The report arrived with nine fields and no content. Title: absent. Source: absent. Type: absent. Domain: "unclassified," annotated with a confidence label of "unassessed." Core thesis: not determined. Information points: an empty list. Entity or protocol: unidentified. Time sensitivity: undeterminable. Source quality: unevaluable.
Nine nulls, a header block, a closing section, and a recommendation.
I have audited a great deal of bad documentation. Most of it fails by containing something wrong — a slippage formula that ignores fragmentation, a metadata update path with no transfer restriction, an attestation report with no UTXO detail. This document failed by containing nothing at all while preserving every structural marker of completed analysis: the field labels, the diagnostic posture, the orderly escalation to "please supply additional material." It is the most honest artifact I have reviewed this quarter and the most dangerous, and those two facts are not in tension. The hazard is not the null. The hazard is the format.
A pipeline that says nothing is a pipeline that failed safe. A pipeline that says nothing inside a report-shaped container is a pipeline that has manufactured the appearance of diligence. In a market where the average asset now clears nine figures before publishing a technical specification, the second failure mode is the one that gets funded.
Context: the industrialization of deconstruction
Every cycle produces a research bottleneck, and every cycle answers it with volume. In 2017 the bottleneck was primary-source reading — I spent three weeks reverse-engineering the 0x whitepaper by hand, cross-referencing its atomic swap proofs against academic literature, and filing a forty-page debrief that received no response. In 2020 the bottleneck was empirical validation, which I answered by writing a Python model of Curve's 3Pool and running a 15% stablecoin depeg through it until the invariant broke. In 2021 it was contract-level review of assets nobody had read line by line. In 2024 it was custody architecture inside regulated wrappers.
In this cycle the bottleneck is not analysis. It is triage. The number of assets seeking capital exceeds the number of qualified hours available to dismiss them by roughly two orders of magnitude, so firms have done what firms do: they have split the work into stages. Stage one deconstructs — extract title, source, type, domain, thesis, information points, entities, time sensitivity, source quality. Stage two reasons over the extracted facts. Between them sits a gate. The gate is supposed to be the safety property: nothing proceeds to deep analysis until the extraction is materially complete.
The gate is the entire architecture. Everything else is formatting.
What I am describing here is not a hypothetical. It is the observed output of a live pipeline running against real material, and the output was a fully rendered template with a null in every load-bearing position. The interesting question is not why the extraction failed — extraction fails constantly, and it should. The interesting question is why the failure emerged wearing the costume of a deliverable, and what that costume does to the humans downstream of it.
Core: the arithmetic of the empty report
Start with the scoring function, because the scoring function is where the dishonesty is encoded. Most pipeline confidence metrics are built on the same benign-looking premise: a missing field is "inconclusive," and inconclusive should be scored neutrally rather than as zero. Scored neutrally, in practice, means scored at 0.5.