagging. The empty array last week was a scraper failure, not an asset failure. Most desks cannot tell the difference, so they route operational risk into the market book and hedge the wrong exposure. Pipeline noise is not price signal. Diagnose upstream before you touch the position.
So here is the forward-looking piece, and it is the only forecast I will make without a confidence interval attached.
Within eighteen months, model provenance becomes a compliance artifact. Any jurisdiction that has already decided algorithmic advice is advice — and MiCA's perimeter is drifting that way — will eventually demand an audit trail for it. Not the model's weights. The model's citations.
When that day arrives, the desks with anchor chains will produce a document. Everyone else will produce forty-seven N/As and a price target, and will have to explain which one they actually believed.
Liquidity doesn't forgive an unmarked gap on the balance sheet. It finds it, prices it, and moves on. The only open question is whether you were the one holding the null — or the one who filled it in.
