Hook
There is a class of trading loss that never appears in an exploit database, never triggers a governance post-mortem, and never receives a CVE identifier. It happens when a number that should not exist is silently substituted for a number that does.
I have watched this pattern four times in eighteen months, and the shape is always identical. An upstream feed goes quiet — a WebSocket drops mid-session, a REST endpoint hits its rate limit, a subgraph indexer falls forty thousand blocks behind. The normalization layer receives nothing. The normalization layer, obeying a default argument written by an engineer who left the firm nine months ago, converts nothing into 0.0. Downstream, a signal generator that has never been taught the difference between equilibrium and absence reads 0.0 and sizes a position against it.
No oracle was manipulated. No private key was compromised. No vault was drained. The system did not lie. It went quiet, and the silence was rendered as a number.
The most expensive failures in crypto infrastructure are not wrong numbers. They are missing numbers wearing the costume of zero. Exploit dashboards will never show you this, because there is no attacker to name.
Context: Where Absence Goes To Die
A crypto trading stack has five layers, and exactly one of them has a correct concept of null.
The venue layer — exchange REST and WebSocket APIs, RPC nodes, indexers — represents absence explicitly. A null, an empty array, an HTTP 429, a closed socket, a reorg that orphans a block. Absence is loud here, and loud is good.
The ingestion layer — retries, backoff, buffering, dead-letter queues — usually handles absence correctly, precisely because the failure is visible and someone gets paged.
The normalization layer is the problem. This is where schema enforcement, type coercion and unit conversion happen. This is where a JSON null becomes float(0.0), because the schema says number and not number or null. This is where absence dies.
The signal layer — indicators, model inference, agent reasoning — receives that 0.0 and has no mechanism to distinguish it from a genuine equilibrium. A zero is a zero.
The execution layer sizes the position.
Layer three is the crime scene, and it exists for defensible reasons. You want deterministic schemas. You want backtests to run without branching on every field. You want float and not Optional[float], because vectorized numpy is faster that way. Every one of those reasons is simultaneously a reason to erase absence.
TradFi solved a version of this and crypto has not. US equities have a consolidated tape, a single official last sale, and exchange-level Limit Up/Limit Down bands with trading halts. When the primary feed breaks, the market stops. Crypto has no consolidated tape, no halt authority, and hundreds of venues that never agree on price. There is nobody to stop you.
Volatility is the tax on uncertainty — and unfalsifiable data is the highest tax rate in the market, because you pay it without ever knowing you were charged.
The regulatory layer arrived between 2024 and 2026 and converted this from an engineering question into a compliance one. Operational resilience requirements under MiCA and the EU AI Act's high-risk classification for automated trading agents both require an auditable, timestamped record of every input that drove an automated decision. An input that was never received is still an input. If your logs show 0.0 where the truth is null, you have not built an audit trail. You have built a fabricated one — and under current reporting regimes, that is an incident in its own right.
Core: Seven Places Nulls Hide
1. On-chain zero and off-chain null are not the same object
Solidity has no null. A uint256 is thirty-two bytes of value. Zero is a legitimate value, and there is no separate absent state. This is not an oversight; it is a consequence of deterministic state machines. But it means the on-chain world possesses no vocabulary for "I don't know."
Off-chain, we inherit Python, JavaScript and JSON — languages in which null, None, NaN and 0 are four distinct things. The bridge between these two worlds is where absence gets laundered into fact.
| Layer | Absence represented as | Rendered downstream as | Loud? | |---|---|---|---| | Solidity uint256 | Not representable | 0 | No | | Solidity balanceOf(badAddress) | Revert | Exception | Yes | | Chainlink stale round | Old answer, old updatedAt | Last known price | No | | REST 429 / timeout | Exception or null | 0.0 under coercion | No | | WebSocket disconnect | Closed socket | Last tick repeated | No | | JSON {"price": null} | null | NaN → 0.0 | No | | Empty array [] | Length zero | sum([]) = 0 | No | | Subgraph indexer behind | Old block height | Stale TVL, unflagged | No |
Read that column again. Only one row fails loudly. Every other row fails in the direction that resembles normal market conditions. A flat funding rate looks like equilibrium. A frozen TVL number looks like consolidation. A repeated last tick looks like a quiet session.
2. The staleness check that everybody skips
Chainlink is the canonical illustration because the interface tells you exactly what to do and most integrations ignore it.