The Null Input Problem: When a 0-Star Risk Report Reads as 'Low Risk'

Last week a nine-dimension risk report crossed my terminal with a clean bill of health on every axis. Technical layer: N/A. Tokenomics: N/A. Regulatory exposure: N/A. Risk matrix: N/A across all six categories — technical, market, operational, regulatory, competitive, narrative. Composite scores: zero stars on technical value, investment value, timeliness, and reference value.

That is not a bearish report. It is not a bullish report either. It is a pipeline that ingested zero bytes of the article it was built to analyze, and then — instead of throwing an error — printed a fully formatted table anyway.

Follow the gas, not the hype. The gas here is zero. And zero gas on a supposed mainnet transaction is the loudest signal in the room.

The Oracle Nobody Checked

Most on-chain research stacks run two stages. Stage one deconstructs source text into structured information points: title, publisher, timestamp, article type, domain tags, named protocols, time sensitivity, source-quality grade. Stage two consumes that JSON and runs the analytical dimensions — technical, tokenomic, market, ecosystem, regulatory, governance, risk, narrative, supply-chain transmission.

Stage one is the oracle. Stage two has no independent access to reality. If stage one returns an empty list and nobody checks the length, stage two will compute on nothing and emit output with the same formatting confidence as a real report.

I built my first version of this in 2018 in Jakarta, during the post-ICO winter — three hundred hours of Python scraping raw Ethereum mainnet transactions, fifty-plus ICO contracts audited by hand for reentrancy patterns the community had glossed over. The lesson that stuck wasn't about Solidity. It was that a clean schema with no row-level validation is just a very orderly way to be wrong.

What Actually Arrived

Here is what the pipeline received. Every field was missing or placeholdered: no title, no source, no article type, no domain classification, no core claims, an empty information-point list, no identified protocols, no time-sensitivity assessment, no source-quality grade.

Critically, the stage-two report knew this. It refused to speculate. It marked every dimension N/A and explicitly warned that reading "no information" as "neutral" or "low risk" would be a category error.

That is correct behavior. It is also the exception, not the rule.

Empty input is a third state. It is not zero. It is not unknown. It is the absence of a question ever having been asked. Systems that collapse all three into one bucket are structurally incapable of distinguishing a safe protocol from an unaudited one.

I enforce one hard rule in every pipeline I ship: no metric without provenance. Every number carries a source URL, a fetch timestamp, and — where the data is on-chain — a block height. Without a block height, you do not have a data point. You have an anecdote with decimal places.

The rule exists because I've been burned. In 2020, tracking liquidity-pool ratios across twenty DEXs and over 100,000 events, a subgraph would occasionally time out. The default return was 0, not null. A pool ratio of 0 reads as a 100% drain. I once escalated a suspected $40 million exit that turned out to be an indexer restarting. The bug wasn't in the math. It was in the type system.

Liquidity mining taught the same lesson from the other direction. An incentivized pool's APY is not a yield; it is a subsidy with a countdown timer. When emissions stop, the TVL it bought leaves within days. Aggregate APY without separating emission-funded yield from fee-funded yield and you are measuring marketing spend, not protocol health.

In 2022 I traced over 500,000 TerraUSD redemption transactions and found the liquidity gap six weeks before the peg broke. That signal existed only because the ingestion layer retained block-level timestamps. Run the same data through an empty-tolerant aggregator and the gap averages into the noise.

After the 2024 ETF approval I aggregated inflows from fifteen issuers against exchange reserve balances. The concentration trend in long-term holder supply was visible only because each reported flow could be reconciled to a specific block range. Reported flow without a block range is a press release.

My 2025 congestion model hits 78% accuracy on gas-fee spikes across the top 100 Ethereum accounts, trained on five years of history. It hits that number only when input windows are complete. Feed it a gap and it will confidently forecast a spike that already happened.

The same shape of error hits Layer 2 comparisons. Dashboards rank rollups by TVL pulled from a single indexer. When that indexer lags, the chain with the most aggressive deployment pipeline looks like it has the best technology. It doesn't. It has the most teams who agreed to ship on it. The technology question is real, but it sits downstream of the distribution question.

And it hits Bitcoin. An ingestion job that drops coinbase-only blocks shows fee revenue flatlining — right through the inscription wave that has been carrying miner economics since 2023. You would conclude the security budget is deteriorating. You would be measuring your own filter.

The Wrong Suspect

Everyone in that incident review will blame the model. Wrong target. The model behaved exactly as a well-formed function should on empty input: it returned the identity element. Nothing hallucinated. Nothing invented a TVL figure.

The defect is architectural. The pipeline was built to degrade rather than to halt. Somewhere, a design decision was made that a partially populated report is more useful than an error. In quantitative finance that assumption gets you liquidated. In a research pipeline it gets you something worse: a document that looks like diligence.

The second failure is semantic. "Zero stars" reads to a human as "bad asset." It means "no asset was named." That distinction does not survive a single hop into downstream alerting. Tag a protocol with a 0-star risk score and an hour later someone's position sizing has quietly anchored to it. The model didn't lie. The interface did.

The report's own final section was the only content of substance: the highest-rated risk in the entire document was input-integrity risk, rated high, with a recommendation to add a null-input circuit breaker — hard-fail on missing required fields instead of silently downgrading output. Process risk is not asset risk. A broken pipe tells you nothing about the water.

Provenance or It Didn't Happen

The fix is boring and it is the whole ballgame: require a minimum viable input — an information-point list, at least one named protocol, a one-line claim, a source, a timestamp — and refuse to run without it.

Next week, watch for provenance. Open any dashboard you trust and ask where the block height comes from. Code is law, but bugs are fatal — and the most expensive bug is the one that makes a system look like it ran.

Whales don't panic. They just stop signing. A pipeline that returns silence and calls it analysis is doing the same thing.