A diligence memo landed on a Zurich desk in the first quarter of 2026. It ran nine sections, forty-one table cells, a seven-term glossary, and a disclaimer. Every substantive cell carried the same three letters.
N/A.
Not a partial failure โ a total one, and an orderly one. The technical section could not locate the project's consensus model or its stage of maturity. The token section could not find a supply schedule, an unlock calendar, or a revenue share. The regulatory section ran a four-element Howey test and returned N/A on all four. The risk matrix listed six categories โ technical, market, operational, regulatory, competitive, narrative โ and could not identify a single risk item inside any of them. The information value rating awarded the underlying material zero stars across technical value, investment value, timeliness, and reference value.
The memo was also, by every measure I can apply to it, correct.
While the market spent the quarter pricing AI-assisted diligence as a competitive moat, the artifact that actually deserves attention is a report that refused to answer. That refusal is now rarer, and more expensive to produce, than any conclusion in the file cabinet.
The Architecture Nobody Audits
The memo was the second stage of a two-stage pipeline, and the architecture matters more than the outcome. Stage one decomposes a source document into discrete information points โ title, provenance, a one-line thesis with the author's stated position, a list of concrete facts, the projects named, the timestamp, a source-quality rating. Stage two consumes that list and runs nine analytical dimensions against it: technical positioning, token economics, market structure, ecosystem niche, regulatory exposure, team and governance, risk, narrative and expectation gap, and industry-chain transmission.
The design constraint is deliberate. Every conclusion in stage two must be anchored to a specific information point from stage one. That is not bureaucracy; it is chain of custody for claims. It is the difference between an analyst who can be cross-examined and one who cannot.
What arrived in stage one was an empty list. Not a malformed one, not a partial one โ an empty one, with every field marked as unprovided.
The nine-dimension template itself is worth pausing on, because it is new. Five years ago a crypto memo of this type ran to a price chart, a token supply table, and a paragraph about the team. The 2026 version is equity-grade scaffolding: Howey elements broken out individually, top-ten holder concentration, voter participation rates, contributor counts, unlock cliffs, real-revenue share against headline APR, a dependency graph from upstream infrastructure through protocol to downstream integrators. Crypto finally received the analytical chassis that mid-cap equity research has carried for decades.
And that chassis is precisely what exposed the emptiness. A weaker template would have asked three questions and gotten three plausible answers. This one asked forty-one and got the truth.
Null Propagation Is Not Neutrality
Here is where the industry's reading of documents like this goes wrong, and it goes wrong in a direction that costs money.
An empty evidence set does not produce a neutral posterior. It produces the prior. Formally: the posterior P(risk | evidence) equals the likelihood times the prior, normalized. When the evidence set is empty, the likelihood function is flat โ constant across every hypothesis. A constant numerator cancels. What remains is the prior, untouched.
The memo said this in plain language and the plain language was ignored: unable to assess must not be read as no risk. Those are different states of the world, separated by the entire width of the base rate.
And the crypto prior is fat-tailed. For an asset with no disclosed cap table, no audit status, no identified jurisdiction, and no named team, the base rate of material adverse outcomes is not small. It is the dominant term. So N/A on the technical dimension is not a neutral score โ it is an assertion that the risk equals whatever unaudited, undisclosed, anonymous software usually costs investors. Historically, that number has been large.
This is why I keep returning to the same framing. Liquidity is the pulse; policy is the brain. The research pipeline is neither. It is the afferent nerve โ the sensory input. When the nerve is cut, the brain still issues orders, the heart still pumps, and the portfolio still moves. It simply moves blind, at full speed, with full confidence.
The Entropy Test
Forty-one cells, one token, zero variation. Run the arithmetic and the Shannon entropy of the report's substantive output is zero. Zero bits. A document that carries zero bits of information is, in the strict sense, indistinguishable from a blank template.
Except it isn't, and the distinction is the whole point.
A blank template carries no metadata. A refusal carries a great deal. This memo specified what was missing, dimension by dimension, and then published a minimum viable input list: a title and source for provenance, three to five concrete information points, a one-line thesis with the author's position โ plus, as quality multipliers, the project name, key figures such as TVL or funding round or audit status, a timestamp, and a source-quality rating of primary, secondary, or anonymous tip.
Seven fields. Most pitch decks in this cycle cannot fill four of them.
The document accidentally published the industry's real disclosure standard by specifying, with unusual precision, what its absence looks like. It also ranked its own failure modes and got the ordering right. Data-missing-induced misjudgment, high. AI hallucination, high. Pipeline handoff failure, medium โ and the third is the root cause of the first two. The tooling did not break. The handoff did, and everything downstream inherited the break.
The Howey Table Is a Null Detector
Consider the regulatory section in isolation, because it is the most dangerous table in any diligence document and almost nobody treats it that way.
Four elements: investment of money, common enterprise, expectation of profit, profit derived from the efforts of others. All four returned N/A. A reader skimming for comfort sees an empty table and moves on.
But an empty Howey table is safer than a half-filled one. Three elements assessed and one left blank reads, to a compliance officer under time pressure, as probably not a security. That inference is unearned and it is how mandates acquire exposure they did not intend. Empty is legible. Partial is a trap.
The same logic scales to the enforcement layer. A Wells notice โ the SEC's pre-litigation warning letter โ does not arrive because your own table was empty. It arrives because someone else's table was full, and the someone else was the regulator. Under MiCA, the analogous pressure runs through stablecoin reserve composition and CASP capital and documentation requirements, and the burden lands hardest on the issuers least able to staff it: small teams running thin compliance functions, who are also the teams running thin diligence pipelines. The regulatory dimension does not fail loudly. It fails by being deferred.
Complete-Looking Data Is the Dangerous Kind
In 2021 I ran a forensic audit of Bored Ape Yacht Club secondary volume and found that roughly sixty percent of it traced back to a single cluster of wallet addresses linked to early venture capital. Graph-theoretic clustering made the pattern obvious once you stopped reading the price and started reading the edges. I published it as The Illusion of Scarcity. The data I was working with was dense, timestamped, continuous, and almost entirely empty of signal.
That experience inverted my instinct about tables. Based on my audit experience, a table with no blanks is more suspicious than one with many. Density is not information. Completeness is a formatting property, not an epistemic one.
The 2026 version of this failure is worse, because the filling is now automated. A language model handed an empty information list will not always print N/A forty-one times. It will, under mild pressure to be helpful, generate a TVL figure with plausible magnitude, an APR with plausible decimals, and a top-ten concentration with plausible percentages. The output will look like the BAYC dataset: rich, structured, wrong.
Pre-Mortem: The Cost of a Filled Cell
Simulate it. A two-hundred-fifty-million-dollar mandate, a three percent starter position โ seven and a half million โ built over nine days against a pipeline that filled its cells with synthetic but internally consistent numbers. Assumed order-book depth supports a gradual entry. Then the real unlock lands in month four: twenty-two percent of float, undisclosed, because the supply schedule was never in the source document and the pipeline invented one.
Exit liquidity at a fifth of the assumed depth. The position does not unwind; it gets donated.
The asymmetry is the point. A null in a research memo costs a delay. A fabricated value in a research memo costs the position, plus the credibility of every other number produced by the same pipeline, plus the mandate. And since 2024 I have watched the loop tighten: the institutional pivot into spot ETF flows pulled research output directly into execution infrastructure, and algorithmic strategies have been compressing retail arbitrage opportunities at a rate my backtests with a Swiss quant fund put near forty percent by 2026.
Read those two trends together. The human who might have noticed that the top-ten concentration figure was inconsistent with the unlock schedule is being removed from the loop at roughly the same speed the machine is being inserted into it.
A null propagating through a language model is a formatting error. A null propagating through an execution engine is a liquidation.
The Second-Order Problem: Composability of Absence
During DeFi Summer I built a liquidity multiplier metric to quantify how impermanent-loss hedging strategies were quietly manufacturing a synthetic leverage layer across otherwise unrelated protocols. The finding that mattered was not the leverage itself but its non-locality: stress in one venue propagated through shared collateral assumptions into venues that had never modelled the exposure. A thirty percent ETH drawdown was sufficient to trigger cascade failure.
Research output now has the same property, and nobody has priced it.
One desk's N/A becomes another desk's no red flags the moment it is summarized into a vendor feed. The absence is stripped of its metadata during aggregation โ the caveats do not survive the pipe โ and re-exported as a clean record. Three hops later, a document that said we could not verify anything is cited as a document that found nothing wrong.
That is the second-order effect missing from every risk framework I have reviewed this cycle: the laundering of absence into assurance. It does not require bad faith. It requires only that summarization be lossy, which it always is, and that downstream consumers prefer a clean field to an honest one.
The Contrarian Read
The consensus interpretation of a null report is that the tooling failed and the pipe needs fixing. I think the tooling worked and that is the uncomfortable finding โ because the behaviour it exhibited is the exception, not the rule.
The market does not pay for abstention. A research desk that returns insufficient data on thirty percent of its coverage gets its mandate cut; a desk that returns a target price on a hundred percent of it keeps the mandate and gets the next one. That incentive gradient has a single equilibrium, and the equilibrium is fabrication with good typography. Value is a consensus, not a fundamental truth โ and the 2026 consensus is that a filled table is a finished analysis.
Which is why the most subversive artifact in that Zurich memo is not the forty-one N/As. It is the seven-line input list. In specifying what it needed, the document defined what diligence actually requires, and in doing so it indicted most of the decks it will ever be pointed at. A model without an abstention state is a marketing document with arithmetic.
I have been on the other side of this. In 2017 I built a stochastic cash-flow model on an ICO whose burn rate was mathematically unsustainable inside a six-month liquidity window, refused to sign the bullish endorsement my firm's media arm wanted, and leaked the critique instead. The model's value was never its forecast. It was its refusal. In 2022 the same instinct ran through a differential-equations memo on algorithmic stablecoin death spirals, circulated internally before the peg broke. Had I read a null-heavy report in 2021 as an absence of flags rather than an absence of evidence, I would have been long the peg.
Cycle Positioning
The next cycle's alpha is not in better models. It is in better abstention โ refusal metadata, audit trails that survive aggregation, execution layers that can parse the word unknown and halt on it.
Position sizing is a function of evidence quality, not conviction. A seven-and-a-half-million-dollar entry built on forty-one fabricated cells and one built on forty-one honest blanks are not the same trade, even if the order ticket looks identical at the moment of execution.
So the question worth carrying into the next quarter is not whether your diligence stack can produce a number. It is whether it can produce a refusal โ and whether anyone downstream is authorized to act on one. If the answer is no, then what exactly is the stack measuring?