128 Nulls: What a Completely Empty Research Report Gets Right
Hook
Last month a nine-section research report crossed my desk. Fourteen tables. Roughly four thousand words. One hundred and twenty-eight discrete data cells.
One hundred and twenty-eight of them read N/A.
The client wanted a refund. He had commissioned an analysis of a token his fund was screening, and what came back, by his reading, was a template. He forwarded it to me expecting confirmation that he had been defrauded. I told him he was holding the most honest document produced by crypto research that quarter, and that he should have it framed.
He did not find this helpful. The empty cells were not a bug in the report. They were the report working.
Context
The architecture is now standard. Nine dimensions: technical positioning, token economics, market structure, ecosystem niche, regulatory exposure, team and governance, risk matrix, narrative expectation gap, supply-chain transmission. Under each, a fixed set of tables. Supply structure by holder class and unlock schedule. Howey elements, assessed individually. Funding rounds with valuation and lock-up. A risk matrix scored by probability and impact across six categories.
This is a venture capital diligence checklist repackaged as crypto-native research. It is not a bad framework. It is close to the structure I used in 2017, when a Sydney legal firm contracted me to audit the tokenomics of a project advertising 1,000% APY. Six weeks of modelling liquidity pools against securities law produced a single conclusion: 40% of supply sat unvested, and the cliff was eleven weeks out. The project was delisted from local exchanges inside a month. The framework worked because the inputs existed, and because the inputs were ugly.
What changed between then and now is not the framework. It is the throughput. More tokens, more desks, more capital, and a publication cadence measured in hours rather than weeks. Under those conditions the commercially rational product is a template that emits a finished document from any input — including no input. A process that occasionally returns nothing is worth less on a word-count contract than a process that always returns something.
That is the incentive. Not a conspiracy; a payout structure. And it produces a specific, measurable failure mode.
Core
Most published crypto research cannot return a null result. Not rarely — cannot. It is architecturally incapable of it.
The pattern is familiar from software. Correct implementation:
def valuation_metrics(token_data):
if not token_data:
return None # absence propagates
return compute(token_data)Absence in, absence out. Now the version that actually ships:
def valuation_metrics(token_data):
try:
return compute(token_data)
except Exception:
return {"verdict": "bullish", "confidence": 0.8}The second function never crashes. It also never tells the truth. Every null becomes a confident output, and nothing downstream can distinguish a computed number from an invented one. Silent coercion of absent data into plausible data is the most expensive bug pattern in financial software, and the research industry has industrialised it as a product feature.
The empty report did the opposite. It propagated the null all the way to the cover page. That is why it looks broken.
Now the asymmetry. Any analytical process makes two classes of error. Type I: recommending something that fails. Type II: rejecting something that succeeds. In a market where a single 100x dominates a portfolio, the incentive gradient is violently skewed toward avoiding Type II errors. Publishing a cautious pass earns nothing. Publishing a hit earns a fund. The entire research layer is therefore optimised to say yes, and the cost is settled in Type I errors — capital allocated to instruments whose economics were never modelled, only described.
Quantify it. An analyst who fabricates a real-yield assumption of 14% on a token that actually yields 6% has introduced an eight-point error. On a $10 million allocation that is $800,000 of annualised mispriced carry, compounding against a position the fund cannot exit without moving the market. The report scored well on every readability metric. It was also a loss.
Even in the best case — a lending market with genuine, non-inflationary yield — the rate is set by a governance parameter curve calibrated by vote, not by a term structure. The number is arbitrary by construction. Fabricating a second arbitrary number on top of the first does not make the analysis wrong. It makes it doubly unverifiable.
The Howey tables in the empty report deserve attention. All four elements marked N/A. That is correct, and it is also the answer to the question the client actually asked. A token with no disclosed issuer, no identified promoter, no distribution schedule and no stated use of proceeds cannot be assessed for securities status, because the information required to apply the test does not exist in the public record. The absence is not a gap in the analysis. The absence is the analysis.
Same with the supply-chain transmission map. The empty report renders it as:
N/A → N/A → N/A
? ?A diagram of question marks looks like a failure. It is more truthful than a diagram with fabricated edge weights. Without a subject there is no causal path. Assigning one is not analysis. It is decoration.
And the risk matrix — six categories, every cell blank. A filled-in risk matrix is a rhetorical device. Probability multiplied by impact is only meaningful if the probability is estimated from a reference class. Most crypto risk matrices are produced by asking a human how likely something feels. That is not a probability. That is a mood with a decimal point.
I have run this exercise in the other direction. In 2023 I evaluated an NFT project claiming virtual real estate yields. I asked for contract access instead of the deck. The "yield" was a redistribution of new buyer deposits; no external revenue touched the contract. 95% of holders were wallet clusters controlled by the team. That report was short, and it also read like a demolition rather than an analysis. Trading volume fell 60% in three weeks. Nothing was fabricated to get there. The absence of a revenue stream was the finding.
The empty report did not go that far, because it had nothing to examine. What it did instead was enumerate, on the final page, the minimum information required to proceed: factual content of at least a few hundred words, the specific protocol name, a timestamp, a source, TVL and holder data, the chain layer. Five required inputs, five suggested. That list is more actionable than any rating the report could have contained, because it is falsifiable. A rating of "buy" is unfalsifiable for a quarter. A list of missing inputs can be verified in a day.
Contrarian
The bull case for the empty report is not that it is thorough. Nine dimensions and fourteen tables is over-engineered; I would have cut four of them. The bull case is that it is the only artefact in the stack capable of being wrong in a way you can detect.
Consider the alternative. A junior analyst under deadline pressure fills the same template with plausible values: a 12% real yield, a 4/10 risk score, a team rating of "experienced." Every cell populated. The document reads better. It clears internal review faster. It generates a position. And it has converted a measurable unknown into an unmeasurable certainty — strictly worse than the null it replaced, because the null at least announces itself.
There is also a screening function hiding here. Run that framework across two hundred tokens and let it return N/A on one hundred and eighty. You have not produced one hundred and eighty failed reports. You have produced a filter with a 90% kill rate and a documented reason for every kill. A filter that knows which of its inputs are load-bearing is worth more than one that assumes every input is free — the same lesson rollup operators are about to relearn when blob space stops being subsidised.
The client's refund demand is rational under a word-count contract and irrational under a capital-preservation contract. Which contract he actually signed is a question about his procurement process, not about the analyst.
Takeaway
The next cycle's differentiator will not be a better model. Everyone will have the same models. It will be better null handling — research pipelines that record absence as a first-class state instead of coercing it to zero, and audit trails that can prove a number was never invented. It is the same problem I spent last year on with an Australian bank: reconciling SQL ledgers against on-chain state, where the hardest case was not a mismatch but a missing row.
So: when your research process finds nothing, does it tell you? Or does it hand you {"verdict": "bullish", "confidence": 0.8} and let you discover the difference later?
In the absence of data, opinion is just noise.
