Garbage In, Nothing Out: The Silent Failure Mode of Crypto's AI Due Diligence

The abort notice landed at 09:14. A nine-dimension due diligence framework — technical surface, tokenomic structure, regulatory exposure, the entire apparatus — returned one verdict across every field: N/A. Not bearish. Not unverified. Empty. The information-point list that was supposed to feed the whole pipeline contained zero entries. Nine analytical dimensions, each a discrete logical unit, all terminated against a null pointer.

No title. No source. No protocol name. No facts.

The front-runner didn't lose the trade. There was no trade. That is the detail most people will skim past. A system designed to dissect distressed crypto assets executed its integrity check and refused to proceed. In a market that rewards volume over veracity, that refusal is the rarest output a machine can produce.

I have spent twenty-nine years watching crypto infrastructure fail at exactly this seam — the point where a system is asked to reason about inputs it never received. The failure is rarely dramatic. It is silent, well-formatted, and confidently wrong.

The context nobody priced

Since 2024, the AI-crypto convergence has produced a specific artifact: the automated due diligence stack. Venture funds, exchanges, and compliance desks now route token evaluations through language models that ingest a news item, decompose it into structured fields, and emit nine-dimensional reports in under three seconds. The pitch is scale. One analyst, ten thousand tokens. The economics are irresistible to anyone who has paid a junior researcher to read a whitepaper.

The 2025 EU AI Act guidelines — which I contributed to, in a small way, through a theoretical framework on trustless AI oracles — classified these systems as high-risk when they touch financial decisioning. The classification was correct. The enforcement lag is not.

Compliance officers I speak with privately describe the same gap. They are legally required to document the basis for a decision. An AI-generated report satisfies the documentation requirement on its face. Whether the underlying input ever existed is a question nobody is incentivized to ask, because asking it would slow the workflow the automation was purchased to accelerate. In a bull market, that incentive compounds. When everything is going up, a fabricated report and a verified report produce the same P&L, and only one of them costs time.

Here is the structural problem. These pipelines have two failure modes, and the industry only monitors one.

The visible failure is fabrication: a model invents a token supply, hallucinates a governance model, assigns a risk grade to a project that does not exist. This gets caught. It embarrasses a fund. It generates a retraction.

The invisible failure is the one in front of me now. A pipeline receives empty input and, instead of halting, fills its template anyway. Nine sections. Confident prose. Zero information. The template becomes the product. The reader cannot distinguish a report built on five verified facts from a report built on none, because both arrive in identical formatting, identical length, identical tone.

That is not a bug in the model. That is a bug in the incentive structure around it.

The core teardown

Follow the incentives. A due diligence analyst is paid to produce a judgment. A vendor selling an AI analysis stack is paid to produce throughput — reports per hour, tokens covered per cycle. Neither party is paid to produce the word "insufficient."

Watch what the metrics reward. A pipeline that abstains on 40% of its inputs looks broken. A pipeline that outputs a report on 100% of its inputs looks robust. The scoreboard cannot tell the difference between "analyzed" and "filled." So the system optimizes for the metric that is visible and sacrifices the property that matters.

I learned this in 2020, debugging MempoolWatch. The tool detected sandwich attacks with high precision, but its latency budget meant it flagged some patterns after the block had already closed. I had a choice: suppress low-confidence detections to keep the interface clean, or surface them with a confidence score and accept a noisier output. I chose the cleaner interface. Fifty high-frequency firms adopted it. A thousand retail users ignored it because it looked uncertain.

The lesson was not about latency. It was about legibility. A system that displays its uncertainty gets discarded in favor of a system that lies cleanly. The market selected for confident emptiness, and the AI analysis stacks of 2026 are simply the industrial-scale version of that selection.

Now apply the cryptographic discipline these pipelines claim to inherit. In cryptography, a verifier does not accept an unauthenticated message because it is well-formed. Formatting carries no evidentiary weight. A signature either validates against a public key or it does not. There is no partial credit for a beautiful forgery.

A bug is just a feature that hasn't been triggered in production. Here, the "feature" is template-filling under null input. It sits dormant until the exact condition that should trigger maximum caution — zero source data — triggers maximum fabrication instead.

The input-validation layer is where the whole stack collapses. I audited the EOS account-creation logic in 2017 and found a race condition that could mint tokens infinitely under specific block producer configurations. The mainstream ignored it; three exchanges delayed delistings. The pattern repeats. The exploit was not in the sophisticated consensus layer. It was in the mundane boundary check that everyone assumed was handled elsewhere. Empty input is that boundary. It is boring. It is unglamorous. It is exactly where integrity fails.

The deeper issue is that verification is expensive and formatting is cheap. To confirm a report, a human must trace every claim back to a primary source — a block explorer, a contract, a filing. That is hours of work. To accept a report, a human must read the header. Seconds. The cost asymmetry between verification and acceptance is the entire attack surface. Any system that produces formatted output at scale will be accepted at scale, because the alternative is uneconomic.

Trace the consequence chain:

First, an empty information set enters the pipeline. Second, the template-filling heuristic activates because abstention is penalized. Third, a formatted report emerges with fabricated substance — a plausible token supply, an inferred governance model, a risk grade. Fourth, a human reader, trusting the format, prices a decision on it. Fifth, the decision propagates into a fund's position, an exchange's listing criteria, a compliance officer's sign-off.

Every link in that chain is a system trusting the output of the link before it without re-verifying the input. That is the same structural fragility that took down Terra in 2022. I proved the LUNA/UST feedback loop was unsustainable at a $10 billion market cap. The math was not hidden. It was ignored because the narrative was legible and the math was not. Sixty billion dollars of retail capital paid for that preference.

The AI due diligence stack reproduces the exact pathology at a different layer. It does not collapse a stablecoin. It collapses the epistemic foundation on which capital allocation rests. And it does so at machine speed, across ten thousand tokens, before any human reads the first line.

There is a further refinement. The pipeline's empty input is itself a signal. A stage-one analysis that produces nothing is not noise — it is a measurement. It tells you the source material was absent, malformed, or never transmitted. That is diagnostic data. The correct output was never a nine-dimension report. The correct output was a single line: no data, no analysis, pipeline halted. The abort notice, in this case, was the only truthful artifact the system produced.

Consider the parallel to the oracle problem I flagged in 2025, when I found that synthetic data injection could manipulate price feeds feeding AI agents. The attack surface was never the model's reasoning. It was the provenance of the input. Garbage in, garbage out is not a slogan; it is a security model. And most of this industry has deployed the front end of that model while leaving the back end unwritten.

The contrarian angle

Here is what the builders got right, and it is not nothing.

Automation of due diligence is correct. Human analysts cannot read ten thousand whitepapers, and pretending otherwise is the same scarcity theater that fragments liquidity across forty Layer 2s while the same small user base gets sliced thinner. Throughput is real value. The pipelines are not wrong to exist.

The abstention behavior — the halt on empty input — is also correct. I spent six months in 2021 after "The Gaming Illusion" absorbing ten thousand downvotes because I insisted Axie Infinity's revenue model depended on perpetual new-user inflows. I was right, and being right cost me an audience. The lesson I drew was that emotional rebuttal is a losing game. What I should have drawn is that honest abstention and honest judgment are the same discipline. A system that says "I cannot assess this" is more valuable than a system that assesses everything.

The bulls are right that the null result is information. They are wrong to assume it is rare. It is the default state of most inputs that reach these stacks — absent, malformed, or stale. The industry has built engines that convert silence into noise, and it calls the noise coverage.

The blind spot is measurement. Nobody tracks abstention rate. Nobody publishes the fraction of reports generated from empty or single-source inputs. The metric does not exist, so the failure is invisible, so the incentive to fabricate persists. Data speaks; noise interprets. The stacks are optimized to interpret.

Takeaway

The next time an AI stack hands you a nine-dimension report in three seconds, ask one question the format is designed to prevent: what did the input actually contain?

If the answer is nothing, the report is not analysis. It is a forgery with good typography. And the only trustworthy output is the one that refuses to exist.

Integrity is the only immutable asset. Everything else — the throughput, the formatting, the confident prose — is a variable, and variables drift toward whatever the scoreboard rewards. Watch the abstention rate. It is the only number in this industry that cannot be faked.