The framework returned nine dimensions. All nine read N/A.
Technical positioning: insufficient information. Tokenomics: insufficient information. Market cycle: insufficient information. Ecosystem position: insufficient information. Regulatory exposure: insufficient information. Team and governance: insufficient information. Risk surface: not assessable. Narrative: insufficient information. Supply-chain transmission: insufficient information.
The deliverable ran to roughly 1,900 words and contained exactly zero assertions about any asset, because the information-point list feeding it contained exactly zero items. Nine empty cells, one blocking statement, three probable causes ranked by likelihood, and a request for the original source.
Most desks read that as an analyst failure. It is not. It is a pipeline failure that the report diagnosed before it wrote a single line of analysis. The ordering is the point. A null result published before the framework is filled in is a finding. A null result published after the framework is filled in is an excuse.
I have watched this industry for 29 years. The most common corruption of research is not bad data. It is confident conclusions drawn from absent data. In the absence of data, opinion is just noise — and noise, once formatted into a table with headers and footnotes, becomes indistinguishable from evidence to everyone downstream who did not build the table.
The pipeline nobody audits
Institutional crypto research in 2026 runs on pipelines, not people. Crawl, parse, extract, analyze. Four stages, each with its own failure mode, each usually monitored by nobody.
The demand side has not adjusted. A portfolio manager asks for deep analysis and assumes depth is a property of the output. It is not. Depth is a property of the input. When the payload is empty, a competent pipeline has exactly two behaviors available: refuse, or hallucinate. Hallucination is cheaper. It is also invisible, because a model asked to produce nine dimensions will produce nine dimensions, in clean prose, with plausible hedging, and no one will check whether any of it traces back to a document.
There is a second-order problem. Pipelines are built by engineers and consumed by analysts. The engineers monitor uptime and error codes. The analysts monitor narrative and valuation. Nobody monitors whether the output's evidence density is above zero, because evidence density is not a metric anyone is paid to track.
This framework had a traceability constraint hard-coded into its instructions: every conclusion must cite an information point (→ basis: [point N]). That single rule is the entire design. It makes the report falsifiable. It also means that when the point set is empty, the framework has precisely one legal output — refusal — and any other output is a violation of its own specification. The source of truth is the point list, not the prose.
The user asked for depth. The input contained nothing. That mismatch is the event worth analysing.
Three failure classes, and the fourth one missed
The report ranked three probable causes. I will reproduce them, then add the one it omitted.
Class 1 — crawl-layer failure. The original content (URL, pasted text, PDF) was never successfully retrieved. The decomposer then ran on an empty input and emitted template defaults. Highest probability in most pipelines.
Class 2 — parse-layer mismatch. Content was retrieved, but the format did not match the parser's logic: scanned image, JavaScript-rendered page, a script the tokenizer does not handle. Every field parses to empty even though the bytes arrived.
Class 3 — transmission truncation. Stage one produced real output, but fields were dropped in the handoff, leaving a skeleton. The tell is that structure survives while content vanishes.
The discriminator is a single test: did stage one return even one real information point? If the title itself is blank, you are looking at Class 1 or 2. If a title exists but the point list is empty, you are looking at Class 3.
Now the omission. Class 0 — no source was ever specified. The triage list assumes a source existed and failed somewhere downstream. But a pipeline cannot fail to fetch a document that was never named. In this case the payload was empty at the point of request: no title, no URL, no text, no publication date. The report's own diagnostic table shows every field blank, including the ones a crawl failure would have partially populated.
Class 0 is not a bug in the crawler. It is a bug in the intake contract. And it is the most common of the four, because intake is the only stage typically owned by a human under time pressure.
The arithmetic of zero
There is a small piece of statistics that research desks consistently ignore. With zero observations, you cannot compute a mean. Any point estimate has unbounded variance. A confidence interval on an empty sample is not wide — it is undefined.
The report handled this correctly, and the handling deserves to be stated plainly because most rating systems get it wrong. It did not score the asset one star out of five. It marked each dimension not assessable, and separated that from assessable and poor. Those are different states. A one-star rating is a claim about a project. A not-assessable rating is a claim about the analyst's evidence base. Collapsing them is how bad assets get laundered as merely unproven, and how good assets get buried as merely undocumented.
The guard clause that should sit at the top of every research pipeline is four lines long:
def analyze(info_points: list[dict]) -> Report:
if not info_points:
return Report.blocked(
reason="EMPTY_INPUT_SET",
allowed_conclusions=[],
next_step="request_source_material",
)No model call. No temperature. No creativity. The absence of a guard clause here is the same class of defect as a missing bounds check in a smart contract: the code runs, the output looks valid, and the failure surfaces three layers later where it is expensive to trace.
Minimum viable evidence
The framework's nine dimensions are useful precisely because they can be inverted into an evidence checklist. For each dimension there is a floor below which analysis is fabrication. I keep this list taped to my monitor.
| Dimension | Minimum viable evidence | |---|---| | Technical | Repository or specification; testnet/mainnet status | | Tokenomics | Supply schedule plus allocation table with vesting | | Market | Price series and venue list; funding data if perps exist | | Ecosystem | At least one dependency edge, upstream or downstream | | Regulatory | Issuer jurisdiction and entity structure | | Team | One verifiable identity, or one named lead investor | | Risk | Contract address or audit report | | Narrative | Dated source of the claim being circulated | | Transmission | A named counterparty or liquidity venue |
If a dimension has no entry, it returns not-assessable. Not zero. Not monitor. Not team appears anonymous but promising. Blank.
I built that habit in 2017, auditing a project that promised 1,000% APY. Six weeks of modelling their liquidity pools against securities law produced one finding that mattered: 40% of the token supply was unvested and unscheduled. The report was not elegant. It was a table. The project was delisted from local exchanges shortly after. The lesson was not that I was right. The lesson was that the table did the arguing, and the table had a basis line for every row.
Two vacuums, not one
The report made a distinction that most readers will skim past, and it is the most important sentence in the document: the information vacuum here comes from the analysis pipeline, not from the analysed object. These two are entirely different in nature and cannot be conflated as risk evidence.
This is where institutional research quietly destroys itself. A missing contract address in your own intake form is an operational defect. A missing contract address in a project's public disclosure is a risk signal, weakly correlated with anonymous teams, unaudited code, and templated whitepapers. The two look identical in a spreadsheet. They are not the same variable, and treating the first as the second manufactures risk where none has been demonstrated, while treating the second as the first buries risk that has.
In 2022 I spent three days pulling transaction-level data from Terra's explorers while the peg was failing. The evidence existed. It was public, dense, and unread. The failure there was not an information vacuum — it was a refusal to read the information that was present, because reading it required admitting the seigniorage mechanism depended entirely on new speculative demand. That is the opposite pathology, and it is more dangerous than a blank field.
What the bulls get right
Here is the counter-intuitive part, and I will give the optimists their due.
A structured null is cheap. A hallucinated thesis is expensive. One pipeline cycle versus one position. The desks that publish refusals will underperform in any given quarter on volume, and outperform over any full cycle on drawdown, because their errors are visible and correctable while everyone else's errors are embedded in prose that reads well.
The framework's traceability rule is also the reason the null report can be trusted at all. It is trusted not because it is right about the asset, but because it is checkable about itself. Every claim carries a basis line. When there is nothing to cite, the framework says so. That is a rarer virtue than accuracy.
The blind spot is incentive. The market does not pay for nulls. A fund that says we do not know loses the mandate to a fund that says we know, and the second fund's model has the same empty input and a lower refusal threshold. Publication bias in crypto research is total: completed analyses get published, refusals get filed. Any desk that samples only published research is sampling a market where 100% of assets were analyzable, which is not a random sample of anything.
Takeaway
Within eighteen months, I expect evidence sufficiency to appear as a scored field in institutional research memos — a number between zero and one, sitting above the recommendation, not below it. The 2025 custody and reporting frameworks I worked on for an Australian bank already require an audit trail from conclusion back to source. An AI-generated analysis with no traceable evidence chain fails that requirement by construction, and it will fail it in front of a regulator rather than a colleague.
That is not a technology problem. It is a governance problem, and governance problems do not resolve themselves when the model gets better.
So the question is not whether your pipeline can produce nine N/As. It can. The question is who in your organisation has the authority to publish them, and whether that person is rewarded or quietly reassigned.