A blank page is a signal. A filled page of N/A is a sedative.
On a Tuesday morning I opened a document that a research pipeline had produced overnight. Four thousand words. Nine analytical dimensions. A six-row risk matrix. A four-element Howey test table. A token supply schedule broken into team, early investors, community and treasury. A competitive landscape grid. A transmission map running from upstream infrastructure to downstream applications.
Every cell said the same thing.
"N/A โ insufficient information."
Not empty. Not blank. Formatted. Aligned. Ruled. The team row had a border. The Howey table had a verdict line. The risk matrix had a composite rating, and the composite rating was "cannot be assessed," printed in the same typeface, at the same weight, in the same column where a real report would have printed "Low."
Then the confidence labels. That was the detail that made me put down the coffee. The document attached "[Confidence: High]" to statements about its own ignorance. High confidence that it could not infer. High confidence that there was no basis for inference. The system was certain about nothing, and it had graded its certainty.
The extractor upstream had returned zero information points. The analyzer downstream had been asked to produce a nine-dimension report anyway. It complied. It built the scaffolding first and looked for facts second. When the facts did not arrive, it shipped the scaffolding.
Somewhere in a compliance folder, that document now exists as evidence of diligence.
Structure precedes profit; chaos demands a fee. But structure without input is not diligence. It is a receipt for a broken pipeline, printed on the same paper as the audit.
Context: What Actually Broke
Let me be precise about the failure, because the sloppy version of this story is "the AI hallucinated," and that is not what happened.
The pipeline has five layers. Fetch, parse, extract, analyze, render. Each layer has its own failure signature, and the signatures look almost identical from the outside.
Fetch fails when the source sits behind a paywall, or behind JavaScript that renders the body after the crawler has already closed the socket, or behind a region block, or behind a login wall on a research portal that charges four figures a month. Fetch failure returns an empty string and an HTTP status that nobody logs.
Parse fails when the document is a PDF with text baked into images, or a Notion page with content in a database view, or a Telegram announcement channel where the substance lives in a pinned message from eleven months ago, or a project's own documentation site where the whitepaper is a Google Doc that requires a request. Parse failure returns a DOM tree with headers and no paragraphs.
Extraction fails when the parser produced text but the extractor's schema expects fields the text does not contain. A funding round announcement rarely contains a token unlock schedule. A token unlock schedule rarely contains a team's legal domicile. Extraction failure returns a set of null fields that are indistinguishable, in most schemas, from fields that were populated and found empty.
Analysis fails when it is handed nulls and instructed to produce output anyway. That is the layer that produced my Tuesday document. It did not invent facts. It invented structure around the absence of facts, which is worse, because invented facts are detectable and invented structure is not.
Render fails last and fails silently. The renderer draws a table because the schema says a table exists. It fills cells because empty cells break the layout. It writes "N/A" because a null in a Markdown table renders as a blank that looks like a formatting bug rather than a data gap.
I have run a version of this stack since 2017. My team measures its own failure rates, because an unmeasured pipeline is a rumor with a cron job. Over the last eight quarters, on a corpus of roughly 14,000 crypto source documents, our fetch layer has failed on 6.1 percent of targets, the parse layer on 4.4 percent of successfully fetched targets, and the extraction layer has returned zero usable information points on 3.8 percent of successfully parsed targets. Multiply the survivors and roughly one in seventy-five documents arrives at the analysis layer with nothing in its hands.
One in seventy-five. Run a hundred documents through the stack and one or two of them will be ghosts. In a bull market, when content volume triples and every fund, exchange and protocol is publishing daily, that number is not a rounding error. It is a stream of confidently formatted documents describing assets nobody actually read about.
The dangerous failure mode in crypto research is not the wrong answer. It is the right answer to a question that received no input.
Now consider what the market looks like from the other side of the table. We are in a bull market. The volume of published research has exploded, the turnaround time has collapsed, and the number of people who read a report end-to-end before acting on it is close to zero. Most institutional readers consume the executive summary, the risk rating, and the position sizing recommendation. All three of those artifacts survive an empty input perfectly intact. The rating column still has a rating in it. It just says "cannot be assessed" in the same font as "Low."
The nine-dimension template itself is not the villain. It exists because institutional diligence requires coverage: technical, tokenomic, market, ecosystem, regulatory, team, risk, narrative, transmission. A checklist that forces you to look at nine things is better than a vibe. My 2017 audit protocol was a checklist. My 2020 liquidation bot ran on a checklist. Checklists are how you convert judgment into something a team can execute without you in the room.
The villain is the missing fourth state. Every schema I have ever inherited has three states: populated, empty, and error. Real pipelines need four: populated, verified-empty, unobserved, and error. "Verified-empty" means we fetched the document, we parsed it, and the document genuinely does not discuss a token supply schedule. That is information. "Unobserved" means we never got the document. That is not information. It is the absence of information, and it must not render as a value.
Collapsing "unobserved" into "empty" is the single most expensive schema decision in this industry. It converts a plumbing failure into an analytical conclusion, and it does so at machine speed, at scale, with a footnote that says the analyst was thorough.
Core: The Asymmetry Principle, and Why N/A Is Not Zero
Start with the principle that governs everything downstream.
In risk analysis, absent data is not neutral. Absent data is maximum uncertainty. The default posture toward a null is not zero risk. It is the highest available risk band, held open.
This is not conservatism. It is arithmetic. Every risk model you will ever build is a conditional expectation: given information set I, the expected loss is E[L | I]. When I shrinks toward the empty set, the conditional expectation does not collapse toward zero. It collapses toward the unconditional expectation, which for a newly issued token in a bull market is somewhere between brutal and total. The unconditional distribution of outcomes for an unexamined crypto asset is not symmetric. It is a barbell: a long thin tail of enormous returns and a fat body of zeros. If you do not know which bucket you are holding, the correct prior is the fat body.
My Tuesday document did the opposite. It treated every null as a deferral, which readers interpret as a pass. Six rows in a risk matrix said "cannot be assessed," and six rows of "cannot be assessed" read, to a human under time pressure, as "nothing flagged."
That is the mechanism. It is not a bug in the model. It is a bug in the interface between the model and the human.
The Rendering Problem
In February 2022 I watched a dashboard render a price feed as 0.00.
The feed had died. The oracle had not updated in nine minutes. The ingestion service had caught the exception, written it to a log nobody tailed, and passed a null downstream. The charting library received the null, applied its default coercion, and drew a zero. Zero is a number. Zero has a shape. Zero drew a line straight down, and the line straight down looked like a crash, and the desk that owned that dashboard spent forty minutes reacting to a crash that had not happened while the actual position sat unhedged.
Three months later, in May, a different desk watched a different dashboard render a different null as a flat line. Flat, not zero. Their charting library coerced null to the last known value, which is a perfectly reasonable engineering decision if you are drawing a sparkline on a marketing page and a catastrophic one if you are drawing a liquidation threshold. Their feed had been stale for six hours. Six hours of flat green. The model did not flag the anomaly because the model never saw an anomaly. It saw continuity.
The Terra collapse did not catch those desks by surprise because the math was hard. The math was not hard. The math had been flagging the imbalance for days. Those desks were caught because their data layer had degraded and their renderer had disguised the degradation as normalcy. Zero and null and stale-last-value all drew as lines.
Code executes what words promise. A renderer that promises a number will produce a number, even when the honest output is a blank.
I rebuilt our charting layer in the third quarter of 2022 and the rebuild had one non-negotiable requirement: null renders as a gap. A visible hole in the series. A grey band across the region where data should be. Not a zero, not a carry-forward, not an interpolation. A hole, with a timestamp label, and a tooltip that says the feed is stale and the reason it is stale.
Complaints came in for two weeks. Traders hated the gaps. Gaps are ugly. Gaps make a chart look unfinished. Gaps force you to notice that your information supply has a hole in it at 03:40 on a Saturday when you are trying to size a hedge.
That is the point. A chart that never has holes is a chart you have stopped reading.
The same rebuild applies to research. If your document has no gaps, it has no honesty. A nine-dimension report where all nine dimensions have content is either a genuinely complete dataset โ rare, and usually expensive โ or a renderer that has been filling holes with the word "N/A" and calling the layout finished.
Four Tiers of Claim, and the Fifth One Everybody Omits
I run every research output through a four-tier labeling discipline. It is the same discipline I imposed on my team in 2017 and it has not changed, because the failure mode it prevents has not changed.
Tier one: explicitly stated. The source document says the team allocation is 18 percent with a twelve-month cliff and thirty-six-month linear vesting. I can quote it. I can link it. If it is wrong, the source is wrong, and the source is on the record.
Tier two: reasonable inference. The source says the round was led by a fund whose standard term sheet includes a twelve-month lockup, and the raise closed in March. I can infer a lockup expiry around the following March, and I label it as inference, with the reasoning attached.
Tier three: high speculation. The team is anonymous, the code is unaudited, the raise was large and the token is not yet trading. I can construct a plausible narrative about what happens at listing. I label it speculation and I do not size a position on it.
Tier four: not observed. We never obtained the document. We do not know whether the supply schedule exists, whether it is public, or whether it contradicts the marketing page. We know nothing.
Here is the fifth tier, the one every template omits: we did not look.
"Not observed" and "did not look" are different failures and they have different fixes. "Not observed" is a fetch problem: paywall, JavaScript, region block, PDF. Fix the fetcher. "Did not look" is a scope problem: the analyst had a deadline, the template had a slot, the slot got filled with a plausible-sounding phrase because the alternative was leaving it empty. Fix the incentive.
My Tuesday document had no tier-five category. It had nine dimensions and no way to say "nobody on this desk ever attempted to retrieve the source." So it said "insufficient information," which sounds like the information was sought and not found, when the truth was that the request never reached the network.
That distinction is worth real money. A fetch failure is a ticket for the data engineering team. A scope failure is a conversation about staffing. A document that renders both as the same grey cell has made the ticket and the conversation invisible, and it will produce the same failure next Tuesday, and the Tuesday after that.
The Schema: Making Emptiness Machine-Readable
I do not fix this with exhortation. I fix it with fields. Exhortation does not survive a deadline. A required boolean survives a deadline, because the pipeline will not run without it.
Our research schema, version six, carries the following on every document object:
input_provenance: {
fetch_status: OK | PAYWALL | JS_RENDER | REGION_BLOCK | PDF_IMAGE | TIMEOUT,
fetch_bytes: integer,
parse_status: OK | NO_BODY | PARTIAL,
parse_char_count: integer,
extract_status: OK | ZERO_POINTS | SCHEMA_MISMATCH,
info_point_count: integer,
input_empty: boolean, // true if info_point_count == 0
human_reviewed: boolean,
reviewer_id: string | null
}Two fields do the heavy lifting. info_point_count makes emptiness a number, and numbers are auditable. input_empty makes emptiness a gate.
When input_empty is true, the analysis layer refuses to emit a nine-dimension report. It emits a one-page incident record: what we tried to fetch, what the fetch returned, which layer failed, and who owns the fix. The nine-dimension template is not available. The renderer has nothing to fill.
This felt excessive when I proposed it. It has since prevented, by my count, thirty-one malformed reports from entering our distribution list. Thirty-one documents that would have carried a risk rating of "cannot be assessed" into a folder where a portfolio manager reads the rating column and nothing else.
Thirty-one is not a large number in absolute terms. It is a large number when each one is a potential allocation.
The 2017 Protocol: Absence Maps to Rejection, Not to Neutral
I learned this in the ICO bubble, and I learned it the hard way, which is the only way these lessons stick.
Late 2017. My team was screening whitepapers at a rate that made individual diligence impossible โ we were seeing forty-plus per week at the peak, most of them structurally identical, most of them promising a token that would settle payments for a marketplace that did not exist. I built a standardized audit checklist and applied it to more than forty papers in the final quarter.
The checklist had a design decision that I did not appreciate at the time. Every item was binary. Pass or fail. There was no "insufficient information" row. If a whitepaper did not state a token supply, it failed the supply item. If it did not state a vesting schedule, it failed the vesting item. If the claimed market size was arithmetically impossible against the total addressable market of the sector, it failed the sanity item.
The papers hated it. The salespeople hated it more. "That information is coming in the next version," they said. "That is a documentation gap, not a red flag." Every one of them said some version of the same sentence, and every one of them was asking me to map absence to neutral.
Twelve of the forty-odd papers failed on mathematical grounds alone. Circulating supply times price implied a fully diluted valuation larger than the GDP of the country the team was incorporated in. Token velocity assumptions required a transaction volume that exceeded the entire payment volume of the sector by two orders of magnitude. One paper's bonding curve, evaluated at the stated parameters, produced a negative price at the tenth purchase. Not an expensive price. A negative one.
I did not have a category for "the math is unclear." I had a category for "the math does not close." Those twelve went into the reject pile, and the reject pile kept my firm out of roughly one and a half million dollars of subsequent losses when the market broke in the first quarter of 2018.
The lesson was not that I was smart. The lesson was that the checklist refused to be neutral. Absence was a fail, and the fail was cheap, and the fail was early.
Compare that to the Tuesday document. Six rows of "cannot be assessed." A human reading that document has to supply the rejection themselves. The document has outsourced its most important function โ the decision โ back to the reader, while retaining the appearance of having performed an analysis.
A checklist that permits neutrality is not a checklist. It is a form.
The 2020 Engine: Halt as a First-Class State
The liquidation bot taught me the engineering version of the same lesson.
We ran it through DeFi Summer on Aave V1, and across a single quarter it processed north of fifty million dollars in bad debt. The interesting number is not the fifty million. The interesting number is the false positive rate.
Our first build guessed. When the oracle feed went stale, the bot computed a health factor from the last known price and either liquidated or did not. Both branches were wrong. Liquidating on a stale price meant closing a healthy position at a bad number and eating the dispute. Not liquidating meant missing a genuinely underwater position because the feed had been down during the exact window when the collateral collapsed.
The fix was not a better oracle. The fix was a third state. When the price feed exceeded its freshness threshold, the bot entered HALT_STALE_FEED and emitted an alert instead of a transaction. It stopped guessing. It told a human that it did not know, and it told the human exactly which input it did not know about and how long it had been missing.
Standardizing that logic โ halt on stale, halt on disagreement between two independent feeds beyond a threshold, halt on any position whose collateral asset had fewer than N observations in the trailing window โ reduced false positives by fifteen percent against the community-built tools we benchmarked against. The gain did not come from more data. It came from the engine's willingness to declare that it lacked data.
When the market corrected and everyone else was improvising, our engine was executing a protocol it had already rehearsed. Not because we were faster. Because we had pre-committed to what we would do when we did not know, and the answer was: stop, escalate, log.
Survival is a function of liquidity, not optimism. But the liquidity that matters most is the liquidity of your decision-making โ your ability to exit a position when the information that justified it evaporates.
Every research pipeline needs a HALT_STALE_FEED. The Tuesday document was a system that had no halt state, so it did the only thing it could do. It ran.
The 2024 Gap: Primary Data Is Where the Edge Lives
Here is the part that connects plumbing failures to actual alpha.
In 2024 I led a quantitative review of the newly approved spot Bitcoin ETF structures, comparing fee models and custody arrangements across five major issuers. The published fee schedules were identical to two decimal places in most comparisons. Everyone had read the fee schedules. The fee schedules were in the marketing material.
The gap was in settlement timing. When I pulled custodian timestamps โ not the summary documents, the raw timestamps from the creation and redemption records โ a consistent 0.05 percent efficiency difference appeared between issuers in how quickly the underlying moved relative to the share creation. It was small. It was systematic. It was invisible in every summary that had been aggregated from those timestamps.
That difference funded a high-frequency arbitrage strategy that produced roughly two hundred thousand dollars a month in alpha for the desk.
Now hold that next to the Tuesday document. The edge in 2024 came from reading primary data that nobody had aggregated. The Tuesday document came from a pipeline that had failed to fetch primary data and had rendered the failure as a table of N/A.
These are the same problem viewed from opposite ends. Arbitrage finds truth where noise ignores it โ and the first thing noise ignores is the provenance of its own inputs. If your research process cannot tell you whether it actually read the custodian timestamp or merely read a summary of a summary, you are not doing diligence. You are doing decoration.
That is also the quiet tragedy of the Bitcoin ETF era. The asset that was designed to settle peer-to-peer now settles inside a creation-and-redemption cycle timed to the New York close, arbitraged by desks that read custodian logs at microsecond resolution. The instrument survived. The intent did not. And the desks that captured the difference were the ones that went one layer below the published document โ into the timestamps, the settlement windows, the small print that no aggregator bothers to parse.
Which means: every time your pipeline fails to fetch and quietly substitutes a null, you are not just missing a data point. You are forgoing the exact layer where the inefficiency lives.
The 2026 Agent: Explainability Means Being Able to Say Nothing
Last year I integrated AI-driven sentiment analysis into the quant stack. I rejected black-box models. Not on principle โ on operational grounds. A model whose reasoning cannot be reconstructed cannot be defended to a compliance team, and a strategy that cannot be defended to a compliance team cannot be scaled.
What I built instead was a transparent decision-tree layer sitting on top of the model's outputs. The model produces a sentiment score and, critically, a coverage score: what fraction of the expected source set did it actually see. I trained the whole thing on ten years of my own P&L, which means the risk parameters it inherited are the ones that survived 2018, 2020, and 2022 rather than the ones that look good in a backtest.
The hybrid approach lifted win rates twelve percent while keeping full explainability. The twelve percent is nice. The explainability is the point. Because the explainability is what let me impose a rule: when the coverage score falls below threshold, the agent's output is suppressed and the desk gets an alert.
The agent is allowed to say nothing. That is a design requirement, not a courtesy.
Most deployed agents are not allowed to say nothing. They are prompted to produce a summary, and a prompt that demands a summary will receive a summary, because the model's job is to satisfy the prompt. If the retrieval layer returned three irrelevant documents and one paywall error, the model will write a confident paragraph about the three irrelevant documents. It is not lying. It is completing an assignment.
The human-in-the-loop question is not "what is the model doing." It is "what is the human responsible for." And a human cannot be responsible for a decision the model was structurally forbidden from declining to make.
The Compliance Exposure Nobody Priced
Now the part that should worry you more than the research quality.
Look again at the Howey table in the Tuesday document. Four rows: money invested, common enterprise, expectation of profit, efforts of others. Four rows, four N/A cells, and a composite verdict of "cannot be assessed."
In a jurisdiction that regulates by enforcement rather than by published rule, that verdict is not a neutral observation. It is a liability.
The SEC's posture over the last several years has not been confusion about the technology. The pattern is more specific than that. Enforcement actions arrive against entities that operated in areas where the published guidance was thin, and the settlement terms are negotiated by parties whose internal documentation either supports their good-faith analysis or does not. A firm that can produce a documented, dated, methodologically consistent securities analysis has something to negotiate with. A firm that can produce a template with four N/A cells and a note saying the jurisdiction was not identified has produced evidence that it did not conduct the analysis at all โ and did it in a formatted, machine-generated document that looks like the output of a process.
The absence of a jurisdiction row is not a defense. It is the prosecution's exhibit.
This is the practical cost of the missing fourth state. When "unobserved" renders as "empty," the compliance file cannot distinguish between "we looked at the token terms and concluded the distribution was not an investment contract" and "our crawler hit a Cloudflare challenge." Those are radically different positions in front of a regulator, and they are the same grey cell in the report.
My rule for compliance-facing outputs is blunt: any dimension with an unobserved input is escalated to a human before the document leaves the building, and the human's name goes in the reviewer_id field. Not because the human will necessarily know the answer. Because the human must be accountable for the gap, and accountability requires that someone be forced to look at it.
Contrarian: The Scarcest Resource Is the Willingness to Publish Nothing
The consensus in 2026 is that AI has democratized research. Anyone with a model and a prompt can produce a nine-dimension report on any token in four minutes. Coverage is no longer a moat. Analysis is no longer a moat. The marginal cost of a well-formatted document has gone to zero, and the market has responded exactly as you would expect: the supply of well-formatted documents has gone to infinity and their information content has gone to zero.
Everyone draws the wrong conclusion from this. They conclude that the edge is now in better models, or proprietary data, or faster inference.
The edge is in the willingness to publish nothing.
Think about the incentive structure on a research desk. An analyst who writes three thousand words and concludes "this asset cannot be assessed on available information" is, in most organizations, a failed analyst. The document has no recommendation. It cannot be forwarded to an investment committee. It generates no meeting. An analyst who writes three thousand words with a nine-dimension template, a risk rating, and a position sizing recommendation is a productive analyst, regardless of whether the underlying input was a press release, a rumor, or a fetch timeout.
The output looks identical. The incentive cannot tell them apart. So the incentive selects for the second one, systematically, every quarter, forever.
That is why the Tuesday document exists. Not because a model malfunctioned, but because a process was designed by people who are rewarded for filling templates and punished for leaving them empty. The model simply learned the local gradient.
Here is the second-order effect, and it is the part that should genuinely unsettle you. The Tuesday document is more honest than most published research. It says, in every cell, that it does not know. The average sell-side note on a newly listed token says the opposite: nine dimensions, all filled, all confident, sourced from a project deck, a Twitter thread, and a spreadsheet with a circular reference in row fourteen.
A report that says "I don't know" forty times is auditable. A report that says "strong fundamentals, favorable tokenomics, experienced team" is not auditable, because there is nothing to check. The confidence is unearned and the sources are unfalsifiable.
So the contrarian position is not "automate less." It is this: the correct output of a diligence process is often a refusal, and a system that cannot refuse is not a diligence process. It is a content generator with a compliance wrapper.
There is a retail-versus-desk dimension here that nobody writes about because it is unflattering to everyone involved. Retail reads the headline. Desks read the provenance. Retail asks "what does the report say about this token." A desk asks "what did the report actually read, and when, and who signed it." These are different documents. One of them is a narrative artifact. The other is a chain of custody.
The gap between those two readings is where the money is made and lost. Retail bought the narrative in 2021 and 2022 and paid for it. The desks that survived had, somewhere in their stack, a field that said human_reviewed: true and a person whose name was attached. Not because they were smarter. Because their process forced a human to look at the gap before the capital moved.
And a final, uncomfortable note on the bull market we are in. Euphoria is the ideal environment for this failure mode, because euphoria suppresses skepticism about process. Nobody wants to hear about extraction rates when the chart is up and to the right. Nobody audits the fetch layer when the narrative is working. The failure gets discovered at the bottom, when the desk tries to reconstruct why it held what it held, and finds a folder of documents that all say "cannot be assessed" in a font that looks like diligence.
Takeaway: Three Rules and a Question
I am not going to tell you to read your research more carefully. That is advice, and advice does not survive a deadline. I am going to give you three things you can implement this week.
One. Add the boolean. Whatever your pipeline is, put input_empty on every document object and wire it to a gate. When it is true, the analysis layer does not run. It emits an incident record with the failing layer named and an owner assigned. This costs you an afternoon of engineering and it eliminates the entire class of failure I have described. The document cannot be produced. That is the point. A control that depends on a human remembering to check is not a control.
Two. Make null render as a gap. Not zero. Not carry-forward. Not a blank cell that looks like a formatting bug. A visible hole with a label and a reason. Then accept that your dashboards will be uglier and your reports will have empty sections, and understand that the ugliness is the control. A chart with no holes is a chart you have stopped reading. A report with no gaps is a report nobody verified.
Three. Map absence to rejection, not to neutral, in every screening decision. This is the 2017 lesson and it is the only one that has never failed me. If the vesting schedule is not stated, the project fails the vesting criterion. If the jurisdiction is not identified, the compliance review halts. If the custodian timestamps cannot be pulled, the arbitrage is not sized. You will reject some good projects on documentation grounds. That is the price of never accepting a bad one on the same grounds, and the price is cheap, because the rejects are early and the accepts are irreversible.
Now the question I want you to carry into next week.
Open your inbox. Count the research documents you received in the last seven days. Now ask, for each one, a single question: what did the extractor actually retrieve, and does the document say so anywhere?
Most of them will not say. Most of them were generated from a pipeline that has no fourth state, rendered by a template that fills cells because empty cells break the layout, and read by a human who looked at the risk rating and the position size and skipped the methodology note.
The market respects discipline, not desire. And the discipline that matters most is not the discipline to analyze. It is the discipline to return an empty page and defend it.
A blank page is a signal. Somebody will read it. That is more than you can say for the nine-dimension report with forty N/A cells, sitting in a compliance folder, looking exactly like diligence.