Something people can contribute
A person picks up an object, turns it over and places it on a board. Their fingers bend around it. Shadows, reflections, clothing and skin change together. If several cameras record the action, they capture different views of the same event.
For the person, this is an ordinary movement. For a system producing artificial evidence, it is a collection of relationships to reproduce: between objects, between viewpoints and between moments in time.
Nova's central idea is to reward work that is easy for people to perform and costly to imitate. Participants carry out simple physical tasks and record them. Public verification models inspect selected parts of the recordings. The aim is to make accepted evidence cheaper to produce through human participation than through its synthetic or robotic substitute.
This addresses a problem known as a Sybil attack. A permissionless network (one that anyone can join) cannot distribute influence merely by counting accounts. One operator can create many accounts. Nova instead ties influence to accepted work.
Participation does not depend on outperforming AI at reasoning or creative work. The task could involve moving a board and handling a few objects according to instructions, without requiring unusual knowledge or strength. Accessible task variants could accommodate different physical abilities, provided they offer comparable verification and resistance to automation.
The useful human contribution is the performance itself. A real event produces a continuous physical history, with many details agreeing automatically. A synthetic producer must create observations that agree well enough to pass the network checks.
Nova's economic case therefore rests on the cost of reproducing the details its verifiers actually inspect. Its objective is a substantial, measurable cost advantage for human work, maintained as technology changes.
Synthetic evidence may pass. That is compatible with the design: acceptance follows public rules, while the task and verification requirements aim to make automated production an expensive way to obtain influence.
Choosing what to inspect
Section 2 describes how the miner commits to evidence and how Proof of Work (PoW) determines an audit. The commitment binds the miner to the submitted material. Once an audit has been obtained, changing that material invalidates the commitment and its associated work.
Assuming video evidence, the complete recording for each task is divided into consecutive video clips. Each clip (except the first one) starts with the frame immediately following the previous clip, so the evidence covers one uninterrupted performance without omitted intervals. When several cameras are used, their streams must follow the same task timeline and be synchronized.
This division makes a long recording manageable to inspect. The task hash (a digital value from which the instructions are derived) determines the assignment. The clip's position in the evidence Merkle tree determines which step should be visible. Therefore, it’s easy to reconstruct the expected actions for any selected clip.
For example, a step might require the participant to hold object O5 over square D8, tilt the board left with the other hand, move O5 toward O2, and then pick up O2 and press it against their chest. The checks would target observable relationships: which object is held, where it moves, how the board tilts and whether the actions occur in the required order.
An illustrative block in the network might disclose twenty short clips in total. Some could be selected in pairs from the same task, allowing a detail visible early in the performance to be compared with its later appearance. Other selections could compare simultaneous views from different cameras.
The public aggregation rule combines the selected verification models' outputs into the result , as specified in Section 2. It must define which conditions are indispensable and how much ordinary measurement variation is acceptable. A useful rule rejects inadequate evidence while reliably accepting correctly recorded human work.
The sampling mechanism matters to the producer's incentives. A miner can search for a more favourable audit, but every qualifying alternative requires the corresponding PoW expenditure. Repairing committed evidence also requires a new commitment and new work. The protocol must price these strategies so that preparing broadly acceptable evidence is more economical than repeatedly seeking a sample that misses its defects.
Three connected kinds of checks
Nova needs a verification bundle designed around its tasks and capture conditions. Existing computer-vision models provide useful building blocks, but they must be adapted. The bundle checks the assigned actions, agreement within and between views, and continuity over time.
Task compliance
The first group of checks measures the required objects, positions, movements, sequence and visible outcomes.
For example, SAM 2 can separate a selected object from the rest of an image and follow its image region through a video. Source: https://github.com/facebookresearch/sam2.
FoundationPose estimates and tracks an object's position and orientation in three dimensions. Source: https://github.com/NVlabs/FoundationPose.
The performer can also be tracked. MediaPipe Face Landmarker estimates facial landmarks, while MediaPipe Pose Landmarker estimates body landmarks. These points provide a way to describe and compare visible positions and movements. Sources: https://ai.google.dev/edge/mediapipe/solutions/vision/face_landmarker and https://ai.google.dev/edge/mediapipe/solutions/vision/pose_landmarker.
Similar measurements would let Nova express a task as a sequence of specific, observable changes. The verifier can check whether O3 crossed square D1 while the board tilted, etc., rather than making one broad judgment about whether the participant appeared to follow the instructions.
Spatial coherence
The second group of checks examines agreement within a frame and across camera views. An object should occupy compatible positions in different views. Fingers should surround the object they appear to hold. The hand, forearm and body should connect plausibly. A shadow should move in a way compatible with the object and the lighting, etc.
For example, MEt3R illustrates how reconstructed geometry can align different views and measure their agreement. Source: https://arxiv.org/abs/2501.06336. DTTNet detects shadows in video and models their changes over time. HandEval assesses the visual quality of hands in generated images. Sources: https://arxiv.org/abs/2511.06925 and https://arxiv.org/html/2510.08978v1.
A practical distinction is important here. Checking whether two views show the same finger position is an agreement problem. Establishing the exact physical force used is a much harder measurement problem. Nova can concentrate on visible agreement and required outcomes without reconstructing underlying physical quantities.
Temporal coherence
The third group of checks follows details through time. An object can disappear behind a hand and reappear, but its appearance shouldn't change. Details like the performer’s body, clothing, and environment should stay perfectly consistent.
Comparisons between distant moments are useful because an AI-synthesized video may drift gradually. A later view of an object can reveal changes that a short check would miss.
For example, TRAJAN evaluates motion using tracked point trajectories and measures properties associated with temporal consistency and realism. VBench includes automated measures of consistency, temporal flicker and motion smoothness. Sources: https://trajan-paper.github.io/ and https://github.com/Vchitect/VBench.
Nova would need targeted comparisons of corresponding surface regions and object shapes. A patch of skin or cloth does not retain identical pixel values when it turns, stretches or moves into shadow. The check must account for legitimate changes in viewpoint, lighting and deformation while detecting inconsistent changes.
Capture rules should make this easier: proper resolution, usable lighting, limited motion blur, visible props and appropriate synchronization between cameras.
These three check groups must be evaluated as one system. Their thresholds need to accommodate real recording conditions while rejecting defects that make synthetic production cheap. Tighter checks help only if people can reliably pass them.
Verification on consumer hardware
The producer may need to prepare a large body of evidence, while each audit examines a small part of it. This is how Nova aims to keep verification affordable even when producing acceptable evidence is demanding.
Full nodes with more resources could use expensive accelerators to perform all checks required for a block. Light clients could download only the evidence randomly selected by them and repeat a smaller number of those checks on ordinary computers. They would simply apply the same models and rules to less evidence.
Finding invalid evidence would establish that the affected block (and the DAG history relying on its validity) must be rejected. A small sample provides probabilistic assurance, whose strength grows to practical certainty with enough random checks.
At an assumed bitrate of two megabits per second, twenty five-second clips attached to a block total approximately 25 MB, before proofs and other metadata; a realistic data budget.
Model storage and compute need equivalent budgets. A small video download does not automatically mean a small inference workload.
One result on different machines
Consensus requires different computers to reach the same result from the same evidence. Ordinary machine-learning software does not guarantee this: PyTorch's reproducibility documentation explains that results can differ across releases, platforms and CPU/GPU execution. Source: https://docs.pytorch.org/docs/stable/notes/randomness.html.
Nova must therefore specify the complete verification procedure: video decoding, preprocessing, model weights, arithmetic, rounding, etc.
I-ViT, which demonstrates integer-only Vision Transformer inference, illustrates one available approach. Such techniques could help build a deterministic bundle, while the complete implementation would also need to take inference speed into account. Source: https://arxiv.org/abs/2207.01405.
What would alternative production cost?
There are three main routes to producing evidence without the intended human work: conventional computer graphics, AI generation and physical automation. An attacker can combine these methods.
The relevant quantity is cost per accepted task, after possible cost optimizations are taken into account. Published benchmarks give concrete starting points for this analysis.
For the following examples, let's assume that one paid hour of human work costs $6 and produces thirty minutes of accepted video evidence, recorded simultaneously by four cameras. The $6 figure is an illustrative labor assumption. The four cameras record the same half-hour, yielding 120 camera-minutes of evidence.
The examples below use advertised GPU rental prices. Operators owning hardware may have different economics.
Conventional computer graphics
Conventional CGI builds images from explicitly represented geometry, materials, lighting and animation. It can preserve temporal consistency (the identity and state of objects) indefinitely. The costly part can be representing and rendering enough detail to satisfy the required realism and task adherence checks.
In a test reported by iRender, an interior scene rendered in Blender Cycles took 8 minutes 12 seconds per frame on an RTX 4090, at 512 samples with denoising disabled. The report does not specify resolution. Source: https://irendering.net/ai-denoising-for-rendering-how-optix-oidn-nlm-cut-your-render-time-by-70/.
At an advertised RTX 4090 rate of $0.34 per hour, this is about $0.0465 per frame. A minute at 30 frames per second contains 1,800 frames: 246 GPU-hours, or $83.64 at that rate. Four equal-workload camera views would cost $334.56 per minute, or $10,036.80 for thirty minutes. Pricing source: https://www.runpod.io/gpu-models.
Optimization changes the result substantially. The same article reports 2 minutes 48 seconds per frame at 128 samples with OptiX denoising, equivalent to $28.56 per minute at 30 fps under the same rental assumption, or $3,427.20 for thirty minutes across four independently rendered views. Source: https://irendering.net/ai-denoising-for-rendering-how-optix-oidn-nlm-cut-your-render-time-by-70/.
This workload illustrates a potentially large cost advantage for physical recording. Nova's task design and verification bundle must preserve that advantage against cheap methods by demanding costly-to-produce details.
Simulate the task, then generate its appearance
A serious attacker could first build a rough simulation of the required task, then use an AI video synthesis model to turn guidance derived from that simulation into realistic video.
This approach can save substantial work through neural generation of a convincing appearance. Nevertheless, each task still requires action sequences and changes to the scene that follow its specific instructions.
A major challenge is preserving the required details consistently across the generated frames. For example, guidance may correctly specify an object’s shape and surface markings, yet the generated video may alter them over time.
Generation costs depend on what a video must demonstrate. A short clip may look convincing while failing to follow a prescribed action sequence or preserve details across time and camera views. For Nova, these are checkable properties: where objects are, how they move, whether their features remain consistent and whether the required interactions occur.
The following examples illustrate two relevant workloads: inexpensive short-clip generation and generation guided by a specified scene structure.
Fast short-clip generation. FastVideo reports an optimized LTX-2.3 pipeline producing five seconds at 24 fps, at 1088 × 1920 resolution, with an end-to-end runtime of 4.55 seconds on one NVIDIA B200. Source: https://haoailab.com/blogs/fastvideo_realtime_1080p/.
At $5.98 per B200-hour, that is approximately $0.00756 per clip, or $0.0907 per minute of output at the same throughput. Scaling the frame volume to four independently generated streams gives about $0.363 per minute of performance, or $10.88 for thirty minutes, assuming equal workloads. Pricing source: https://www.runpod.io/gpu-models.
This demonstrates how inexpensive short-clip production can be.
Generation with structured control. One way to improve task adherence is to provide the video model with guidance showing the required activity. NVIDIA’s Cosmos-Transfer2.5 supports control inputs such as depth maps, edges and segmentation masks. Segmentation masks identify object regions in each frame, providing guidance about their positions, outlines and movement. Source: https://huggingface.co/nvidia/Cosmos-Transfer2.5-2B.
NVIDIA’s inference guide reports 186.92 seconds end to end on one B200 for a 121-frame segmentation-control input, involving two 93-frame generation chunks. The benchmark uses 720p video at 16 fps, with guardrails disabled. Source, “Segmentation” table and its footnotes: https://github.com/nvidia-cosmos/cosmos-transfer2.5/blob/main/docs/inference.md#segmentation.
The model card specifies output of the same temporal length as the input. Using 121 frames at 16 fps gives approximately 7.56 seconds of output. At the same rental rate and throughput, the calculated cost is $2.46 per output minute, approximately $9.85 per minute across four independently generated streams, or $295.61 for thirty minutes across four streams, assuming equal workloads. Model specifications: https://huggingface.co/nvidia/Cosmos-Transfer2.5-2B.
Structured guidance does not ensure that every required property survives generation. NVIDIA reports imperfect control adherence, temporal inconsistency, unstable motion and imprecise interactions. Source: https://huggingface.co/nvidia/Cosmos-Transfer2.5-2B.
| Published workload | One minute, one stream | One minute, four independent streams | Thirty minutes, four independent streams |
|---|---|---|---|
| Cycles interior, 512 samples, no denoising; extrapolated to 30 fps; resolution unspecified | $83.64 | $334.56 | $10,036.80 |
| FastVideo / LTX-2.3 short clips; 1088 × 1920, 24 fps | $0.0907 | $0.363 | $10.88 |
| Cosmos-Transfer2.5 segmentation-guided generation; 720p, 16 fps | $2.46 | $9.85 | $295.61 |
Under the illustrative human-work assumption, the corresponding labor cost is $6 for the thirty minutes of performance shown in the last column. Nova’s goal is to make cheap production inadequate by checking task adherence, realism and consistency together. The relevant cost is what an optimized producer must spend to satisfy those checks in its final evidence.
Why control and consistency can add cost
A synthetic video must keep details consistent as objects move or come back into view. Video synthesis does not automatically achieve this: the generator may change an object’s texture or invent details that disagree between frames.
Warped Diffusion provides a measured example of the extra computation used to correct these problems. In its video-upscaling experiments, it uses RAFT, an AI model, to estimate movement between frames. During generation, it checks whether corresponding details agree after accounting for that movement, then uses the disagreement to guide corrections.
These corrections make generation more expensive. On one NVIDIA A100, the researchers report using 50 denoising steps instead of 25, while processing speed falls from 4.32 to 1.62 steps per second because each step includes additional calculations. Using the step counts and processing speeds reported in Appendix F.3 gives an estimated 5.33-fold increase in time spent on denoising calculations.
In the real-video upscaling experiment, the automated measure of inconsistency between frames falls from 0.812 without the correction mechanism to 0.649 with it, a reduction of approximately 20% in the measured error. Source: Table 3 and Appendix F: https://arxiv.org/html/2410.16152v1.
This illustrates the mechanism relevant to Nova: the consistency rule is known to the producer, yet using it to improve the generated evidence still adds substantial computation.
Public verifiers and adaptation
Nova’s verifiers are white box: their rules and models are public. An attacker can test outputs and train a specialized producer against the verifier bundle. These adaptations can reduce production costs.
Knowing the requirements still leaves the problem of producing evidence that satisfies them together, while a physical performance supplies these relationships naturally. Nova seeks to turn that advantage into a measurable production-cost gap.
What about physical automation?
A humanoid robot could simply perform the task physically. This bypasses digital rendering and produces realistic and continuous evidence. The attacker would need to pay for machinery, operation and maintenance. If the network verification bundle also checks signs of a human performer, the robot would need to reproduce those observations convincingly.
Production forecasts make this a relevant long-term threat:
Bank of America, March 2026: 1.2 million annual humanoid shipments in 2030, 10 million in 2035, and 300 million robots in operation by 2040. Source: https://institute.bankofamerica.com/content/dam/transformation/physical-ai-part-2.pdf.
Berg Insight, September 2026: 26 million annual humanoid shipments by 2040. Source: https://www.berginsight.com/annual-humanoid-robot-shipments-to-reach-26-million-units-worldwide-by-2040.
Morgan Stanley, May 2025: approximately one billion humanoids in use by 2050. Source: https://www.morganstanley.com/insights/articles/humanoid-robot-market-5-trillion-by-2050.
For Nova, the relevant subset is machines that can both perform the assigned task and satisfy the required observations of the performer. Reproducing human appearance and bodily behaviour under detailed examination is an additional demand beyond manual capability.
Better capture can expose more of the difference
Robotics and verification will develop over the same decades. Better hardware sensors could let Nova inspect more of the human body's complexity: skin reflectivity and deformation, eye responses, temperature patterns, and other bodily changes that depend on a task. The objective would be to make imitation expensive while keeping measurement practical.
World, formerly Worldcoin, provides a concrete example of combining bodily signals. Its Orb uses iris imagery alongside infrared imaging and depth sensing, then checks for genuine human presence. World describes layering these checks to increase the cost and complexity of attacks. The same design principle could inform Nova's future capture systems. Source: https://whitepaper.world.org/achieving-proof-of-human.
An evolving system
Nova's long-term task is to maintain the desired economic relationship between accepted human work and its cheapest substitutes. Its decentralized update process gives individual contributors a role in deciding how the network responds as synthetic generation, robotics and evidence verification methods improve.
A network update could change the task, capture requirements, verification models, model score aggregation rule or audit and PoW parameters. Its purpose would be to restore or widen the cost advantage of human work while preserving affordable participation and verification.
Proposals should be accompanied by reproducible tests. The community needs to see how often ordinary participants pass, what recording equipment they need, how much validation costs and how cheaply an optimized attacker can produce accepted submissions.
This makes public evaluation central to governance. Contributors can examine an update and decide whether to adopt it without accepting its supplier's claims on trust.
The desired cost margin should be large enough to provide time to respond to improvements in attack methods. Monitoring those improvements and updating the network before the margin disappears are continuing responsibilities.
Nova
Behind Nova lies a simple question: what connects our agreement to reality?
A social or digital consensus process can follow every rule of its protocol and still gravely misrepresent the world outside it. Nova seeks to draw the strength of its proof from reality itself.