Hook
Last month a counterparty sent me a research memo on a mid-cap rollup token. Nine analytical dimensions. A weighted scoring matrix. Eleven risk flags, each carrying a severity rating. Every cell in the table resolved to the same value: N/A โ insufficient information.
The document ran roughly 1,400 words. It contained zero facts.
It was forwarded twice more inside the firm as completed diligence. Nobody flagged the void, because the void had been formatted. I counted the rows twice. Nine rows, nine nulls, zero verifiable inputs โ and a distribution list of six people who read it as work.
This is the failure mode I have spent eight years warning about, and it is now the default output of an industry that industrialized the production of research faster than it industrialized the verification of it.
Context
In a bull market, bad research is expensive but survivable. Fabricated throughput numbers, phantom integrations, a tokenomics pie chart drawn in a design tool โ these produce losses a position can absorb. The asymmetry forgives.
In a bear market the asymmetry inverts. A single wrong claim about liquidity depth, an unlock cliff, or a market-maker agreement can end a fund. So the rational researcher, operating under institutional review pressure, learns to produce something safer than a claim: a framework.
Frameworks cannot be wrong. A taxonomy of nine dimensions can be applied to anything, including nothing, and it will still return a valid structural artifact. That is what I received. Not a lie. A container.
The incentives reinforce it. The current discoverability regime rewards documents that look like they contain information gain โ novel headings, scored rubrics, flagged risks โ without requiring that any of the underlying fields be populated. Structure has become cheaper to manufacture than evidence, and it reads better in a committee packet.
I have watched this pattern compound since 2018, when I spent four months on the tokenomics of a privacy coin called Project Aether. I found a deflationary burn schedule that would evaporate liquidity within eighteen months, wrote forty pages, and recommended rejection against significant internal pressure. What I could not do โ with the data available at the time โ was model the second-order effect of that burn on market-maker inventory. I marked the gap explicitly in the memo. That single line of honest absence was cited more often in later committee reviews than any of my forty pages of conclusions. The absence was the deliverable.
Core
Here is what the memo got technically right, and why it still failed.
Every LLM-assisted analysis stack I have reviewed this cycle โ seven in the last six months, three of them inside banks โ exhibits the same three defects, in the same order:
- Null propagation with schema survival. The input layer returns nothing; the output layer still emits a well-formed object. The structure is inherited from the prompt, not from the world. Formally: the document is a function of the schema, not of the underlying system.
- Conflation of "not available" with "not applicable." These are distinct states and should never share a token. "Not applicable" is a conclusion โ a claim that a dimension does not bear on the asset. "Not available" is a failure โ a claim about the researcher's own instrumentation. Every pipeline I have examined collapses them into one string.
- Refusal rendered as output rather than error. The report did not crash. It did not throw. It rendered N/A nine times and returned HTTP 200. Downstream readers had no signal to distinguish a completed analysis from an empty one.
The oracle precedent. In November 2020, a lending market on Ethereum lost roughly $10 million to an oracle vector that was not price manipulation in the ordinary sense. The issue was latency and staleness. A feed that stops updating does not go silent. It keeps returning its last value. And a stale price is not a neutral price โ it is a specific, exploitable state with a defined direction.
I built a quantitative model of that failure in 2020, simulating oracle update intervals against liquidation thresholds across historic volatility regimes, and published the methodology. Five thousand stars on GitHub. The reason it mattered was never the exploit itself but the general form it exposed: absence presented as data produces confident, directional, wrong decisions โ not cautious ones.
That form transfers directly to research. Nine N/As in a scoring matrix do not read as ignorance. They read as a completed matrix. A reader scanning a table sees rows. Rows imply work. Work implies a stronger prior than blank space ever would. The formatting layer is doing persuasion that the content layer did not authorize.
The statistics of a null set. Math doesn't lie. But math has nothing to say about nothing. Run a Sharpe calculation on an absent price series and you get an undefined denominator, not a zero. This is the part institutional readers consistently miss: undefined is not the midpoint between positive and negative. It is a third state, and it does not average. A nine-dimension rubric scored across nulls produces a weighted composite that is arithmetically well-defined and epistemically empty.
In 2024 I built a spot-versus-futures premium arbitrage model that back-tested to roughly 12% annualized alpha across 2017โ2021. That number was real because the basis history existed โ eight years of observable premium and discount. When our chief strategist asked me to extend the estimate forward into a window where the instrument did not yet trade, I declined and gave him the confidence interval on the null window instead. The reallocation to structured product still went through. It went through smaller, and it went through hedged.
Why this matters more now than in 2021. May 2022 taught the market to fear feedback loops in stability mechanisms. I spent six weeks modeling the UST/LUNA reflexive relationship and published the drain-rate estimate three days before the terminal move. The lesson institutions extracted was "algorithmic stablecoins are risky." The lesson they should have extracted was narrower and more transferable: the speed at which a system fails is a function of how quickly it can mistake its own internal state for ground truth. A death spiral is a stale-data event. The mechanism stopped observing the world and started observing itself, and the reflexive loop had no external reference to break it.
Every research pipeline returning a formatted N/A is running a slower version of the same loop. It is not failing to observe the asset. It is observing its own prompt.
Contrarian
The prevailing threat model is wrong. The industry worries about hallucination โ models inventing protocols, fabricating TVL, citing papers that do not exist. Hallucination is dangerous precisely because it is loud. It generates checkable claims attached to specific entities, and specific entities have counterparties who will object. It gets caught, eventually, at cost.
The quieter failure is schema compliance. A pipeline optimized on output validity will always return a valid object, and validity is indistinguishable from accuracy at the formatting layer. Code is law, until it isn't โ and here the code is delivering syntactically perfect objects with no referent. Nothing throws. Nothing alerts. The monitoring layer watches for errors, not for emptiness.
โ Scenario: When debunking a project, the most valuable line in the entire report is usually the one that says nothing can be concluded. It is the only line carrying information about the researcher's own position rather than the subject's. I have rejected more deals on that line than on any other, and I have never regretted one.
So the document I opened last month may in fact be the most honest artifact I received this quarter. It admitted that its instrumentation failed. It simply did so in a font that looked like diligence, which is why six people forwarded it.
Takeaway
The next question I ask any counterparty is not what they concluded but what they had. Show me the input manifest: which sources, at what timestamps, with what coverage, under what staleness tolerance. A memo that cannot produce one is not research. It is typography.
Position for the cycle ahead: verifiable data provenance becomes the scarce asset. Protocols that publish lineage โ signed feeds, timestamped attestations, reproducible numerics, explicit confidence intervals on the null window โ will price at a premium to those that publish narrative. The bear market is not killing the projects with bad numbers. It is killing the ones that never had numbers at all โ and it is taking their research vendors with them.