Forty-seven fields. Nine sections. Zero values.
I was handed a research artifact last week. A second-stage deep analysis report on an unnamed crypto asset. It had the full furniture: a technical assessment matrix, a tokenomics unlock schedule, a Howey test breakdown, a nine-cell risk matrix, a supply-chain transmission graph. Every cell rendered. Every cell empty.
The header was the interesting part. Analysis status: Insufficient information. Deep analysis cannot be completed.
Then it did something unusual. Instead of stopping, it printed the entire framework anyway โ section by section, with N/A in place of every input. One line read [Upstream dependency] โ [This project] โ [Downstream integrator], with N/A on both ends and a lone pipe where the protocol should be.
That is not a failure. That is a schema. And a schema with nothing in it tells you more about the analyst than the asset.
Where this thing comes from
It is a two-stage pipeline. Stage one ingests source material โ a news article, a governance post, a listing announcement โ and extracts a title, a list of information points, a core thesis, project names, and a source-quality score. Stage two takes that structured output and runs it through a fixed scaffold: technical, tokenomics, market, ecosystem, regulatory, governance, risk, narrative, supply-chain transmission.
The design is sound. Extract, then evaluate. Separate the messy string from the typed struct.
The problem is the interface. Stage two has no input validation on the handoff. Stage one returned nulls. Stage two did not throw. It executed against an empty struct and dutifully rendered the template, all forty-seven cells of it.
I have built systems with exactly this shape. In 2024 I led a team analyzing Lido DAO's treasury management contracts. We pulled governance proposals, decoded the calldata, diffed parameters, then replayed the changes in Hardhat. Our first pipeline had the same defect: it would happily produce a report about a proposal that did not exist, because extraction returned empty and analysis treated empty as a valid state.
We fixed it with a hard gate. Fewer than three extracted information points, abort. Not render-with-blanks. Abort.
The artifact in front of me is the render-with-blanks version. It took ten minutes to realize I was looking at something more honest than most published crypto research.
The template is the product
So the real question is not why it failed. The question is whether the framework holds up.
Judge a due diligence template by what it refuses to say, not by what it asks.
Nine sections. Let me score each as an instrument.
Technical assessment. Four rows: innovativeness, maturity, security assumptions, performance. Each row demands a competitor comparison. That is the right shape. Innovation without a comparator is a press release. Maturity without a versioned commit history is a claim. Security assumptions without an enumerated trust model are marketing. Performance without a declared load condition is a benchmark somebody chose.
But there is a gap, and it is structural. None of the four axes is adversarial. None asks what the worst sequence of calls a funded attacker can execute inside a single block looks like. In 2023 I spent three months reverse-engineering Arbitrum Nitro. The output was a fifty-page memo. The part that mattered was not throughput. It was the precompile boundary โ the seam where the WASM-to-EVM handoff produced behavior that neither the EVM specification nor the Nitro design document fully described. A four-row matrix has nowhere to put that finding. The template is incapable of holding it.
Tokenomics. This section is better. It asks for the four-way allocation split, the unlock schedule, and โ critically โ a Ponzi-structure risk flag. That flag is what separates an analyst from a listing memo.
I audited an AVS specification in 2025 and found that the slashable amounts were mathematically insufficient to deter sybil behavior during low-liquidity windows. That is a tokenomics finding. It belongs in exactly this row.
But the row requests a flag, not a model. A flag is binary. Security is a curve. The correct output is not HIGH or LOW. It is the cost of corrupting the system plotted against the cost of acquiring the stake, sampled across the liquidity distribution. Anything less is a vibe with a border around it.
The Howey test. Here the template gets genuinely interesting, because this is the only well-formed section in the document. Money investment. Common enterprise. Expectation of profit. Efforts of others. Four elements, each scored, then a composite judgment.
Howey is a legal instrument, not a technical one, and that matters. It is the single place where the framework admits the answer may be decided by someone other than an engineer. The Tornado Cash sanctions reset how developers reason about code-as-speech, and they were decided by a court, not a compiler. Any framework that treats regulatory exposure as a footnote is lying by omission. This one gives it a full section.
The same failure mode shows through, though. Four N/A cells and a composite judgment of N/A. The template will emit a legal conclusion from zero evidence. That is not a rendering bug. That is a conceptual one.
Risk matrix. Six categories โ technical, market, operational, regulatory, competitive, narrative. Five columns โ item, severity, probability, impact, mitigation.
This is the weakest section, and it is the one every report leads with.
Probability and impact are multiplicative, so a six-by-four grid yields twenty-four cells. Unless you calibrate the scale, every analyst grades on a different curve. One team's medium probability is another team's likely. I have read hundreds of these grids, and the variance between two analysts on the same asset exceeds the variance between assets. The matrix is measuring the author, not the protocol.
A risk matrix without a calibration protocol is a personality test with columns.
Compare that to what I do. When I benchmarked Nitro's precompiles, I graded nothing. I ran the transactions and recorded the gas. The number had no opinion. Code is the only law that compiles without mercy. When the output is a label โ HIGH, MEDIUM โ the author re-enters the system through the back door.
Supply-chain transmission. The graph is [Upstream] โ [Protocol] โ [Downstream], with impact direction, magnitude, and time frame per segment. The time-frame column is what saves it. Most contagion analysis fails because it conflates a same-block effect with a two-quarter effect. MEV extraction is instant. Unlock pressure is scheduled. Governance capture is a slow process that only becomes visible when a quorum finally fails.
The upstream node in this artifact was empty, and that is the honest version. Most real reports fill that node with "miners," which in a post-merge, post-restaking world is a category error. The upstream of a modern DeFi protocol is not hardware. It is staked capital and oracle feeds. If your supply-chain diagram still has a pickaxe in it, your model has not been updated since 2021.
Ecosystem position. Developer signals โ contributor count and trend, contract deployments. User signals โ DAU/MAU, retention.
This is the one section where the empty version is actively preferable to a filled one. Contributor counts are trivially gameable and universally misread. A repository with forty contributors and one maintainer pushing ninety percent of the commits is not a healthy project, and a commit graph will never tell you that. The instrument you want is bus factor and merge-authority distribution. Neither is in the schema.
DAU has the same problem. In a bull market, DAU is a function of incentives, incentives are a function of emissions, and emissions are a function of price. The metric is downstream of itself. With euphoria running as hot as it is right now, every one of those numbers is inflated by farmers who will exit the block after the emission cliff. You cannot evaluate retention in a market where leaving is irrational.
Market and governance. Two sections I skipped, and both deserve a line.
Market asks for cycle positioning, pricing-in degree, expected volatility, funding rates. Funding is the only hard number in the section, and it is a real-time measure of leveraged positioning โ of crowding, not of value. When funding runs persistently positive, longs are paying to stay long. In a euphoric tape, that figure tells you how much pain is queued, not how much upside remains.
Governance is thinner. Voting participation, top-ten concentration, proposal quality. Participation is meaningless without a denominator โ turnout says nothing without the quorum threshold and the delegation distribution. A DAO where five delegates hold veto power and turnout is eighty percent is less decentralized than one where turnout is twelve percent and no one holds a bloc.
The scoring layer. Star ratings, one to five, across technical value, investment value, timeliness, and reference value. All five stars empty here.
A five-star system with no rubric is a review score with no reviewer.
I built something adjacent in 2026 โ a Technical Viability Score for AI-crypto projects. The difference is that mine had a denominator. Computational overhead per verified inference, measured against a baseline oracle. Latency in milliseconds. Error as mean absolute deviation from ground truth. When I ran the prototype, the numbers came back ugly for high-frequency use. ZK proof generation swallowed the entire latency budget. That is a score. It has units.
Stars have no units. Stars have moods. Code is the only law that compiles without mercy.
The contrarian read
Everyone who sees this artifact will call it a failure. A pipeline that produced nothing. Empty grids. A wasted run.
I think it is the best crypto research document I have read this quarter.
Because the alternative โ the filled-in version โ is the thing that actually damages people. Same nine sections. Same borrowed erudition. Same risk matrix. But with confident cells: Medium, High, Moderate, 3/5 stars, Attractive entry. Generated from the same thin source data. Rendered without the gate. That document gets screenshotted. It gets quoted. It becomes the seed of a position.
The honest N/A is a system that knows its confidence is below threshold. The confident report is the same system with the guard removed. Same data. Different behavior. One of them fails safe.
There is a second-order effect too, and it is the part that bothers me. The existence of this template tells you what the market is buying. Nine standardized sections means nine standardized reports means thousands of assets reviewed with an identical instrument, where the buyer's only signal is a star count. That is not research. That is a rating agency with worse incentives and no liability.
I have argued before that liquidity fragmentation is not a real problem โ it is a manufactured narrative that exists to sell new products. The same mechanism is at work here. The fragmentation is not in the liquidity. It is in the attention. Each of these reports is engineered to be skimmed in ninety seconds, which means the structure that makes them comparable is the same structure that makes them interchangeable.
A template that fits everything measures nothing.
What I am watching next
The pipeline that aborts. Not the one that aborts loudly โ the one that aborts and then tells you what it needed. This artifact did that by accident. Title. Information points. Core thesis. Project names. Source quality. Data. The required-fields list is not an error message. It is a specification. It is a demand for input.
The next generation of research tooling will be graded not on how many reports it can produce, but on how many it refuses to. The refusal rate is the confidence interval. In a bull market, where every protocol ships a deck and every deck ships a token, the analyst who says "insufficient information" is the only one whose "sufficient" means anything.
Code is the only law that compiles without mercy. So does a null check.