Last month a governance analytics vendor sent me a 41-page health report on a DAO I advise. Nine dimensions. Confidence scores. A remediation roadmap with a budget attached. The input array behind it was empty. Not corrupted β empty. No proposal data had resolved, no voter records had loaded, no treasury queries had returned. The pipeline did not throw an error. It imputed. Every field came back filled with something plausible, written in the flat, unbothered tone of a radiologist reading a scan of an empty room.
I have spent nineteen years in this industry watching abstraction outrun evidence. That morning was the first time I saw the failure mode automated, packaged, and invoiced.
This matters more in a bear market than it ever did in a bull one. Through 2024 and 2025 the governance tooling layer exploded β delegate dashboards, health scores, treasury runway models, all of it aggregated and summarized and sold back to communities that were too small to build their own. Now the budgets funding those tools are the same budgets being cut. Grant committees, delegates, and multi-sig signers are reading these reports to decide which teams get another quarter and which get wound down. In a bull market a bad metric costs you alpha. In a bear market it costs a team its runway, and the people running it their health insurance.
So it is worth being precise about how the numbers go wrong. Three mechanisms, in my experience, account for most of it.
The first is metric monoculture. Voter turnout, proposal count, TVL, unique voters. All four are trivial to query, which is exactly why they became the definition of "governance health." All four are also easy to game and none of them measure what the label claims. Identity isn't a row in a voter table, and turnout isn't a measure of legitimacy β it's a measure of how many wallets touched a Snapshot contract. In a sample of nine mid-cap DAOs I audited last year, stripping delegated wallets out of turnout cut the reported figure by roughly a third in six of them. The underlying number didn't change. The meaning did. Proposal count is worse: it rewards spam, and I have watched a treasury get drained by a governance process whose only visible symptom was a healthy proposal cadence.
The second mechanism is architectural, and it is the one nobody wants to hear. Modern report generators are built on schema-validated outputs with a retry loop. The model produces JSON, a parser validates it, and if a field is null the parser rejects the payload and asks again. Ask a helpful model for a non-null treasury runway number often enough and it will give you one. The imputation isn't a bug in a single product; it's the default posture of every AI-assisted report generator shipped in the last eighteen months. Helpfulness bias is baked into the contract between tool and user.
The third is provenance, or the absence of it. Every claim in a governance report should carry a query, a block range, and a snapshot hash β enough that a hostile reader can reproduce it. In a spot check of eleven vendor reports in Q4, I traced 214 factual claims and could find a query path for 83 of them. The rest were unfalsifiable. A claim you cannot falsify is not a finding. It is decoration.
We didn't set out to build a better dashboard. We built a stricter one. Three rules: every metric declares its query, its block range, and its null policy, and a null renders as null. The first run against a live DAO returned 38 percent empty fields. Nobody on the call liked the output. But the absence was the finding β the treasury data genuinely didn't exist at the granularity the report had been claiming for two years, and knowing that changed what the community funded next.
The same silent-imputation pattern shows up in rollup economics. Operators publish cost per transaction, but that number is proving cost amortized over projected batch occupancy. Proving cost is a bill paid in ETH every day. Occupancy is a forecast. When occupancy halves, real cost per user transaction doubles while the dashboard keeps rendering the amortized figure. In modeling work I did for two mid-size rollups last year, the true per-transaction proving cost at observed occupancy was a multiple of the number in the investor deck. I'm not naming them, because the pattern is the point and the pattern is everywhere.
It shows up in DeFi too, and Uniswap V4 makes it sharper. Hooks move fee logic, LP accounting, and a good chunk of risk into per-pool contracts. Aggregators index the factory and the pool manager. They do not read your hook. So TVL and volume still render, confidently, while the actual fee dynamics sit outside the indexer's field of view. Liquidity isn't the number in the pool table; it's a promise that someone will still be there at three in the morning on the worst day of the quarter. Composability without observability is just a longer list of things that can surprise you β which is why I still think the hook surface will scare off most of the developers who try it.
Lightning is the oldest example. Seven years of channel-capacity charts that looked healthy while routing failed. Capacity isn't liquidity. Channel count isn't a network.
The reflex, when you lay all this out, is to demand more transparency. Publish the queries. Open-source the models. I think that's wrong, and it's the part of this I'd most like to argue about. Transparency scales the imputation rather than fixing it β more dashboards, more surfaces, more confident nonsense with better citations. What's missing is falsifiability, which is a much stricter requirement: every number must be reproducible, every absence must be visible, and the tool must be allowed to fail loudly. A metric you can't falsify is decoration. And an empty dataset is a finding, which is precisely why communities resist it β a null forces a decision that a plausible number lets you postpone. We didn't have a data problem. We had a decision problem wearing a data costume.
Which brings me to the part that keeps me up. AI agents are already co-signing on DAO treasuries, and the near-term risk isn't a malicious agent. It's a helpful one. An agent that cannot distinguish an empty array from a solvent treasury isn't an auditor. It's a very fast intern with signing rights.
The DAOs that come out of this bear won't be the ones with the prettiest dashboards. They'll be the ones whose tooling can say three words out loud: I don't know. Legitimacy isn't the size of the vote. It's the presence of consent.