A blank template passed as analysis is not a failure of effort. It is a failure of architecture. When an automated research pipeline outputs 标题: 未提供, 来源: 未分类, and a field literally reading "identify projects from the above information points" — while there are zero information points above — the system has done something more interesting than error. It has revealed its own dependency graph.
I have spent eighteen years watching data pipelines fail in crypto, and based on my audit work on oracle networks and proof-of-reserves infrastructure, I can tell you: the most dangerous output is not a wrong answer. It is a confident answer built on a null input.
This is what the pasted document exposes. A stage-two analytical process was invoked on a stage-one artifact that contained no facts. No title. No source. No timestamp. No information points. The template instructions (请从上述信息点中识别) survived into the output because nothing overwrote them. The pipeline had no validation gate between ingestion and analysis.
Here is the forensic reconstruction.
Context: The Research Stack Has No Circuit Breaker
In traditional market surveillance — the discipline I practice daily as a 7x24 analyst — every data feed has a circuit breaker. If a price feed returns null, the system flags stale data and halts downstream computation. If a settlement instruction arrives without a counterparty reference, it is rejected, not enriched with guesses.
Crypto research infrastructure rarely implements this. The incentives point the other way. Volume of output is rewarded; null-handling is invisible. A pipeline that produces a thousand reports with 50 blank fields still looks more productive than one that halts on bad input.
I watched this pattern crystallize during the 2017 Parity multisig audit. The contract's initWallet function had a well-known pattern: it checked whether the wallet was already initialized, but the check could be bypassed because the function was callable by anyone. The compiler warning existed. The check existed. But there was no enforcement at the boundary — no require that made the precondition load-bearing. The bug was not in the logic. It was in the missing gate.
Research pipelines have the identical failure mode. The analytical stage assumed the input stage had succeeded. No require(input.isPopulated()). The result: an analysis that could only have been produced by fabricating facts, and which correctly refused to do so — but only after the pipeline had already committed CPU cycles and, more importantly, already produced a document that looks like a deliverable.
That last part matters. This document was formatted. It had headers. It had a remediation section. It had a text code block containing a schema with fields like 作者立场与利益相关. The presentation layer executed flawlessly on zero data. Presentation succeeded where substance failed, and that is the exact inversion a bull market rewards.
Core: What the Null Output Actually Reveals
The pasted text is itself a piece of on-chain evidence. Let me parse it the way I parse a failed transaction.
Three fields — 一句话摘要, 作者立场, 文章目的 — are marked empty or 未判断. One field, 涉及项目/协议, contains the template instruction rather than a value. 标题, 来源, 类型, 领域 are all 未提供 or 未分类. The 信息点列表 is empty.
This is not a corrupted template. This is a template that was executed with an injection point left unpatched. The instructions embedded in the schema were never replaced by runtime data. In smart contract terms, this is the equivalent of a contract that ships its constructor arguments in the bytecode comment — readable, present, and completely inert.
The document's own conclusion is technically correct: building analysis on this input would require fabricating content, and fabrication is worse than abstention. I agree with that conclusion. But the conclusion is also a tell. *A pipeline that can detect "this input is empty, I cannot analyze" but cannot detect it before generating a formatted response has its validation in the wrong layer.* The check happens after the expensive operation, not before.
Based on my modeling of cascading failures in Aave and Compound, this is the single most replicated design error in DeFi risk systems. Liquidations are checked after the position is underwater, not when the borrow is initiated. Oracle feeds are validated after the price is used, not before. The check exists, but it sits downstream of the damage.
The remediation the document proposes is a two-path solution: provide the source text (path A), or provide a populated stage-one result with a defined minimum schema (path B). Look closely at what that second path requires. Title, source, publish time, article type, one-sentence summary, two to five core arguments each with original-text evidence, three to ten information points each labeled by source tier — first-hand (official docs, code repos, on-chain data, financials), second-hand (media, analyst secondary sourcing), or unsourced assertion — plus author positions and conflicts (holdings, project affiliation, paid placement), plus named projects and protocols.
That schema is the entire analytical philosophy of a serious researcher compressed into a JSON block. The tiering of sources — first-hand versus second-hand versus unsourced — is the load-bearing wall. It is the thing that separates a research pipeline from a rumor mill. And its absence is why the pipeline failed closed.
There is a deeper insight buried here that the document itself does not surface. Notice the required field 作者立场与利益相关（是否持仓、是否项目方、是否付费稿）. This is conflict-of-interest disclosure encoded as a mandatory input field. In every serious research operation I have audited, this field is the one most often left blank, because it is the one most likely to embarrass the author. The template treats it as non-negotiable. That is a choice, and it is the correct one.
Now the timing.This document is dated in a bull market. The irony is structural. In euphoria, research volume spikes. Every fund, every protocol, every influencer runs a pipeline. The demand for analysis outpaces the supply of verifiable data. So pipelines normalize empty fields, soften 未分类 into plausible categories, and ship. The document you are reading refused to do this — but only at the final step, and only because its prompt happened to catch the null.
Most pipelines do not have that prompt. Most would have generated a confident, fluent, entirely hallucinated analysis: a project name inferred from context, a thesis assembled from priors, a 作者立场 guessed as neutral. That output would pass every surface-level check. It would be cited. It would move a small cap. Predictability is a myth; only volatility is real — and the most persistent volatility source in crypto research is not market manipulation, it is fabricated analytical outputs that circulate as fact.
Contrarian: The Fix Is Not Better Prompting
The obvious remediation — add validation to the analysis prompt — misses the point and would make the system worse.
Adding a if input is empty, return error instruction to the analysis stage does not fix the architecture. It adds a second check where a first one should exist, and it leaves the expensive analysis stage still coupled to an unvalidated input stage. What the pipeline actually needs is a hard partition: the ingestion stage emits a signed, schema-validated artifact, and the analysis stage refuses to import anything that fails schema validation. Not a prompt instruction. A type system.
The document proposes exactly this, somewhat buried in its two-path remedy: the minimum input specification is a schema. But it frames the schema as "what you need to give me," not "what the pipeline must enforce before it runs." The distinction is the difference between a polite request and a circuit breaker.
There is a second contrarian point that the document gets right almost by accident. It says the pipeline should perform stage-one deconstruction — including source quality grading and time-sensitivity tagging — before stage-two analysis, because "skipping the former makes the conclusion untrustworthy, and errors compound downward."
This is the correct principle, but it applies more broadly than a single pipeline. In crypto, error compounding is the dominant risk mechanism. Look at 2022's Terra collapse through my forensic timeline: the seigniorage model assumed demand for UST would grow monotonically. Every data point that contradicted this was filtered at ingestion — the growing mint/burn imbalance, the reserve depletion, the anchor yield subsidy becoming the only demand source. By the time the recursive death spiral became legible, the input layer had been compromised for months. The analysis was downstream of a poisoned feed.
History does not repeat, but it rhymes in binary. The empty template is the Terra problem in miniature: an input layer that emits placeholders where facts should be, and a downstream system that keeps computing anyway. The document's own failure to flag this at the architectural level — rather than the individual-prompt level — is the same blind spot that let UST reach zero.
One more thing the document does not say, and I will. An empty template passed into a research pipeline is not a user error. It is a signal that someone, somewhere, automated a step that should have required human verification. Somebody wired a stage-two prompt to a stage-one output that had not been checked. Somebody chose to optimize for throughput. The document format — headers, schema block, remediation section — was designed by someone. That design assumed the input would always be populated. That is the original sin: no assert(input != null) at the boundary where the design crossed from theory into execution.
Takeaway: The Next Circuit Breaker Will Be Boring
Watch the next six months for two signals. First, schema-validated research artifacts — signed JSON, versioned, with mandatory source-tier fields — will start shipping from the more serious funds. They will look boring. They will not go viral. They will be the thing that saves a portfolio when a fabricated analysis would have taken it down. Second, the pipelines that lack these gates will fail in the most expensive possible way: confidently, at scale, in public, during a bull market, when the incentive to ship anything is highest.
The question worth sitting with is not whether your research pipeline can produce an article. It is whether it can refuse to. The document that reached me refused. Most will not.