The schema came back full. Every field was populated. And every single one of them was a lie.
I was auditing a two-stage extraction pipeline for a crypto research desk last month when I hit the failure mode that has quietly corrupted more on-chain intelligence workflows than any exploit I have documented. Stage one is supposed to distill raw article text into structured information points: discrete factual units with source and timestamp, a stance classification, a domain tag. Stage two consumes that structure and produces analysis. When I opened the intermediate output, I found 未提供 in the summary field, 未分类 in the domain field, 未评估 in the epistemic-weight field, and an empty list where the information points belonged. This is the software equivalent of a clean bill of health written by a doctor who never entered the room.
Here is why this matters more in crypto than in most data disciplines. Roughly 70% of the market intelligence that reaches retail and even institutional desks is machine-summarized — governance forum scrapers, protocol newsletter aggregators, token unlock alert services, LLM-driven 'sentiment' feeds. The whole premise of due diligence is that somewhere upstream, a human or a well-tested model actually read the source. When the pipeline silently substitutes a template for an extraction, the defensibility of every downstream conclusion evaporates, and nobody notices because the JSON is valid.
Context: How a Template Pretends to Be Data
The pipeline in question was a two-stage design. Stage one: an LLM reads raw text and emits structured extraction — information points (statement, source, timestamp), a one-line abstract, an author stance, a purpose classification, and domain labels with confidence scores. Stage two: an analyst-facing model consumes that structure and produces a deep analysis. The contract between the stages assumes that stage one either succeeds or fails loudly. It does neither.
My reconstruction of the likely causal chain, in order of probability: the stage-one model call timed out under load, the orchestrator fell back to default values to avoid breaking the run, and no exception propagated. A variant: the code passed the stage-one prompt schema into stage two instead of the stage-one output, so the downstream model dutifully reported 'not provided' for every field because the schema defines the fields but carries none of the content. A third variant, which I have seen in production scrapers: the article body was empty — a paywall, a JS-rendered page, a rate-limited fetch — and the extractor, receiving nothing, returned defaults with a 200 status code. A fourth: field-name drift between services, where upstream emits core_viewpoints and downstream reads 核心观点, and every miss falls silently to the default branch.
I have made this mistake myself. In 2022, while auditing the post-Terra ruins, I built a reentrancy scanner that returned a green 'no findings' report on three lending protocols. The scanner had a regex bug that matched nothing. I shipped the report to a small fund before I caught it. The lesson I internalized — and the one this pipeline has not learned — is that a silent default is more dangerous than a crash, because a crash forces inspection while a default invites trust.
Core: The Placeholder Problem Is an Integrity Problem
The technical failure is mundane. The systemic failure is not. Look at what the pipeline cannot distinguish: whether the original article genuinely contained no abstract-worthy content, or whether extraction simply never ran. These two states demand opposite remediation — the first should lower your confidence in the analysis; the second should rerun the pipeline entirely. When both collapse into the same token, 未提供, the downstream consumer has no way to route the decision correctly. The system has manufactured ambiguity and called it data.
This is the same disease I documented in the spot Bitcoin ETF prospectuses in 2024. The custody-risk disclosure and the actual cold-storage architecture diverged by about 15%, and the marketing language was technically compliant while operationally misleading. The regulated document and the placeholder schema share a pathology: they satisfy the form of disclosure without transmitting the substance. A field that is present but empty is not transparency. It is the appearance of transparency, which is worse, because it consumes the reader's trust budget without repaying it.
The second structural issue is that placeholder values were treated as legal output. If 未提供 and 未评估 are valid enum members, then the contract between stage one and stage two has no definition of failure. There is no status: failed, no null that forces a branch, no confidence threshold that blocks downstream execution. In smart-contract terms, the pipeline has no require() statement guarding the call. It is a function that returns true on every input, which means it returns nothing at all.
The fix I recommended, and which I would push onto any team running multi-stage extraction against on-chain or market data, is a status envelope that makes the failure mode legible: