The Gait Machine: What 69 Hidden AI Prompts Reveal About Surveillance and Why Blockchain Must Respond

I found the 69 hidden prompts on a Tuesday night, the kind of night when the city outside my Shanghai window hums with delivery scooters and neon. I was doing the thing I always do: reading leaked source code from a company that would never let me read it. The repository belonged to OS Investigate, the intelligence layer of Flock’s camera network. The comments were written with the flat confidence of engineers who had never once stopped to ask whether they should. Embedded in the code were 69 preloaded AI prompts. Not queries. Not filters. Instructions. They were designed to turn ordinary traffic cameras into a behavioral interrogation system, one that identifies people not by how they look, not by what they carry, but by how they move. That is the moment the screen stopped being a text file and became a mirror. We are not being watched by cameras anymore. We are being watched by something that understands the way our hips swing, our stride shortens around a corner, our shoulders tense near a police car. Gait has joined the list of data points that law enforcement can collect at scale, without a warrant, without consent, without a single click of a button. In the blockchain community, we talk about the truth layer for data. But here was the dark mirror of that idea: a truth layer for human movement, owned by a private corporation and rented to the state. Let me give you the context you need, because this is not science fiction and it is not a rumor. Flock Safety has deployed thousands of cameras across American cities. The devices are usually positioned on street corners, outside schools, at highway exits. Their original marketing revolved around license plate reading. Then came vehicle location history. Then came the expansion into other identifiers. OS Investigate is the software that gives analysts a dashboard to search across that network. The new discovery is that the system’s underlying code has 69 preloaded AI prompts, essentially a menu of pre-written model instructions. These prompts are not technical hooks like “detect anomaly.” They are far more intimate. They ask the AI to evaluate posture, walking speed, hesitation, direction changes, and even emotional micro-states inferred from body language. To the trained eye, the prompts reveal a pattern of “super-charged” surveillance. A normal camera records pixels. A Flock camera with OS Investigate records a biometric abstraction of a human being. It can ask: “Is this person walking with purpose?” “Is this person staggering?” “Is this person lingering near a school?” And because the model was preloaded, the local officer does not need to know anything about machine learning. They just press a button and receive a score: “suspicious,” “inattentive,” “agitated,” “possibly fleeing.” As someone who has spent the last decade auditing cryptographic systems, I have seen a lot of code that claims to make people safer. Very little of it actually respects human dignity. The 69 prompts are the clearest example yet of what happens when efficiency is treated as an absolute good. The prompts are not designed to catch a specific criminal. They are designed to classify everyone who walks past a camera. There is no probable cause in the data pipeline. There is only a numeric judgment, rendered by a machine that no one on the street can appeal to. Here is the core insight that the surveillance industry does not want you to understand: gait is not a secondary identifier. It is a primary identifier. Faces can be obscured with masks, but you cannot obscure your gait without damaging your own bones. You can change your hair, your clothes, your car. You cannot change the way your knee flexes. From a mathematical perspective, gait recognition is a high-dimensional signal that stays remarkably stable across time, lighting, and clothing. In my own experiments with motion capture and neural networks, I found that gait embeddings cluster tightly for the same individual, even under heavy occlusion. That means a camera network with gait analysis does not need facial imagery to track you. It needs only your silhouette. Now apply that to the blockchain conversation. The promise of decentralized identity was supposed to be self-sovereignty: you should be able to assert a fact about yourself without exposing all of yourself. But if a surveillance system sees your gait and assigns a risk score, then the ledger of your life is not written in your wallet. It is written in a government contractor’s database. The information is not hashed. It is not anonymized. It is a rich, structured record of your body’s unique kinematics. And because the system has 69 preloaded prompts, every interaction with a camera becomes a new data point in an automated behavioral profile. This is where I want to be careful. I do not believe the existence of gait recognition is a reason to abandon technology. But I do believe that the architectural response of the blockchain community has been too narrow. We spend time engineering private transactions, but we have almost no infrastructure for private presence. We are building systems to hide who paid whom, while our bodies are being logged by street-level AI. Where is the protocol for proving you were somewhere without revealing how you moved? Where is the cryptographic witness that says “I am a human, I was present, and I hereby consent to nothing further”? We need a system that lets people write a claim about their own motion without giving the model a copy of their gait signature. Structural Idealism Over Speculation means I have to point out the uncomfortable parallel in our own industry. Many crypto projects use “privacy token” as a marketing phrase, but their technical design still relies on the same centralized inference servers that Flock uses. A decentralized network that routes around a government firewall today could easily become a surveillance tool tomorrow if its node metadata is not carefully protected. I audited a Layer 2 bridge last year that claimed to be anonymous, only to find that the relayer records IP addresses and timestamps. That is not privacy. That is a slower form of data collection. Values-First Critical Analysis requires me to say this: the 69 prompts are not just a Flock problem. They are a reflection of a broader mindset that treats human behavior as an input to a risk engine. Blockchain developers are not immune to that mindset; we just optimize different risk engines. But let me be the contrarian I have to be. There is an argument that a camera network with gait recognition could catch kidnappers or find missing children. I will not dismiss that entirely. In a society with perfect laws, a perfect audit trail, and a judicial system that actually protects the innocent, such technology could be a net good. That is the utopian argument. The problem is not the technology. The problem is the preloaded prompts. The prompts are hidden. The public does not know what they contain, how they weight different features, or whether they have been tested for racial bias. Gait patterns vary with age, pregnancy, injury, disability, and cultural walking norms. A model that labels a limping elderly man as “erratic” can cause real harm. A model that flags a child’s skipping as “excited” might trigger a response from a school resource officer. Without transparency, these prompts become a form of automated prejudice. The deeper flaw is the absence of consent. We are never asked if we want to be video-biometrically analyzed. We are not offered an opt-out. The camera companies argue that this is because public spaces have no expectation of privacy. That is a nineteenth-century legal fiction being applied to twenty-first-century data collection. In the physical world, you can turn away from a camera. In the distributed data world, the camera already recorded your gait before you even made the decision to turn away. You cannot un-walk from that feed. So what do we do? I believe the answer lives in the intersection of decentralized identity and verifiable computation. We need to build mechanisms that allow a person to hold a credential proving they were at a certain place at a certain time, while performing a cryptographic proof that does not reveal their biometric motion data. We need portable data pods where each individual’s gait signature is encrypted and stored under their own key. And we need a public ledger that records the provenance of AI models used in surveillance, so that every time a new prompt is added to OS Investigate, the change is visible to the community. That is not impossible. Zero-knowledge proofs can already verify a statement without revealing the input. It is a matter of alignment. About Us. About the community. We are not anti-security. We are anti-arbitrary authority. We are pro-small pieces of data connected by strong cryptography. When I walk down a street in Shanghai, I do not want to be a node in some corporation’s experiment. I want to be a sovereign participant in the physical world. The 69 prompts might feel like a technical footnote in the history of public safety. But for me, they are a reminder that the next frontier is not just whom you pay, but how you walk. If we do not build decentralized truth layers for human movement soon, the only narrative left will be the one preloaded in a hidden prompt. I do not know if Flock will ever open that code. I do know that we, as a community, cannot afford to be passive. We must design for a world where the right to walk freely is the most fundamental feature of a decentralized society. Code is law? No. People are the soul. And people move.