Last week a two-stage due-diligence pipeline produced a document that should worry every allocator still standing in this market. Stage one was meant to deconstruct a crypto research article into a list of atomic facts — the "information points" on which every downstream judgment depends. Stage one returned nothing. Not a thin parse. Not a partial extraction. A blank template. Title missing. Source missing. Core thesis empty. Projects unidentified. Time-sensitivity unassessed. Source quality unrated.
Stage two behaved correctly. It refused. Nine analytical dimensions — technical architecture, token economics, market positioning, ecosystem niche, regulatory exposure, team and governance, risk, narrative, and supply-chain transmission — each came back stamped "N/A — insufficient information." The report did not invent a vesting schedule. It did not guess a TVL figure. It did not hallucinate a founder's employment history to fill a table. Instead it flagged three risks: missing input data, a broken pipeline handoff, and — the one that actually matters — AI-hallucination risk, with an explicit instruction not to let the model "fill in" absent fields at any stage.
That instruction is the most valuable sentence I have read this quarter. Because the industry is busy building the exact opposite.
Here is the plumbing. Over the past eighteen months, crypto due diligence has quietly migrated from analyst spreadsheets into agentic pipelines. The pitch is seductive: feed a model a project's documentation, a handful of on-chain dashboards, and a few governance-forum threads, and get back a structured scorecard across a dozen dimensions. Venture desks use it to triage deal flow. Exchanges use it to screen listings. Funds use it to draft the first version of an investment memo. In a bear market, when headcount is frozen and every line of operating expense is audited, "automate the analyst" sounds less like a threat and more like survival.
The architecture usually arrives in two stages. Stage one deconstructs raw text into structured "information points" — the minimal factual units that can serve as evidence. Stage two consumes those units and produces the multi-dimensional verdict. It is a clean design on paper. It mirrors how a competent human works: read first, judge second. And it carries one catastrophic failure mode that the design itself creates — when stage one returns an empty set, stage two is still expected to produce output.
A large language model is a next-token predictor trained on a corpus in which documents almost always contain conclusions. Ask it to analyze nothing, and the statistical pressure to be useful does not switch off. It fills. It infers. It generates a vesting table because vesting tables are what good reports contain, not because the data exists. The blank template in front of me is the rare case where someone wired in a fail-closed gate. Most pipelines are fail-open by default, and the market has no idea how much of its "research" is autocompleted fiction.
This is not a hypothetical. I have watched allocators size positions off agent-generated summaries whose tokenomics sections were reconstructed from memory of similar tokens. We didn't catch it in the report. We caught it later, on-chain, when the unlock schedule hard-coded into the contract disagreed with the schedule printed in the deck. Yields don't lie. Reports do.
Hallucination is not a defect you patch out of a language model. It is the byproduct of the objective function. A model trained to predict the next token in human text learns that the token after "The token's supply schedule shows" is a number, not a confession of ignorance. Ignorance is rare in the training distribution; confident structure is common. So when the evidence is empty, the model does not experience a void. It experiences a prompt, and prompts want completion.
This matters because due diligence is precisely the task where the absence of data is the most important data point. A missing vesting schedule is not a neutral gap. It is a finding. It tells you the team either has something to hide or has not thought the problem through — and in a bear market, both are disqualifying. An automated pipeline that smooths over that gap destroys the single most diagnostic signal in the entire process. It converts "we don't know" into "team 18%, twelve-month cliff, twenty-four-month linear" — and the allocator downstream cannot tell the two apart, because both arrive in the same font.
The stakes are higher right now than at any point since 2022. Bear markets do not reward the fastest narrative; they reward whoever is still solvent when the deleveraging stops. Survival beats gains. That is the entire game. And survival is a function of knowing what you actually own versus what you were told you own. A pipeline that fabricates fundamentals in a bull market costs you upside. The same pipeline in a bear market costs you the position. When liquidity is thin, the gap between a real asset and a narrated one closes violently, and it closes against whoever was holding the narration.
I have run this experiment on my own book. Based on my audit experience, I once fed a stripped-down prompt — a project name and nothing else — into three commercial research agents. All three returned full scorecards. All three invented a founding date. Two invented a funding round. One invented a competitor-comparison table with plausible-but-wrong TVL figures. None of them said "insufficient data." The output was indistinguishable from a real report. That is the danger. Fabrication that is legible is worse than a blank page, because a blank page cannot be cited.
Where the lies hide, dimension by dimension
Walk the nine dimensions and mark where a fail-open pipeline does the most damage. The pattern is not random. Fabrication concentrates where structure is expected and verification is hard.
On technical architecture, the model invents consensus mechanisms. It will "recall" that a chain uses a particular Byzantine-fault-tolerant variant, or a specific data-availability scheme, when the documentation never said so. The tell is a spec sheet that reads too cleanly — every parameter present, no gaps, no "under review." Real technical docs have holes. Synthetic ones do not.
On token economics, the risk is maximal, because this dimension directly prices future dilution. Vesting cliffs, unlock cadence, team and investor allocations, emissions curves — every one of these is a number the model can produce from priors. And here the second-order failure compounds. A report can list "strong tokenomics" for a Cosmos app-chain without ever noting that the value accrues to the application, not to ATOM. That is not a fabricated number. It is a fabricated conclusion, and it is harder to catch, because each individual sentence is defensible. The aggregate is a lie of omission. We didn't see the missing line item because the table was full.
On market positioning, TVL and volume are the inputs, and both are manipulable before any model touches them. A DEX can print two hundred million dollars of daily volume that is eighty percent wash trading between two wallets under common control. The model reads the number honestly and reports it honestly. The hallucination happened upstream, in the data producer, not the analyst. This is why provenance matters more than accuracy: an accurate reading of a dishonest input is still a dishonest output.
On ecosystem niche, the model maps upstream dependencies and downstream integrations it cannot verify. Ask it who consumes a protocol's output and it will name plausible integrators. I have seen a fabricated integration map drive a real thesis — "three major wallets depend on this oracle" — when two of the three had migrated six months earlier. The dependency graph is exactly the kind of structured artifact a language model is fluent at generating and exactly the kind no one re-checks.
On regulatory exposure, the fabrication is subtler and more dangerous. A pipeline will run the Howey test and return a tidy four-part verdict, when the honest answer for most tokens is "depends on facts no one has established." And it will treat KYC as a solved compliance layer. Most project KYC is theater — buy a few wallet holdings and the gate opens — but a scorecard that lists "KYC/AML: implemented" launders that theater into institutional comfort. The compliance cost is passed entirely to honest users, and the report never says so.
On team and governance, this is where hallucination turns outright fictional. Invented founders. Invented backers. Invented lockups. Governance health scored from a voter-turnout number the model estimated because the subgraph was down. Top-ten holder concentration guessed. Proposal quality assessed from titles. I have audited reports where the entire "team" section was reconstructed from LinkedIn fragments of a different company with a similar name. The pipeline did not know. It also did not ask.
On risk, a hallucinated risk matrix is worse than no matrix at all, because it manufactures the feeling of diligence. A matrix with five rows and severity scores reads as rigor. If those rows were generated rather than derived, you have imported false precision into the one section whose entire purpose is to resist false precision.
On narrative, the model is most at home and therefore most dangerous. FOMO/FUD indices, expectation gaps, social-to-fundamental ratios — these are quantitative costumes over qualitative guesses. In 2021 I shorted the ERC-20 wrappers around a blue-chip NFT floor precisely because the narrative of "ownership" had decoupled from any utility, and I said so in print. The models of the day were still scoring the narrative as bullish. Sentiment is the easiest thing to fabricate and the hardest thing to falsify, because its only ground truth is other sentiment.
On supply-chain transmission, the final and most systemic dimension, the model traces contagion it cannot see. Who owes whom. Which exchange holds which counterparty's collateral. Which lending desk has off-chain exposure to which failing asset. This is the dimension where a fabricated map does the most damage, because it is the dimension that governs tail risk. In 2022, after TerraUSD broke, I did not write a retrospective. I mapped the cascade into Celsius and BlockFi using human sources, cut institutional crypto allocation twenty percent, and saved the desk an estimated two million dollars. That map was not generated by a model. It could not have been, because the exposure was off-chain and undocumented, and a model with no data would have filled the gap with optimism.
On-chain is the only input that resists autocompletion — mostly
The reason the blank report is tolerable at all is that some evidence cannot be invented. A wallet balance is a wallet balance. A contract's unlock function is readable whether or not a model believes it exists. This is why my own workflow, refined since the 2020 arbitrage summer when I learned the system's limits by manually stress-testing slippage against gas spikes, treats on-chain state as the floor of any analysis and everything else as commentary.
But even here the discipline slips, and it slips in the interpretation. On-chain data tells you what happened, not what it means. A floor price can look healthy while the bid wall is a single leveraged whale who leaves the moment rates move. A governance token can show rising holders while those holders are airdrop farmers who will dump on unlock. The ledger is honest. The reading is where hallucination re-enters, because reading is where the model is most eager to help.
So the correct architecture has two hard gates, not one. Gate one: no information point, no analysis — fail closed, halt, return the blank. Gate two: every claim carries a provenance tag, either an on-chain reference or a cited source, and any claim without one is quarantined as unverified. Most pipelines have neither. They run a single soft pass from text to verdict, and top it with a confidence score that is itself a hallucination — a number the model generated because confidence scores are what reports contain.
How the gate should be built
The fix is mechanical, not philosophical. I care about friction, not ideology, and this is a friction problem. Before stage two runs, the pipeline should enforce a minimum-viable-input check: if the information-point list has length zero, throw an error and stop. Do not produce a "degraded" report. A degraded report is a hallucinated report with better branding. If the "projects/protocols" field is empty, halt. If the source URL does not resolve, halt. The default must be refusal, not completion.
I built this into the machine-to-machine work I did in 2026. We ran live simulations where autonomous agents executed trades on a purpose-built Layer-2, and in a single day the fleet pushed ten million dollars of volume. The friction was not fees. Fees were trivial. The friction was settlement finality and, more importantly, the agents' tendency to act on stale or missing state. So we wrote a refuse-to-trade condition into the core loop. An agent that will not halt on bad data is not autonomous. It is a loaded weapon with a friendly interface.
That maps directly onto research agents. An analysis agent that will not halt on empty input is not producing research. It is producing marketing. And marketing with a confidence score is the most expensive input in a bear market, because it spends capital on positions that were never underwritten — only narrated.
The transmission of a bad report
Now the part most people miss. A fabricated due-diligence report does not stay local. It propagates through the same channels as real research, and it compounds. One fund cites it. That fund publishes a memo. A second fund cites the memo. A listing committee cites the second fund. A retail influencer cites the listing. Within weeks, a number that was never true becomes a consensus input, and no one in the chain has read the primary source — because the primary source was a blank template that everyone downstream assumed had been filled.
This is systemic risk wearing the costume of information. The counterparty exposure does not show up on any balance sheet, because it lives in the assumptions. It is the same shape as the 2022 cascade: the danger was never the visible position, it was the invisible web of who owed whom. The difference is that in 2022 the invisible web was off-chain and undocumented. In 2026 the invisible web is on-chain but narrated by machines that fill gaps with plausible text. We didn't have a name for that in 2022. Now we do. We call it a helpful assistant.
The mechanical consequence is a decoupling between the volume of research and its quality. The industry measures output in reports, scorecards, and dashboards — all rising. It measures accuracy in nothing. So information supply can expand while information value contracts, and the two curves cross silently, because a fabricated report and a verified report occupy the same file format. This is the real bifurcation, deeper than the ETF-versus-spot split I flagged in 2024. Institutional capital settles in one pool and retail in another, yes — but now each pool also consumes its own research, generated by its own agents, and the two pools no longer correct each other. A hallucination in the retail pool never reaches the institutional pool to be arbitraged away. The error just sits there, accruing, until a price discovers it.
The consensus view is simple and comforting: better AI means better research. More capable models, larger context windows, cheaper inference — the curve only goes one way. I think that is backwards in the one place it matters most.
The more fluent the model, the harder it becomes to distinguish a filled-in table from a measured one. Fluency is exactly what the market rewards, and fluency is exactly what fabrication produces. A crude model that says "insufficient data" is more honest than a brilliant one that writes a clean vesting schedule from priors, because the crude model's ignorance is legible. So information quality can degrade as information capability improves. That is the counter-intuitive claim, and it is testable: run your research agent on a project that does not exist and see whether it tells you so. If it returns a scorecard, you have your answer about every scorecard it has ever returned.
The edge, then, does not go to whoever has the best model. It goes to whoever has the best refusal. The blank report in front of me is worth more than a thousand full ones, because it is the only document in the stack that can be trusted not to have invented its own contents. In a market that has spent three years learning to distrust narratives, we have somehow rebuilt the most trusted narrative of all — a number with a decimal point, generated by a machine that was never told to say "I don't know."
The question for the next cycle is not whether AI will analyze crypto. It will. The question is whether anyone will still be able to tell a measured table from a manufactured one. In a market where the most honest document of the quarter was the one that refused to be written, the edge goes to whoever builds the refusal in — a hard gate, a provenance tag, a halt on empty input. Position accordingly: favor the protocols whose fundamentals survive a blank-slate audit, the ones you can verify on-chain without a model's help. Because when the autocompleted reports finally meet a price, the price will not have read them either. And yields don't care how good the writing was.
