Computational integrity is a fundamental property that ensures the output of a
computation is provably correct and has been executed as intended.
The status quo today forces users to trust centralized AI operators to run
models correctly without manipulating their inputs or cheating them with worse
models.
Verifiable computing, powered by the computational integrity gadgets below,
enables any computation—whether conducted by a trusted or untrusted party—to
be verified for accuracy and correctness, without redoing the often complex
computation itself.
Ritual takes a credibly-neutral approach to computational integrity by enabling users to leverage different gadgets based on their app specific needs and their willingness to pay.
Ritual's modular design and flexible underlying architecture empower user choice.
Supported gadgets
Zero Knowledge Machine Learning
Strong cryptographic guarantees of correct model execution, at expense of
added overhead, complexity, and cost.
Optimistic Machine Learning
Optimistic acceptance of model execution, with model bisection based
verification only when disputes arise.
Trusted Execution Environments
Model execution with hardware-level isolation in enclaves, at expense of
trust in chip manufacturers and hardware attacks.
Probabilistic Proof Machine Learning
Low overhead and cost-efficient statistical guarantees of model execution,
at expense of consistently perfect verification.
Eager vs Lazy consumption
Ritual enables both eager and lazy consumption of proofs from supported gadgets. Lazy consumption enables use cases where computational integrity is only required in the sad path: Save costs: Lazy proofs are
generated only when disputes or errors occur Improve performance: Minimize proof verification
for applications with infrequent disputes Better developer experience: Build
simpler, easier to audit applications with fewer hot paths
Gadget trade-offs
A one-size-fits-all paradigm to computational integrity creates inherent trade-offs between security, cost, and performance. Each gadget has its own trade-offs and best use cases:Zero Knowledge Machine Learning
Zero Knowledge Machine Learning (ZKML) builds on zero-knowledge proofs to cryptographically assert correct execution of an AI model. Ritual’s ZK generation and verification sidecars enshrine this gadget natively, enabling users to make strong assertions of model correctness, with robust blockchain liveness and safety. Robust security: Offers the
strongest correctness guarantees via cryptography High complexity: Computationally
expensive, demands high resources, and is slowest Limited support: Only simple models are supported by modern ZKML proving systems today
Optimistic Machine Learning
Optimistic Machine Learning (OPML), inspired by optimistic rollups, assumes model execution is correct by default, with verification occurring only when disputes arise. At a high level, the system works as follows:- Model execution servers stake capital to participate
- These servers then execute operations, periodically committing intermediary outputs
- If users doubt correctness, they can contest outputs via a fraud proof system
- The system views models as sequences of functions and uses an interactive bisection approach, checking layer by layer, to identify output inconsistencies
- If model execution is indeed incorrect, server stake is slashed
Cost effective: Especially efficient for use cases where disputes rarely occur Extended support: Bisection approach better
supports large, complex models (like LLMs) Weaker security: Relies on
incentivized behavior rather than cryptographic security Complex sad path: Dispute resolution is lengthy, complex, and demands some re-execution
Trusted Execution Environments
Trusted Execution Environments (TEEs) provide hardware-based secure computing through isolated execution zones where sensitive code and data remain protected. Ritual’s TEE Execution sidecar enshrines this gadget natively by executing AI models in secure enclaves enabling data confidentiality and preventing model tampering. Performant: Enables sans-gadget competitive performance for most AI model types Real-time: Better suited for real-time
applications with limited proving complexity or overhead Vendor trust: Requires trust
in chip manufacturers and secure enclave software Hardware attacks: Susceptible to sophisticated side-channel hardware attacks
Probabilistic Proof Machine Learning
Most model operations are computationally complex, especially when performing resource-intensive operations like fine-tuning or inference of modern LLMs. To better support these operations with a low computational overhead tool, Ritual has pioneered a new class of verification gadgets, dubbed Probabilistic Proof Machine Learning. The first of this line of tools is vTune, a new way to verify LLM fine-tuning through backdoors. Computationally cheap: Time and cost-efficient for even the most complex model operations Third-party support: Suitable for trustlessly
verifying third-party model API execution Statistical correctness: Not suitable for when perfect verification guarantees are necessary