In January 2026, an AI research agent I run for the fund returned a 4,000-word due diligence memo on a restaking protocol. It was confident, well-cited, and entirely fabricated. Every field in its source table was null.
The agent hadn't malfunctioned. It had executed exactly the architecture I signed off on: parse, weight, summarize, opine. Nobody specified what to do when the parse returned nothing. So it did what language models do with absence โ it filled the gap with the most probable next token.
In 2017 I walked away from a $500,000 advisory contract because a whitepaper's consensus section was mathematically incoherent. That decision was cheap. January was expensive โ not because we lost capital, but because for six hours our risk dashboard displayed a portfolio that did not exist, and two analysts nearly acted on it.
Ignore the price. Watch the input.
Almost all crypto analysis now runs through a four-stage pipeline: a chain indexer (Dune, The Graph, a bare archive node), an aggregator (DefiLlama, CoinGecko, Kaiko), a presentation layer, and โ increasingly โ an autonomous agent that reads the presentation layer and writes conclusions.
Every hop makes the same silent assumption: that the previous hop returned a value. Not a correct value. Just a value. That is a schema-level bet, and nobody prices it.
In a standard ETL stack, a failed API call usually does not throw. It returns an empty set. An empty set summed by a downstream model evaluates to zero. And zero is a legal number. At the schema level, "TVL unavailable" and "TVL is $0" occupy the same byte.
Price oracles solved this years ago. Chainlink feeds revert on staleness because a wrong price liquidates positions. Analytical feeds never adopted that discipline, because being wrong about a chart historically cost nothing. Then agents started executing on charts.
Three failure modes matter, and they compound.
Null-as-zero. A protocol's TVL reads null for one block. The cached value expires. Twelve consecutive nulls get charted as zero. The chart now describes a 100% drawdown that never happened. Any risk model keyed to a drawdown threshold fires โ and it fires at the exact moment its counterparties are firing. In 2020 I structured synthetic hedges against stablecoin depegs because I did not trust the peg. I did not anticipate that the risk system itself could depeg.
Stale-as-fresh. Dashboards cache aggressively to control API cost; 30 to 60 seconds is standard. That cache window is precisely the interval you need during a depeg. In March 2023, several dashboards still displayed $1.00 for USDC minutes after Curve's 3pool had cleared 0.87. Anyone whose bot read the dashboard was trading a number that had been true ten minutes earlier. Bets are cheap; exits are expensive.
Aggregation laundering. Summing TVL across chains and venues. A venue that fails to report does not reduce the numerator. So a market-share percentage gets computed with a venue present in the numerator and absent from the denominator. The error is small, and it never gets smaller โ it gets inherited by every dataset downstream. An aggregator that ingests a laundered zero cannot later un-launder it.
The fix is not better models. It is attestation. Sign the payload at the source, hash it, anchor it, and require the consumer to verify signature and freshness window before the value enters a calculation. That is deterministic, cheap, and almost nobody is building it as a product, because it carries no token narrative.
I published a paper in 2026 arguing that machine-to-machine micropayments would become a $10 billion market. The number is still right. I put the money in the wrong place. The payment rail is the easy part โ a signature scheme and a settlement layer, most of which we have had since Lightning and ERC-20 allowances. The hard part is knowing what the agent actually bought. An agent that pays for a data point without verifying its provenance has simply purchased a more expensive guess.
This is no longer an internal tooling problem. Post-ETF, the marginal consumer of crypto data is an allocator with a compliance department. They do not run archive nodes. They read a PDF and a dashboard. They are structurally incapable of distinguishing a null from a zero, and their mandates do not authorize them to try.
When I run my own eight-dimension screen before capital moves, a failed input produces one output: N/A. I have published entire frameworks where every dimension reads "insufficient information" rather than manufacture a conclusion. That is not humility. It is an exit strategy.
Here is the part the infrastructure crowd will not like. The industry is spending 2026 on data availability layers, rollup throughput, and decentralized compute. Render, Akash, a dozen DA entrants. All of it solves moving bytes. None of it solves knowing whether the bytes mean what you assume they mean.
Dedicated DA layers remain oversold. Most rollups do not generate enough data to justify them, and the ones that do are not the ones buying. What rollups do generate, in enormous volume, is unverified state that downstream consumers treat as ground truth.
The liquidity fragmentation thesis behaves identically. Every new aggregator launched to "solve fragmentation" inserts one more hop where a null can become a zero. Fragmentation is a distribution problem, not a data integrity problem, and the products addressing it add surface area without removing ambiguity.
The AI-crypto convergence narrative makes this worse rather than better. The current thesis is that agents need payment rails. Correct, and insufficient. An agent with a wallet and no calibration is a faster way to be wrong at scale. The scarce resource in 2026 is not compute, or bandwidth, or liquidity. It is the architectural capacity to return "I don't know" โ and to be structurally incapable of returning anything else.
Follow the gas, not the hype. Then follow the signature.
Every number in this market is now one failed API call away from being a fabrication, and the entities consuming those numbers increasingly cannot tell the difference. The next systemic event will not resemble a hack. It will resemble a well-formatted report built on an empty table โ and it will be priced in before anyone checks the source.
When an agent hands you a figure, ask who signed it, when, and what it returns if the source goes dark. If there is no answer, you are not reading data. You are reading a guess with good typography.