Over the past seven days, a deconstruction pipeline ingested a source document and returned zero information points. It also returned nine analytical dimensions. Eight of those dimensions consume the information point array as their only data source. The output was structurally complete: title field, source field, summary field, project field, time-sensitivity field. All present. All empty or marked N/A.
A template with a shape. A shape with nothing inside it. The pipeline did not fail loudly. It completed.
The only substantive event in that log is the refusal downstream. The second stage detected an empty array, declined to fill it with invention, and escalated with three remediation paths. That refusal is worth more analysis than the nine dimensions it prevented. A system that can return nothing is more valuable than a system that always returns something.
The architecture is standard for 2026. Two stages. Stage one parses the source and extracts an information point list โ each point carrying content and provenance โ plus project tags and a time-sensitivity assessment. Stage two runs nine dimensions on top: market structure, risk matrix, industry-chain transmission, competitive positioning, regulatory exposure. Stage two is downstream. It inherits.
When the array is empty, stage two has three possible behaviors. Fabricate. Emit the template. Or halt. The first two are indistinguishable from the outside. Both produce a document that looks finished. Only one of them is honest about the fact that it is not.
I have run this pattern by hand for a decade. In 2024 I co-produced a fifty-page whitepaper mapping twelve regulatory pain points for institutional custodians. The rule was mechanical: every claim carried a citation or it was cut. Not softened. Cut. The document ended up shorter than the brief and considerably more useful. That discipline is not a personality trait. It is a schema constraint, and schemas are the only thing that survive when the analyst is tired.
In 2017 I audited the emission architecture of Golem and Status with a Python script that tracked distribution schedules against live liquidity pools. Golem's claimed distribution mechanics were off by roughly fifteen percent. The number existed because the script compared the raw schedule to the claim. Had I run a template instead of a script, the output would have read distribution mechanics: verified. The discrepancy was not discovered by analysis. It was discovered by an array that was not empty.
The dependency graph is the finding. Eight of nine dimensions route through a single array. That is not redundancy; it is a single point of failure dressed as modularity. Zero information points means eight dimensions have no input, which means their outputs are either generic or invented. There is no third substance. A framework without data does not degrade gracefully. It degrades into prose.
The failure mode is not hallucination. It is hollow compliance: producing the correct shape of an answer to an input that cannot support one. Hallucination is loud. It invents a number you can check. Hollow compliance invents nothing and still ships a document โ headings, transitions, a confident closing paragraph, all of it empty. The second is more dangerous because it passes review. Reviewers check structure. Structure was never the problem.
There is an exact on-chain analogue. An oracle feed that returns a valid-looking price with no underlying data sources. In 2020 I built a model of Aave V2 under a thirty percent ETH drawdown. Forty percent of users went undercollateralized. The interesting part was never the price path. It was the assumption embedded in the feed โ that the number arriving at the contract had a source. Template research is an oracle with a hardcoded price. It answers instantly, forever, and about nothing.
This is why the remediation paths matter more than the failure. The pipeline specified its own minimum viable input: one summary sentence, at least three information points each with a named source, and at least one identified project or an explicit statement that none exists. That is a specification. Specifications are the artifact that outlives the cycle. The ledger remembers what the bubble forgets โ and what it remembered here was that the input threshold was zero.
The three paths are also a taxonomy of research honesty. Rerun stage one with a compliant input. Or hand over the original text and let one pass do deconstruction and nine-dimension analysis together, closing the loop without a handoff. Or supply only a project name and an event description and let the pipeline reconstruct from on-chain and public data, marking every gap as N/A with an attached confidence level. The third path is the one practitioners actually use at 3 a.m. during a liquidation cascade. It is also the one that fails silently if the N/A convention is not enforced.
The escalation itself is the deliverable. Three options, ranked, with the recommendation stated: rerun stage one, or hand over the source, or supply the event and let the pipeline reconstruct. A recommendation is a falsifiable claim about what will work. A template is not.
The 2026 context sharpens this. I have modeled the economics of autonomous agents settling micro-payments on-chain and projected that by 2028 roughly thirty percent of internet traffic will be machine-to-machine. That model assumed something I did not state loudly enough at the time: that agents can verify each other's inputs. Research traffic will scale faster than payment traffic, because research is cheaper to fake and harder to price. An agent economy that settles payments against unverified research is a settlement layer built on a template.
There is an economy to hollow output. A template costs a fraction of an audit and prices at the same rate, because the buyer cannot see the array. The operator in this log named the failure correctly โ filling nine templates to manufacture completion is overstated, not analyzed. Overstatement is not a moral defect. It is a pricing defect. The market pays for the shape of diligence and discounts the substance, right up until the substance is the only thing that clears.
The 2026 discovery layer enforces the same rule from the other side. Ranking systems now reward information gain โ at least one insight the reader did not have. A template has zero information gain by construction, because everything in it was derivable from the schema alone. So the hollow document fails the human reader and the ranking system for the same reason: it contains nothing that could not have been written before the source arrived.
A compliant stage-one array in this market is narrow and specific. Protocol name. Event. Direction of flow. Magnitude with a unit. Timestamp. Source. Anything else is decoration. During the 2022 unwind the arrays that mattered had four fields: which stablecoin, what collateral ratio, which venue, how many hours of redemption queue. Everything else โ governance, roadmap, partnerships โ was noise that read like signal.
Bear-market relevance is direct. In a cycle where survival outranks gains, the cost of a hollow report is not embarrassment. It is position sizing. A desk that ships templates will allocate on air, and air has no depth. Liquidity is not depth, it is just delayed panic. Coverage is not analysis, it is delayed ignorance. A nine-dimension framework applied to an empty array produces exactly the kind of confident, well-organized, unusable document that circulates in the third month of a drawdown, when everyone is desperate for a structure to hold onto.
In 2022 I modeled stablecoin de-pegging probability during the Celsius unwind and found that sixty percent of algorithmic designs lacked sufficient over-collateralization buffers. I shorted leveraged tokens and held USDC. That decision was cold and it was correct, and it rested entirely on an array that had entries in it. Nothing about my framework changed between that call and a template. The data did.
The consensus view is that AI research agents fail by hallucinating, and that better models will fix it. That view is comfortable and wrong. Hallucination is the failure mode that announces itself. The one that compounds is quieter.
The blind spot is in the benchmarks. Every evaluation rewards completion. A pipeline that returns null on an empty array scores zero on coverage, zero on dimension count, zero on throughput. A pipeline that returns nine filled sections scores well โ even when eight of them rest on nothing. The incentive structure therefore selects for exactly the behavior the industry claims to fear. You do not fix that with a larger model. You fix it with a schema that permits null and a downstream consumer that rejects any non-null value when the array is empty.
The second blind spot is attribution. A template has no author. Nobody signed the empty field, which is why nobody is accountable for it. The refusal in this log is notable precisely because it was signed โ a named escalation, three options, a recommendation. Accountability is the only feature that does not ship in the framework.
By 2028 the provenance layer will be the settlement layer. Agents will pay each other for data, and the ones that survive will be the ones whose counterparties can prove an input existed. The firms that clear the next cycle will not be the ones with the most dimensions. They will be the ones whose pipelines can return a blank and be believed.
The question is not whether your framework has nine dimensions. It is whether it has one that knows when to stop. When your agent hands you a complete report on an empty array, who in the chain is authorized to say nothing?