Last quarter my desk processed 62 research escalations. Not one of them was about price. I audited the requests themselves rather than the assets behind them, and 44 of the 62 β 71% β arrived with the input layer blank. No tagged information points. No identified counterparties. No time-sensitivity grade. No source-quality assessment. A title, a ticker, and a demand for ten-dimension deep analysis.
The framework was never the bottleneck. The intake was.
That 71% sits uncomfortably close to a figure I published two years ago while building data integrity standards for autonomous agents: 30% of AI-driven execution errors traced to corrupted oracle inputs, not to faulty agent logic. Deterministic systems do not improvise. They amplify. Feed a pipeline an empty field and it will return a filled page, because structure reads as rigor whether or not anything underneath it exists.
Research desks run on escalation templates now. The standard institutional version, adopted broadly after 2022, carries ten dimensions: technical architecture, token economics, market microstructure, ecosystem position, regulatory exposure, team and governance, risk surface, narrative and expectation pricing, supply-chain transmission, and a synthesis layer.
That list is good. I use a version of it. But a template is a transmission protocol, not an oracle. A protocol moves what you feed it, at the fidelity you feed it, and it does so without complaint. When my team standardized due diligence in the weeks after the Terra collapse β mandatory on-chain verification, no exceptions, no narrative overrides β the change that mattered was not the ten dimensions. It was the requirement that every dimension be reachable from a sourced, dated, addressable input.
In practice the escalation path looks like this. A portfolio manager flags a position. A junior analyst receives a ticker and a deadline. The ten dimensions get populated in order. The template fills, the deck ships, and the provenance of every number inside it is a Slack message. Third-party template vendors sell the same structure to a dozen funds, which means a dozen funds inherit the same blind spot simultaneously.
Analysts keep upgrading the model. The model was never the problem.
Start with three failure modes, because they cascade.
Failure one: the untagged information point. A claim without a source and an original position cannot be graded. "TVL up 40% this month" is not a data point; it is a rumor with a typeface. Tagged properly β source, timestamp, contract address, original text location β the same claim becomes auditable, and roughly a third of the time it collapses. Double-counted deposits. Incentive-driven inflows. A migration counted as organic growth.
Failure two: the unidentified counterparty. You cannot run ecosystem analysis on "a protocol." Tokenomics without contract addresses is fiction with a chart. I have watched a nine-figure ecosystem fund deploy against a token model reconstructed from a blog post, because the escalation form had a field labeled "project" and someone typed a brand name into it.
Failure three: the ungraded time sensitivity. A 48-hour liquidity event analyzed as a structural trend produces a thesis that is wrong in both directions β too alarmist for a week, too complacent for a quarter. Time sensitivity is not a label. It sets the sampling window, and the sampling window determines which conclusions are even available to you.
Run all three failures against a ten-dimension framework and the output is not "incomplete." It is confident. That is the danger. Empty fields do not stay empty; they get filled by whoever asks loudest, which is how an analyst's priors enter a report that will later be described as a deduction. Nobody writes "I assumed this" in a synthesis section.
I learned this shape of failure in 2018, during a six-week audit of the Zcash shielded transaction protocol. Three zero-knowledge proof implementation flaws, each capable of balance inflation, found by tracing consensus rules line by line. No whitepaper mentioned them. No summary contained them. Code does not lie, only developers do β and a consensus layer does not accept "the documentation says" as an input. Neither should a research desk.
The counter-example is 2022. When algorithmic stablecoin reserves began showing anomalous composition, my fund liquidated 80% of exposure inside 48 hours. Peers running comparable analytical frameworks bled out. The difference was not intelligence or conviction. It was that our reserves data had been ingested, tagged, and graded weekly, so the decision was mechanical by the time it was needed. Standardization survives the chaos of collapse. Frameworks that live only in a slide deck do not.
By 2026 the same principle had migrated to machine execution. The verification protocol my team built for autonomous agents β zero-knowledge validation of oracle inputs before an agent acts, deployed across three DeFi lending protocols β cut oracle-related losses 45%. The valuable component was never the verifier. It was the sequencing decision: verify before execution rather than reconcile after loss. Most desks still do the latter, and they call the reconciliation "post-mortem analysis."
What I enforce now is a five-field intake schema, and it is deliberately small. Title. At least three tagged information points, each with source and original position. Identified counterparties with contract addresses. Time-sensitivity grade. Source-quality grade. If a field is empty, the request returns to sender. Our rejection rate sits above 40%, and the portfolio managers who complain loudest are the ones whose positions we have not had to unwind.
There is a measurement discipline underneath this. When a growth narrative arrives, the first cross-check is pool depth, because liquidity is the current of truth β volume stories that cannot survive contact with actual depth were never data to begin with. The graph clarifies what sentiment confuses, provided the graph is drawn from claims that had provenance before they were plotted. Provenance first, visualization second. The reverse order produces beautiful charts of nothing.
The comfortable assumption is that better frameworks produce better research. The correlation is real; the causation is not. Frameworks and accuracy move together because both track input quality. Upgrade the model while the pipeline stays broken, and you get the same errors with more slide transitions. Ledger lines reveal what noise obscures β but only if someone bothered to write the lines down.
The bull market makes this worse, not better. Urgency substitutes for intake when everything is going up, and completeness gets punished as delay. An empty field therefore gets treated as a scheduling problem rather than an epistemic one. It is neither. It is a liability with a delivery date, and it compounds quietly until the position it supports has to be exited at whatever the book will take. Efficiency is the only permanent alpha, and the most efficient thing a research desk can do is refuse work it cannot source.
Next quarter, measure the rejection rate, not the coverage rate. The volume of escalations a desk sends back is a better leading indicator of drawdown avoidance than the volume of reports it ships. The signal to watch is whether ingestion schemas start appearing in institutional due-diligence templates the way post-mortem frameworks did after 2022. If they do, the industry has learned something durable. If they do not, the next collapse will be fully documented β in a template with empty fields.