Last Tuesday, at two in the morning Chengdu time, I opened an analysis pipeline and found a table of nulls.
Nine dimensions. Every field empty — no title, no source, no claims, no information points, no jurisdiction, no counterparty. The system had not crashed. It had not hallucinated a substitute. It had stopped, and it had told me it stopped, in plain language, across page after page of honestly blank tables.
I have read machine-assisted research for most of a decade. I had never seen a document quite like it: an engine that refused to lie.
My first reaction was not admiration. It was irritation — the reflex of anyone who has shipped against a deadline. Nine empty tables is not a deliverable. Nine empty tables is a shrug with formatting. I closed the file, made tea, and opened it again at four in the morning, because something in it kept pulling, and the thing pulling was not the emptiness.
It was how rare the emptiness was.
This industry produces abundance. We have never been worse at proving where any of it came from.
The bear market has turned certainty into a consumer product. Nobody wants a treatise on mechanism design right now; they want to know whether the thing holding their savings is still solvent next quarter. That want is legitimate. It is also being met, aggressively, by a research economy that has never been larger, cheaper, or harder to trace. Liquidity leaves protocols and does not come back. The commentary around that departure multiplies like sediment. Over the past two years I have watched the ratio of published analysis to verifiable primary data widen into something obscene.
I learned the discipline of provenance in 2017, at thirty-three, as a senior strategist at Polymath, drafting a forty-page whitepaper on tokenized equity as digital citizenship. I spent weeks in rooms with securities lawyers, and what I remember from those weeks is not the economics. It is the citation discipline. Every claim had a parent. Every number could be walked backward to a filing, a statute, a recorded conversation with a named person on a named date. When I could not find the parent, the claim came out of the document. Not softened. Removed.
That habit is what provenance means, and it is not a legal formality. It is the entire difference between a document and a rumor.
Blockchain people should understand this better than anyone, and mostly we do not. The reason a chain is trustworthy is not that the data on it is good — a great deal of on-chain data is garbage, wash-traded, spammed, self-dealt. The reason is that you can verify how the data got there: which address signed, in what order, against which prior state. Trustless does not mean trustworthy. It means auditable. The audit trail is the product.
What we call research in crypto has almost none of that. A thread cites a screenshot. The screenshot cites a dashboard. The dashboard aggregates an API whose upstream source was, at some point, a message in a channel. And at the end of that chain sits a reader allocating money they cannot afford to lose.
The null-field pipeline sat at the opposite end of that spectrum. It had been handed nothing, and it returned nothing, and it documented the nothing with the same rigor it would have applied to a finding.
When a data pipeline meets an empty input, it has exactly three options, and I have audited enough of these systems to say that two of them are catastrophic.
The first is silent truncation. The upstream fetch fails, the field resolves to an empty string, and the empty string is rendered as a fact. I have seen liquidation engines do this with price feeds, which is where the lesson becomes expensive: a missing price is not zero, and code that cannot distinguish them will happily close a solvent position at a price that never existed. The same failure in an analysis document is less dramatic and more corrosive. The reader sees zero. The reader believes zero.
The second is confident interpolation. The model — or the analyst, or the analyst using the model — notices the gap and fills it with the most plausible thing. Plausible is the operative word. Not verified. Not sourced. Plausible, which in a market that runs on narrative means expected. This is the failure mode that produces a beautiful, coherent, internally consistent document about a project whose team was anonymized three weeks ago.
The third is the honest null. The pipeline returns empty fields and a note explaining that the gap is upstream, not downstream. This is the only one of the three that costs the producer anything, because it looks like failure. It looks like failure to the client. It looks like failure to the audience. And if the producer's incentives run quarterly, it is failure.
The honest null is the most expensive output in this industry, and it is the only one worth anything in a downturn.
A null is a location. It tells you exactly where the map ends. A fabrication, by contrast, extends the map into blank water, draws a coastline there, and everyone downstream sails toward it.
We solved a version of this problem years ago and then forgot the solution. Oracle design has always understood the difference between an absent value and a null value, because the cost of confusing them was measurable in liquidations. A well-built feed reverts. It refuses to answer. It forces the consumer to handle the absence explicitly, at the call site, where a human decision still exists. That is an architectural choice with a moral shape to it: the system would rather break loudly than lie quietly. Every research operation in crypto should be built on that principle. Almost none are. We have built feeds that would rather lie.
Provenance fails in governance the same way, only there the null field is a person.
During DeFi Summer in 2020 I ran a governance working group at MakerDAO and read through more than five hundred voting proposals. What I found was not fraud. It was aggregation. The risk parameters were individually defensible; each had a rationale, a sponsor, a traceable line of reasoning. But the sum of them fell hardest on the smallest collateral holders — people whose positions were too small to justify the gas of a governance vote, too small to appear in any delegate's calculus, too small to be a constituency. They were not overruled. They were never counted, because the counting mechanism had no slot for them.
Algorithmic neutrality is often just the absence of anyone whose job is to notice. I wrote about that in an essay called The Quiet Collapse of Equity in Code. Fifty thousand people read it, and the reason it resonated was not the math. It was that I admitted a system I had helped design had a moral blind spot, and I named it, and I did not soften it with the word unintended. The blind spot was structural. The structure was mine.
That, too, is a provenance question. Before you can trust a governance outcome, you need to be able to audit who was structurally incapable of participating in it. Almost no DAO reports that field.
The same erasure happens to authors, and it is happening right now under legal cover.
The Tornado Cash sanctions marked a line a great many developers had already crossed in their heads long before any court said a word: writing code can be treated as conduct, and conduct can be charged. Whatever you think of the merits, the practical consequence inside the industry has been a quiet recoiling — contributors stripping their names from repositories, maintainers keeping private notes about which commits were theirs, people in their thirties asking lawyers whether a grant they accepted in 2019 is a liability now.
The record of authorship — the provenance of a piece of code — has become something individuals hide rather than publish. When provenance becomes dangerous, it does not disappear. It goes underground, where it can no longer be audited by anyone, including the people who might one day want to defend you. We are watching an industry that built its identity on verifiability decide, under pressure, that verifiability is a personal risk.
I do not have a clean answer. What I have is a strong sense that any regime which makes authorship itself the offense has not thought carefully about what authorship is: a claim about who did the work. You cannot prosecute a chain of custody without also destroying the only tool we have for holding anyone accountable for anything.
In 2021 I ran a small experiment in the opposite direction. I curated an invite-only DAO called The Ethereal Archive — one hundred and twenty members, no token, no roadmap, no public mint. Over three months I personally verified the artistic intent behind three hundred digital pieces: not the floor price, not the rarity tier, but who made the thing, when, with what tools, and whether the story attached to it survived contact with the actual files. Some did not. I removed them.
When the market collapsed in 2022, the archive held, and it held for a reason that had nothing to do with liquidity. Every piece in it had a documented parent. It was, as I wrote at the time, curating the soul in a world of derivative clones.
That phrase has aged into a thesis. Look at what the last two years produced: near-identical forks with new branding, near-identical PFPs with new palettes, near-identical research reports with new bylines. The mechanism never varies. Take something with a verifiable lineage, strip the lineage, keep the shape. A clone is not a copy. A copy preserves the original's provenance by reference. A clone deletes it.
Consider what happened to creator royalties. A system was built in which an artist's name lived permanently on-chain and their economics lived off-chain, in an enforcement policy any marketplace could switch off at will. When the largest venue softened its enforcement, the attribution survived and the value did not. Every artist's name is still there. The name is now a museum label on a wall from which the painting has been quietly sold.
Attribution without enforcement is decoration. That is the whole lesson of the PFP era, and it took five years and several billion dollars to learn. There is still no durable on-chain business model for creators. There is a provenance layer we built carefully and a payment layer we never built at all.
The same severing of label from substance is happening now with settlement layers. I have spent time with teams building what they call Bitcoin layer twos, and I have read documentation in which the settlement asset is a wrapped token issued on a different chain, custodied by a multisig, with a bridge operator set that has never once been described in a governance document. The word Bitcoin appears in the brand, in the marketing, in the fundraise deck. It does not appear in the trust model.
This is not a technical critique of those teams; several of them are competent. It is a provenance critique. When you remove the lineage and keep the name, you have manufactured a derivative clone with a premium ticker. The people who built this ecosystem over fifteen years mostly do not recognize these things as theirs, and that non-recognition is itself a data point — the kind you only get if you bother to ask the people who were there.
Here is the part I keep circling, and it is the part that makes me unpopular at conferences.
We keep asking how to make analysis more abundant. We should be asking how to make it more scarce.
The instinct to fill a blank field is not a bug in our tools. It is a perfectly rational response to our incentives. A published document gets read; a retraction gets read once. A confident call gets quoted; a we-do-not-know-yet gets scrolled past. Every mechanism in this industry — the token, the follow, the grant, the raise — pays for output and pays nothing for abstention. So we get output. Enormous, fluent, beautifully structured output, produced at a cost near zero, about subjects whose primary data was never fetched in the first place.
The bear market did not create this problem. It removed the price signal that was hiding it. In an uptrend, a fabricated thesis and a verified one both go up, and nobody can tell them apart. In a downtrend, the difference between a reader who knew which of their protocols had audited code and a reader who knew which had good threads is denominated in the remainder of their savings.
And here is the blind spot: we treat silence as a product defect. A null output reads as a broken tool. But in a market where the marginal cost of producing plausible text has gone to zero, the only scarce thing left is a claim with a parent. Scarcity is the signal. The producer who will not fill the blank is not failing to deliver. They are the entire delivery.
I am aware of the cost of that position. I have watched honest analysts lose retainers because they refused to write a number they could not source. I have been that analyst. It is a lonely way to work and it does not scale, and I have decided that not scaling is, in this specific case, the point.
So what should we build next?
A provenance standard for claims, not for tokens. Every published number carrying a parent: the file, the block, the signature, the conversation, the date. Not a disclaimer at the bottom. A chain of custody at the top. A provenance is a promise with a timestamp on it, and this industry knows perfectly well how to keep those.
We spent fifteen years teaching the world to verify state instead of trusting institutions. It would be a strange and bitter ending if we exited this cycle having taught ourselves nothing about verifying the words we write — and if the only entity left in crypto willing to return an empty field when it was handed nothing turned out to be a machine.
